Methods for digital phenotyping and relation thereof to genotypes

Abstract

Provided herein is a method of training an artificial intelligence (AI) model to generate digital phenotypes. The method includes providing a training set including digital descriptors and training the AI model to generate the digital phenotype using the training set, wherein the training is configured to relate the digital descriptors to a health condition to generate the digital phenotype. Also provided herein are methods of generating a digital phenotype for a subject using the trained AI model and methods of predicting a health status of a subject based upon the generated digital phenotype.

Description

METHODS FOR DIGITAL PHENOTYPING AND RELATION THEREOF TO GENOTYPESCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] The present application claims priority under 35 U. S. C. § 119(e) to U. S.Provisional Patent Application No. 67 / 719,916, filed November 13, 2024, which application is incorporated herein by reference in its entirety.STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

[0002] This invention was made with government support under R01 AA031959, K23AA026890, and U54AA027989 awarded by National Institutes of Health (NIH). The government has certain rights in the invention.BACKGROUND OF THE INVENTION

[0003] Brain diseases and disorders affect a large fraction of the population worldwide. They are often difficult to characterize due to their complex phenotypic and genetic architecture. For example, psychiatric disorders of childhood and adolescence currently affect 1 in 7 youths in the United States and globally. Externalizing disorders such as attention-deficit / hyperactivity disorder (ADHD), and internalizing disorders such as anxiety, are among the most prevalent and represent a wide spectrum of dysfunctional behavior patterns. Treatment barriers are complex and multifaceted but major contributors include our limited understanding of psychiatric phenotypes and difficulty identifying youth individuals that experience these disorders.

[0004] Traditionally, psychiatric disorders have been conceptualized as categorical macrophenotypes based on clinical manifestations of a disease, which are defined according to the number and type of symptoms and the presence of distress or impairment. While this has practical benefits in terms of reliability and ease of diagnosis, it poses several challenges to the research of these disorders and, consequently, to the development of treatments. Furthermore, given the high heritability of psychiatric disorders, dissecting their underlying genetic architecture is of interest to researchers. While cost-effective and accurate genotyping technologies in large cohorts of individuals have significantly advanced the field, barriers associated with missing heritability and the need for improved phenotyping strategies are stillpresent. In fact, many psychiatric genome-wide association studies (GWASs) to date rely on subjective, dichotomized (i.e., binary) traits. However, psychiatric disorders are complex and often comorbid, and this high degree of heterogeneity is not always accurately translated into categorical diagnostic labels, which may be defined by arbitrary cut-offs.

[0005] Accordingly, there remains a need in the art for articles and methods that improve upon existing articles and methods for detecting and treating psychiatric disorders. The present disclosure meets this need.SUMMARY

[0006] In one aspect, a method of training an artificial intelligence (Al) model to generate digital phenotypes includes providing a training set including digital descriptors and training the Al model to generate the digital phenotype using the training set, wherein the training is configured to relate the digital descriptors to a health condition to generate the digital phenotype.

[0007] In some embodiments, the digital descriptors are generated from digital information recorded from one or more sensors. In some embodiments, the one or more sensors include at least one of a biosensor, a wearable sensor, or a combination thereof. In some embodiments, the wearable sensor comprises at least one of a smartwatch, a fitbit, an accelerometer, or combinations thereof.

[0008] In some embodiments, the digital information comprises data that reflects at least one of physical processes of the subject, physiological processes of the subject, inferred higher-order behavioral events of the subject, temporal dynamics, or combinations thereof. In some embodiments, the physical or physiological processes include at least one of movement, pulse, metabolic intake, or combinations thereof. In some embodiments, the higher-order behavioral events include at least one of sleep, exercise, or a combination thereof.

[0009] In some embodiments, the digital information further includes at least one of genotype data and demographica data.

[0010] In some embodiments, the digital information is recorded from control individuals and case individuals. In some embodiments, the control individuals are individuals without the health condition and the case individuals are individuals with the health condition.

[0011] In some embodiments, the method further includes generating the digital descriptors from the digital information. In some embodiments, the generating of the digital descriptors comprises feature engineering the digital information. In some embodiments, the feature engineering comprises at least one of generating time-invariant static features from the digital information and generating time-varying dynamic features from the digital information. In some embodiments, generating the time-invariant static features includes applying a statistical measure over an arbitrary time window to summarize the corresponding digital information. In some embodiments, generating the time-varying dynamic features includes preserving a timevarying nature of the corresponding digital information by processing the corresponding digital information using signal imputation and alignment, and then concatenating to form a multichannel time series.

[0012] In some embodiments, prior to the feature engineering, the method includes at least one of preprocessing the digital information, applying a quality control (QC) process to the digital information, imputing missing values in the digital information, or combinations thereof. In some embodiments, the digital information includes at least two modalities and the preprocessing includes combining the at least two modalities into a single matrix. In some embodiments, the QC process comprises filtering the length of a time series to be included in downstream analyses and identifying a time window that provides best alignment across individuals. In some embodiments, the QC process comprises filtering, from a sample pool, individuals that did not provide a minimum portion of valid measurement within an identified measurement window. In some embodiments, the imputing step comprises a two-fold imputation strategy. In some embodiments, the two-fold imputation strategy is employed for both categorical and quantitative data.

[0013] In some embodiments, the training comprises a dual-architecture approach employing machine learning for static features of the digital descriptors and deep learning for dynamic features of the digital descriptors. In some embodiments, the machine learning comprises XGBoost. In some embodiments, the deep learning comprises Xception.

[0014] In some embodiments, the health condition comprises a mental health condition, a neurocognitive disease, or a neurodegenerative disease. In some embodiments, the mental health condition comprises at least one of a psychiatric disorder, a mental disease, anxiety, attentiondeficit hyperactivity disorder (ADHD), or combinations thereof. In some embodiments, the neurodegenerative disease comprises Parkinson’s disease.

[0015] In some embodiments, the method further comprises predictive training of the Al model to form a predictive Al model, the predictive Al model configured to predict a health status of the subject based upon the digital phenotype. In some embodiments, the predictive training comprises providing a predictive Al training set including digital phenotypes and training the predictive Al model using the predictive training set. In some embodiments, providing the predictive Al training set includes creating the predictive Al training set. In some embodiments, creating the predictive Al training set includes generating the digital phenotype for each individual represented by the digital descriptors in the training set. In some embodiments, the predictive Al model is trained for disease prediction. In some embodiments, the predictive Al model is trained to predict disease progression. In some embodiments, the predictive Al model is trained to predict disease progression rate.

[0016] In another aspect, a method of determining a digital phenotype for a subject includes providing digital descriptors for the subject, inputting the digital descriptors to the Al model trained according to any of the embodiments disclosed herein; and receiving the digital phenotype of the subject as an output of the Al model based upon the digital descriptors.

[0017] In some embodiments, the method further comprises generating the digital descriptors for the subject. In some embodiments, the generating of the digital descriptors for the subject comprises providing digital information for the subject and feature engineering the digital information to generate the digital descriptors.

[0018] In another aspect, a method of identifying genetic associations includes applying the digital phenotype of a subject generated according to any of the embodiments disclosed herein as a univariate or multivariate response variable in a genome-wide association study.

[0019] In another aspect, a method of predicting a health status of a subject includes providing digital phenotype for the subject, inputting the digital phenotype to a predictive Al model trained according to any of the embodiments disclosed herein, and receiving the predicted health status of the subject as an output of the predictive Al model based upon the digital phenotype.BRIEF DESCRIPTION OF THE DRAWINGS

[0020] For a fuller understanding of the nature and desired objects of the present invention, reference is made to the following detailed description taken in conjunction with the accompanying drawing figures wherein like reference characters denote corresponding parts throughout the several views.

[0021] FIG. 1 shows a schematic illustrating the leveraging of clinical, digital, and genetic data of an ABCD cohort to improve characterization of psychiatric disorders. Panel A shows a framework schematic describing how digital phenotypes from wearable-derived data are leveraged to better understand the association between macrophenotype and genotype. The link between digital phenotype and macrophenotype serves as construct validity and aid in diagnostics. Wearable GWAS is performed through genotype-to-digital-phenotype association studies. Panel B shows that the Adolescent Brain Cognitive Development (ABCD) cohort contains 11,878 individuals spanning nine different categorical macrophenotypes based on clinical diagnosis from the Kiddie Schedule for Affective Disorders and Schizophrenia-5. A breakdown of the counts of each disorder is shown in the bottom bar graph, with anxiety disorder and ADHD being the most prevalent. “Bipolar” refers to bipolar or psychotic disorders. Panel C shows digital data from FitBit biosensors are collected for 5,339 individuals. The collected time series data are then processed into dynamic and static features, with information spanning various physiological and higher-order processes. Panel D shows that genetic data are collected by the ABCD consortium through Smokescreen genotyping array. Imputed genotypes are used for downstream GWAS analyses. The genotype arrays are subjected to best-practice processing and QC to ensure included individuals and SNPs are of high quality. PCA performed on 8,791 individuals and 157,556 genotyped SNPs reveals distinct ancestral clusters across the cohort and the inferred genotype principal components (PCs) are used as covariates in downstream analyses. (Details in FIGS. 30-33.)

[0022] FIG. 2 shows a schematic illustrating workflow for data processing, feature engineering, and model architecture. Panel A shows that ABCD cohort metadata including various demographic features, cognitive test scores, and clinical characteristics are used as covariates and represent the input features used in our baseline comparison model. Features shown in this plot correspond to the filtered set of individuals with wearable data. (Details in FIGS. 11 and 12A-L and Data S40-S42.). Panel B shows that digital data collected by wearablebiosensors are used to generate dynamic features after signal processing and imputation steps. Together with the processed covariates, these time series features represent the input features for the dynamic model. (Details in FIGS. 13-15.). Panel C shows that summary statistics applied to digital data collected by wearables are used to generate a total of 258 static features. In addition to the covariates, these are the input features used in the static model. The static model leverages the machine learning framework, XGBoost, for downstream tasks such as wearable combination score generation and classification. (Details in Data S46-S47 ). Panel D shows that hierarchical clustering of the static features yields seven distinct physiological clusters of wearable data. (Details in Data S48-S49 ). Panel E shows how the dynamic model is based on the Xception deep learning framework, and uses the generated 48 channels from the dynamic features and covariates as input into a convolution-like model. The architecture consists of six inception layers and residual connections. Global average pooling and a fully connected layer allow for similar downstream tasks as mentioned in Panel C.

[0023] FIG. 3 shows performance and interpretability of psychiatric phenotype classification models. Panels A-B show model performance for baseline, static, and dynamic models employed for classifying individuals with ADHD (blue, top) or individuals with anxiety disorder (purple, bottom) versus healthy controls. P values were calculated using one-sided t-test. (Details in FIGS. 19-22.). Panels C-D show feature importance based on ablation studies for the dynamic model for ADHD (blue, top) and anxiety disorder (purple, bottom) classification. Wearable-derived dynamic features are shown in red font and clinical features (covariates) are shown in black font. Feature importance is equivalent to the decrease in model performance (AUROC) after removal of the given feature. (Details in FIGS. 23-27.). Panels E-F show temporal importance during a 48-hour period for dynamic features in ADHD (blue, top) or anxiety disorder (purple, bottom) classification based on the GRAD-CAM interpretability module. Importance is represented as the GRAD-CAM score, based on each time point’s contribution towards model performance. (Details in FIGS. 28-29.)

[0024] FIGS. 4A-C show Manhattan plots summarizing the results of multivariate and univariate GWAS for ADHD. (A) Left panel: Schematic describing for a given SNP the frequency of healthy controls or individuals with ADHD for each genotype. Right panel:Resulting Manhattan plot from a case-control GWAS on 1,191 individuals from the ABCD cohort. We employed the clinical diagnosis label as the binary univariate response variable forthe GWAS (TIADHD = 137, ncontroi = 1,054). No genetic variants passed the genome-wide significance threshold (p value < 5 IO'8; blue line). Genetic variants with a suggestive p value (< 10'5) are represented as green dots. In all panels, proximal genes related to ADHD are highlighted in dark blue, and genes related to other psychiatric disorders are highlighted in pink (evidence obtained from OpenTargets). Brain-related traits associated with genetic variants overlapping the genome-wide significant loci are highlighted in orange. GWAS associations were obtained from the EBI-NHGRI GWAS catalog. A detailed list of genome-wide significant loci for all panels is provided in Table 1 and Table S2-S3. In this figure, we only show results related to autosomal chromosomes. (Details in FIGS. 48-49 and Data S56.). (B) Left panel: Schematic describing for a given SNP, the relationship between a multivariate set of n wearable- derived features (dependent / response variable) and an interaction term represented by the genotype and the disorder status of the individuals (independent / predictor variable). Right: Resulting Manhattan plot using clusters of wearable-derived features as the multivariate response variable in a GWAS that encodes the interaction term genotype disorder (where disorder is a binary feature such as 0 = Control, 1 = ADHD; gxm). The GWAS was performed on the same set of 1,191 individuals as in panel A. We identified 2 and 174 loci passing the p value thresholds of 5 • 10'8and 1 1 O'5, respectively. Locus chr6:53, 240, 429-53, 356, 412 is proximal to genes CILK1, ELOVL5, FBX09 (highlighted in dark blue), which have been associated with ADHD previously. The inset panel shows that individuals with ADHD exhibit different levels of residualized (i.e., covariate-adjusted) sedentary time (maximum) depending on the genotype at lead variant rsl 86003 (chr6:53,320,326). In contrast, healthy control individuals show no difference among genotype groups (*** p < 0.001, **: 0.001 <p < 0.01, *: 0.01 <p < 0.05, ns: p > 0.05; two-sided Wilcoxon Rank-Sum test) (Details in FIGS. 37-38 and Data S53). (C) Left: Schematic showing the relationship between the wearable combination score (dependent / response variable) and genotype (independent / predictor variable). Right: Resulting Manhattan plot using the wearable combination scores (trained on classification of individuals with ADHD) as the response variable in a GWAS for ADHD. The GWAS was performed on the same set of 1,191 individuals as in panel A. We identified 10 and 414 loci passing the p value thresholds of 5 IO'8and 1 10'5, respectively. Loci chrl: 111,372,165-111,482,359, chrl7:7, 101,607-7, 101,608, and chrl7:32, 256, 997-32, 283, 356 are proximal to genes ADORA3(72 Kb), DLG4 (86 Kb) and PSMD11 (174 Kb) (highlighted in dark blue) respectively, which have been previously associated with ADHD.

[0025] FIGS. 5A-C show images and graphs exploring the genetic-physiological-psychiatric axis with wearable GWAS. (A) Using the 258 wearable-derived static features as a continuous multivariate response variable, the GWAS was performed by pooling a set of 2,410 individuals (both healthy controls and individuals with any disorder). We identified 4 and 198 loci passing the p value thresholds of 5 • 10‘8and 1 10'5, respectively. A detailed list of genomewide significant loci is provided in Table 1 and Table S4. Neuropsychiatric-related genes proximal to the identified loci are highlighted in pink. Brain-, and heart-related traits with associated variants overlapping these 4 loci are highlighted in orange. (B) Left panel: rs365990 (chrl4:23,392,602, A / G) is located in exon 25 ofMYH6 and is associated with changes in wearable-derived heart rate features (multivariate GWAS p value = 5.33E-09). The boxplots show distributions of covariate-adjusted mean and interday coefficient of variation (CV) for heart rate across genotype groups at rs365990 (AA n individuals = 1,228; AG n individuals = 1,509; GG n individuals = 519). p values for each pairwise comparison are also displayed, encoded as follows: ***: p < 0.001, **: 0.001 <p < 0.01, *: 0,01 <p < 0.05, ns: p > 0.05 (two-sided Wilcoxon Rank-Sum test) For visualization purposes, outliers are not shown. Right panel: enrichment, displayed as odds-ratio (log2(OR); y-axis) of the minor allele (G) in individuals with different psychiatric disorders (x-axis) compared to healthy controls. OR estimates and 95% confidence interval (error bar) are displayed. The red horizontal dashed line indicates no enrichment. The G allele is significantly more enriched in individuals with bipolar / psychotic disorder compared to healthy controls (two-sided Fisher test p value: 8.00E-03; FDR-adjusted p value: 7.00E-02). (C) Similar representation for rsl 13525298 (chr7: 1,791,353; AAn individuals = 2,294; AG n individuals = 101; GGn individuals = 15). rsl 13525298 is located 125 Kb from ELFN1, a gene that encodes for a postsynaptic protein involved in the temporal dynamics of interneuron recruitment. Elfnl mutant mice exhibit hyperactivity that is treatable by psychostimulant medication. The G allele at rsl 13525298 is associated with increased minimum number of first-out-of-bed minutes and decreased minimum number of total-vigorously-active minutes (multivariate GWAS p value = 5.09E-09), and is significantly more enriched in healthy controls compared to individuals with ADHD (two-sided Fisher test p value: 9.00E-04; FDR-adjusted p value: 6.00E-03). (Details in FIGS. 43-46 and Data S55).

[0026] FIG. 6 shows a schematic illustrating an overview of the study design, according to an embodiment of the disclosure. Panel A shows a schematic describing potential causal relationships linking environment, genotype, macrophenotype (psychiatric diagnosis), and digital phenotype. The direction of the arrows signifies causal direction. Solid lines in the figure denote causal directions explored in this study, and dotted lines represent other possible directions of causality. Pathways with both solid and dotted lines may show bidirectional causal association (e.g. positive feedback loops). Panel B shows an example of raw wearable-derived dynamic data, showcasing time-series profiles of the seven channels collected over a selected period of 96 hours for three individuals of the AB CD cohort.

[0027] FIGS. 7A-B show schematics illustrating processing pipeline and deep learning architecture. (A) Flowchart of data preprocessing, imputation and covariates integration. First, seven modalities of data are combined and transformed into n data frames corresponding to n samples. An iterative examination is performed to find out all possible time windows with continuous time-series data avoiding noise or outliers, and the optimal time window is selected based on statistical power and data quality. Some samples are filtered out due to inadequate measurements compared to the defined threshold. During the data imputation process, missing categorical data is marked as ‘not recorded’ and missing quantitative data is imputed through sk-time drift method. To indicate whether data points at the current time step are imputed or not, a binary indicator channel, ‘*_flag’, is added to each modality. This channel enables the model to distinguish between imputed and non-imputed data, improving the model’s performance.Covariate data is also converted into time series format as time-invariant values. Finally, digital signature, imputation indicator and covariates are combined for modeling. (B) Deep learning architecture of XceptionTime Encoder. The architecture starts with Id-CNNs with various sizes of convolutional filters to create multi-level receptive fields. It is followed by MaxPooling to reduce dimensionality and Batch Normalization to normalize the activations. Finally, the ReLU activation functions are applied to add non-linearity.

[0028] FIGS. 8A-B show graphs illustrating model comparisons and loss curves. (A) Comparison of AUROC performance for predictive models of ADHD. The comparison includes XGBoost model (details in FIGS. 19-27), MiniRocket, a recurrent neural network (RNN) architecture, and the best performing Xception model. MiniRocket is a non-deep learning algorithm that uses convolutional filters (varying in weights, sizes, strides, dilations, andpaddings) along the original dataset to generate deterministic summary statistics of the transformed time-series. The summary statistics are then fed into a logistic regression model. (B) Training and validation loss over 30 epochs of the ADHD Xception model. The close alignment of training and validation loss demonstrates that the model generalizes well to unseen data and has a low risk of overfitting.

[0029] FIGS. 9A-C show schematics and graphs illustrating statistical details of the multivariate and univariate GWASs for ADHD. (A) Schematic outlining the types of association tests implemented in each GW AS. On the left (Case-Control GWAS), we use a traditional framework to test for associations between genetic variants (genotype, g) and ADHD diagnosis (macrophenotype, m while adjusting for covariates (c) Here, m is a binary univariate response variable (presence or absence of ADHD), and g and c serve as independent variables. On the right (Wearable GWAS), we perform two types of association tests. 1) Continuous & Multivariate Response: In this case, the response variable is a vector wearable-derived features, which are not directly related to ADHD. To evaluate the relevance of genetic associations to ADHD, we introduce an interaction term g x m, which allows us to assess genetic effects on wearable-derived features specifically in the context of ADHD. Here, the wearable feature vector t / ^serves as the continuous multivariate response, while g, m, g x m, and c serve as the independent variables. 2) Continuous & Univariate Response: In this case, the response variable is the wearable-derived combination score (dscore), which represents a non-linear combination of wearable features. This score is generated through our supervised modeling framework aimed at classifying individuals with and without ADHD, thus making dscore implicitly related to tn. (Details in FIGS. 36, 39-41, and Data S54). (B) Hierarchical Two-step Test of association implemented for the multivariate GWAS. Consider p wearable features (in our case p = 258) measured across n individuals. When testing for genetic associations with each independent feature (Feature-wise Test, left panel), the study -wide significance threshold is (5 - 10‘8 / p) (i.e., the genome-wide significance divided by the number of independent tests). However, it is possible to reduce this multiple testing correction burden by performing a Hierarchical Two-Step Test (right panel). In step 1, p features are grouped into Q clusters of correlated features, and we test the association between each multivariate cluster of features q and g. The correlated nature of the features within a particular cluster enhances association power. The study-wide significance threshold in this case becomes 5 - 1 O'81 Q. Given that Q «p, this strategy translates to an improvement of ~1.5 orders of magnitude in the stringency of the threshold, compared to the Feature-wise Test. Since Step 1 does not specify which individual features within the cluster drive the significant association with the variant, in Step 2 we perform post-hoc univariate tests to independently test each of the r features within the significant cluster and assess differences in a given wearable feature between individuals carrying different genotypes (two-sided Wilcoxon Rank-Sum test with Benjamini -Hochberg FDR correction). (Details in FIGS. 37-38 and 43-46 and Data S53 and S55). (C) Dot plot showing the power (y axis) of continuous and binary response variables for increasing genotype effects (beta coefficient; x axis). The type of response is color-coded to reflect the binary and continuous nature of macrophenotype and digital phenotype, respectively. Power is defined as the proportion (%) of linear (continuous) and logistic (binary) regression tests with BH-adjusted p value < 0.05 out of 10,000 simulations. We simulated SNP genotypes considering six different Minor Allele Frequency (MAF) values.

[0030] FIGS. 10A-D show a schematic and graphs dissecting genetic associations with behavioral traits. (A) Schematic describing the Wearable GWAS for behavioral traits, where the wearable feature vector d ^serves as the continuous multivariate response, and the genotype g and covariates c serve as the independent variables. (Details in FIGS. 42 and 47-49). (B) Enrichment of chromatin features at GWAS locus for heart rate features (chrl4:23,392,601-23,418,974). Right panel: Barplot showing the number of epigenomics ENCODE experiments (x axis) with peaks at the locus across different organs and tissues (y axis). Only organs I tissues with a Benj mini -Hochberg adjusted p value < 0.1 (two-sided Fisher test) are shown. The type of experiment is color-coded (purple: DNase-seq; blue: ATAC-seq; magenta: Mint ChlP-seq; orange: Histone ChlP-seq). Left panel: Dot plot showing the log2 odds-ratio (OR) of the significance of peak enrichment across different biosamples. (C-D) GTEx expression levels of proximal genes and eGenes associated with locus chrl4:23, 392, 601-23, 418, 974. Scatterplot showing the expression level (y axis, log2 scale) of proximal genes (upper panel) and GTEx eGenes (lower panel) across 54 GTEx tissues (x axis). In this case we show the two eGenes (CMTM5 and HOMEZ) that have GTEx eQTLs coinciding with four of the five GWAS variants within the locus. (Details in Table S4). Human tissues are color-coded based on the GTEx palette.

[0031] FIG. 11 shows a schematic illustrating a flowchart of the ABCD modeling framework. The framework incorporates data from two sources: raw wearable data generated by the Fitbit device and covariates data including information on demographics, cognitive scores, and behavioral checklists. After data preprocessing, the data are further processed using summary statistics to generate static wearable features. These are then clustered into seven categories. The wearable data, considered as the digital signature, is preprocessed as a timeseries with optimal window selection, missing data imputation, and covariate integration. This procedure creates the dynamic wearable feature-time series. The XGBoost model on static features and the Xception deep learning model on dynamic time-series features are used to generate wearable combination scores for classification. Models are interpreted by model performance, feature importance, and step importance plots.

[0032] FIGS. 12A-L shows graphs illustrating the covariates distribution of psychiatric disorders across 12 pairs of density plots. The left plots showcase eight different cognitive scores and their distribution across each disorder. The right plots show two kinds of CBCL scores and their distribution across each disorder. (A) Cognitive and CBCL scores distribution for all disorders. (B) Cognitive and CBCL scores distribution for severe disorder. (C) Cognitive and CBCL scores distribution for anxiety group. (D) Cognitive and CBCL scores distribution for depression group. (E) Cognitive and CBCL scores distribution for bipolar group. (F) Cognitive and CBCL scores distribution for eating group. (G) Cognitive and CBCL scores distribution for OCD group. (H) Cognitive and CBCL scores distribution for pstd group (I) Cognitive and CBCL scores distribution for sleep group. (J) Cognitive and CBCL scores distribution for ADHD group. (K) Cognitive and CBCL scores distribution for panic group. (L) Cognitive and CBCL scores distribution for nonclinical controls.

[0033] FIG. 13 shows a graph illustrating percentage of missing / recorded data. It reveals the presence of missingness across various features. The figure shows values from all individuals before QC, filtering, and processing.

[0034] FIG. 14 shows a graph illustrating sample size variation across non-missing data thresholds. The figure shows the relation between number of samples (y-axis) and thresholds for non-missing data (x-axis), for individuals with ADHD or anxiety, and healthy controls

[0035] FIG. 15 shows a heatmap illustrating correlation between wearable PCs and covariates. This heatmap represents the correlation matrix between the top 22 PCs and the 26 covariates in the dataset. Colors show the strength and direction of the correlation.

[0036] FIG. 16 shows box plots illustrating distribution of wear time of wearable Fitbit devices across control, ADHD, and anxiety groups. The Kruskal-Wallis test indicates no significant difference in wear times across the three groups shown in the figure above (p value: 0.546). These findings are consistent with the Mann-Whitney U test results revealed in Data S43, which also did not show any significant differences when each phenotype was compared to the control group individually. The consistency among different tests indicates that wear time does not affect the classification outcome by phenotype in the context of this study.

[0037] FIG. 17 shows graphs illustrating distribution of sports and activities participation in phenotype groups. (Top Panel) The stacked bar chart in the upper panel shows the percentage of the occurrence of each sport / activity across the three groups (control, anxiety disorder, and ADHD). The legend on the right shows the colors representing different sports and activities. (Bottom Panel) The lower panel features boxplots that depict the distribution of the number of sports / activities that individuals participate in across each of the three groups. The Kruskal-Wallis tests the number of sports / activities that individuals participated in across each of the three groups (p value: 0.513). These findings are consistent with the Mann- Whitney U test results revealed in Data S44, which did not show any significant differences when each phenotype was compared to the healthy control group.

[0038] FIG. 18 shows images illustrating ADHD medication analysis workflow. (Panel A) Pipeline for comparing wearable features across healthy controls, ADHD (no Rx), and ADHD (w / Rx) ADHD (no Rx) denotes a subset of individuals with ADHD not treated with medication and “ADHD (w / Rx)” denotes a subset of individuals with ADHD who were treated with medication (e.g., Adderall, Concerta, Vyvanse). P-values were determined using a Wilcoxon Signed-Rank Test. (Panel B) Comparison of resting heart rate across the three different groups. Healthy controls were significantly different when compared to either the “ADHD (w / Rx)” or “ADHD (no Rx)” groups (p=0.017 and p=0.00163, respectively). No significant difference was determined when comparing between “ADHD (w / Rx)” and “ADHD (no Rx)” groups. (Panel C) For each feature, a p-value was determined for each of the pairwise comparisons between groups. Overall, the most significant differences across all features werefound when comparing healthy controls to either the “ADHD (w / Rx)” or “ADHD (no Rx)” groups. Comparison of “ADHD (w / Rx)” or “ADHD (no Rx)” groups yields the least significance across all features.

[0039] FIG. 19 shows box plots illustrating the performance of different ADHD prediction models (with CBCL incorporated) in ten experiments. The performance metrics displayed are AUROC, Accuracy, Precision, Recall, and Fl Score. The explanation of the model index can be found in Data S52

[0040] FIG. 20 shows box plots illustrating the performance of different ADHD prediction models (without CBCL incorporated) in ten experiments. The performance metrics displayed are AUROC, Accuracy, Precision, Recall, and Fl Score. The explanation of the model index can be found in Data S52

[0041] FIG. 21 shows graphs illustrating a side-by-side comparison of model performance with covariate set 1 (displayed in Data S40) and set 2 for the classification of ADHD and Anxiety disorders. Compared to set 1, set 2 incorporates additional covariates: sports / activities participation (see also Data S42) and wear time. The boxplots display the distribution of AUROC across baseline, static, and dynamic models for each disorder. Adding new covariates does not produce a significant change in the performance of the models for both ADHD and anxiety.

[0042] FIG. 22 shows graphs illustrating the performance of different Anxiety prediction models in ten experiments. The architecture of the first two models is XGBoost while the remaining two are built on Xception, and all four models apply CBCL. The performance metrics displayed are AUROC, Accuracy, Precision, Recall, and Fl Score. The detailed explanation of each indexed model can be found in Data S52.

[0043] FIG. 23 shows a graph illustrating feature importance by showing the change in AUROC scores after removing each feature from the Xception model, with negative values denoting greater importance of a given feature. The features are ranked from left to right, in order of decreasing importance, including demographic information, family history, cognitive score, and time-series wearable data. The error bars show the standard error across multiple experiments.

[0044] FIG. 24 shows graphs illustrating ADHD - feature importance and grouped feature importance of XGBoost COV. The left plot shows the feature importance of covariatesincluding demographic information, family history and cognitive scores in the XGBoost model, with higher values indicating greater importance of a given feature. The features are arranged from left to right, in ascending order. The feature importance is computed based on how much each feature contributes to the overall prediction of the model. The error bars show the standard error across multiple experiments. The right plot is an aggregated version of the top one. The features are categorized into groups and their individual importance values are summed to quantify the importance of the corresponding group.

[0045] FIG. 25 shows graphs illustrating ADHD - feature importance and grouped feature importance of XGBoost COV + different wearable static feature clusters. The plots on the left show the importance, in the XGBoost model, of different covariates, including demographic information, family history, and cognitive scores, as well as the importance of features from each of the seven clusters of wearable-derived static features. Higher values indicate greater importance of a given feature towards model performance. The features are ranked from left to right in ascending order. The number after ‘C’ in superscript in the upper right corner of the letter ‘W’ indicates the cluster of wearable features, and the subscript ‘Static’ in the lower right corner indicates static features. Feature importance is computed based on how much each feature contributes to the overall prediction of the model. The error bars show the standard error across multiple iterations. The plots on the right are aggregated versions of the left plots. The features are categorized into groups and their individual importance values are summed to quantify the importance of the corresponding group.

[0046] FIG. 26 shows graphs illustrating ADHD - feature importance and grouped feature importance of XGBoost COV + all wearable static feature cluster. The first two plots show the importance, in the XGBoost model, of different covariates, including demographic information, family history, and cognitive scores, as well as the importance of features from each of the seven clusters of wearable-derived static features. Higher values indicate greater importance of a given feature towards model performance. The features are ranked from left to right in ascending order. Feature importance is computed based on how much each feature contributes to the overall prediction of the model. The error bars show the standard error across multiple iterations. (Due to the larger number of features, the feature importance is plotted into 2 figures, where the first figure shows top 50% features and the second contains the remaining.). The last plot is an aggregated version of the two upper plots. The features are categorized intogroups and their individual importance values are summed to quantify the importance of the corresponding group.

[0047] FIG. 27 shows a graph illustrating anxiety feature importance by showing the change in AUROC scores after removing each feature from the Xception model, with negative values denoting greater importance. The features are ranked from left to right, in decreasing order, and do not include CBCL data. The error bars show the standard error across multiple experiments.

[0048] FIG. 28 shows a graph illustrating the importance of time ‘step’ intervals (with each step being a 1-hour time interval in a consecutive 48-hour period) in relation to the change in AUROC score observed upon their removal from the ADHD model. The y-axis shows the change in AUROC score after a time interval is excluded, and the x-axis are intervals.

[0049] FIG. 29 shows a graph illustrating the importance of time ‘step’ intervals (with each step being a 1-hour time interval in a consecutive 48-hour period) in relation to the change in AUROC score observed upon their removal from the Anxiety model. The y-axis shows the change in AUROC score after a time interval is excluded, and the x-axis are consecutive time intervals.

[0050] FIG. 30 shows graphs illustrating ancestry composition of ethnicity score groups Distributions of proportions (y axis) for different ancestries (AF: African; AM: American; EA: East Asian; EU: European; x axis) across five ethnicity score groups (1-5). We obtained ethnicity score groups and ancestry proportions for genotyped individuals from the ABCD consortium metadata. Based on these plots, we assigned the following labels to the five ethnicity score groups: European (group 1); African (group 2); European / American (group 3), European / East Asian (group 4); Mixed (group 5).

[0051] FIG. 31 shows a graph illustrating frequency of ethnicity score groups. The barplot shows the number of individuals (y axis) that belong to each of the five ethnicity score groups (x axis).

[0052] FIG. 32 shows graphs illustrating ancestry decomposition of genotype principal component analysis. Analogous representation to FIG. 1 (Panel D) where the proportion of the four different ancestries are color-coded for each individual.

[0053] FIG. 33 shows graphs illustrating number of genotyped and imputed genetic variants. Barplot showing the number of variants (y axis) across the 23 chromosomes (x axis) that we included in the GWAS analyses.

[0054] FIG. 34 shows histograms illustrating the distribution of estimated imputation accuracy (R2) for all genetic variants (genotyped & imputed) across the 23 chromosomes.

[0055] FIG. 35 shows Histograms showing the distribution of empirical R2(ER2) values for genotyped variants across the 23 chromosomes.

[0056] FIG. 36 shows a quantile-quantile (QQ) plot of the observed -logio p values (y axis) from the continuous multivariate GWAS for ADHD, against the expected -logio p values (x axis). The red line indicates the distribution of p values under the null hypothesis. The QQ plots are shown only for the GWAS runs that reported genome-wide significant loci. For each run the genomic inflation factor was also computed. As the multivariate GWAS p values do not follow a normal distribution, Ax was employed (instead of the commonly used AG) to estimate the genomic inflation factor (see Methods). The Axvalues are 1.18 and 1.16 for Cluster 3 and 5, respectively. Although these values are larger than the recommended threshold (0.95 < A < 1.05), it was found that most of the inflation was contributed by variants with p value > 10'2. As a result, when only considering variants with p value < 10'2, smaller Axvalues were obtained(l.16 and 1.05 for Cluster 3 and 5, respectively). Note that these QQ plots correspond to variants located on autosomal chromosomes. For variants located on chromosome X, see FIGS. 48-49.

[0057] FIG. 37 shows boxplots illustrating the distributions of Cluster 3 wearable-derived static features across genotype: disease groups at the lead variant for locuschrl: 103,922,953-104,046,796 (rs75092661). The number of individuals in each genotype:disease group is: 116 (GG: ADHD), 20 (GA: ADHD), 821 (GG:ctrl), 219 (GA:ctrl), 14 (AA:ctrl). As > 10 individuals are required in each genotype:disease group, the set of individuals with ADHD and genotype AA (i.e., AA: ADHD) was not included (n = 1). For each feature within Cluster 3 a post-hoc univariate test was conducted comparing the feature distributions between each genotype group within the set of individuals with ADHD as well as for control individuals. P values for each pairwise comparison are also displayed, encoded as follows: ***: p < 0.001, **: 0.001 <p < 0.01, * 0.01 <p < 0.05, ns: p > 0.05 (two-sided Wilcoxon Rank-Sum test). A Benjamini -Hochberg False Discovery Rate (FDR) correction was applied on these p values and reported only those features with FDR < 0.1 in at least one pairwise comparisonamong the genotype groups of individuals with ADHD. A detailed list of the significant features for this locus is provided in Data S53 (for a description of each feature see Data S45-S46). For visualization purposes, outliers are not shown.

[0058] FIG. 38 shows boxplots illustrating the distributions of Cluster 5 wearable-derived static features across genotype: disease groups at the lead variant for locuschr6:53, 240, 429-53, 356, 412 (rsl 86003). The number of individuals in each genotype: disease group is: 33 (AA: ADHD), 71 (AC: ADHD), 33 (CC: ADHD), 325 (AA:ctrl), 490 (ACctrl), 239 (CC:ctrl). A detailed list of the significant features for this locus is provided in Data S53 (for a description of each feature see Data S45-S46). This is an analogous representation to FIG. 37 for locus chr6:53, 240, 429-53, 356, 412.

[0059] FIG. 39 shows barplot illustrating the number of loci (x axis) identified by the ten univariate continuous GWASs (y axis) for ADHD. The number of loci identified at the conventional genome- wide significant threshold (p value < 5 - 10'8) as well as suggestive associations (p value < 1 • 10'5) are reported. Note that these results correspond to variants located on autosomal chromosomes. For variants located on chromosome X, see FIGS. 48-49.

[0060] FIG. 40 shows a quantile-quantile (QQ) plot of the observed -logio p values (y axis) from the continuous univariate GWAS for ADHD, against the expected -logio p values (x axis). The red line indicates the distribution of p values under the null hypothesis. The QQ plots are shown only for the GWAS runs that reported genome-wide significant loci. For each run the genomic inflation factor (AG) was also computed. From left to right, the A. G values are: 0.98 (XGB + CBCL v2), 0.98 (XGB + CBCL), 0.99 (Xception v2), 1.01 (Xception), 0.97 (CBCL ext). Note that these QQ plots correspond to variants located on autosomal chromosomes. For variants located on chromosome X, see FIGS. 48-49.

[0061] FIG. 41 shows a graph illustrating enrichment of EBI-NHGRI GWAS catalog traits on the set of significant loci from the continuous univariate GWAS for ADHD. For each brain-related trait (description of EFO term; x axis), the proportion of SNPs from the EBI-NHGRI GWAS catalog (y axis) that overlap the univariate continuous GWAS hits (colored dot) are reported. The color of the dot corresponds to the percentile of the observed enrichment with respect to the null distribution (gray violin plot). T-tau: A-Beta: t-tau:b eta-amyloid 1-42 ratio measurement. Cortical SA: Cortical Surface Area. Brain msrmnt.: Brain Measurement.

[0062] FIG. 42 shows a quantile-quantile (QQ) plot of the observed -logio p values (y axis) from the continuous multivariate GWAS for behavioral traits, against the expected -logio p values (x axis). The red line indicates the distribution of p values under the null hypothesis. The QQ plots are shown only for the GWAS runs that reported genome-wide significant loci. For each run the genomic inflation factor was also computed. As the multivariate GWAS p values do not follow a normal distribution, Ax was employed (instead of the commonly used G) to estimate the genomic inflation factor. From left to right, the Ax values are: 1.02 (HR-related Cluster), 1.02 (Cluster 1), 1.02 (Cluster 2). Note that these QQ plots correspond to variants located on autosomal chromosomes. For variants located on chromosome X, see FIGS. 48-49.

[0063] FIG. 43 shows boxplots illustrating the distributions of Heart Rate (HR)-related wearable-derived static features across genotype groups at the lead variant for locus chrl4:23, 392, 601-23, 418, 974 (rs365990). The number of individuals in each genotype group is: 1,228 (AA), 1,509 (AG), 519 (GG). For each feature within the cluster of HR-related features a post-hoc univariate test comparing the feature distributions between each genotype group was conducted, p values for each pairwise comparison are also displayed, encoded as follows: ***: p < 0.001, **: 0.001 <p < 0.01, *: 0.01 <p < 0.05, ns: p > 0.05 (two-sided Wilcoxon Rank-Sum test). A Benjamini -Hochberg False Discovery Rate (FDR) correction was applied on these p values and only those features with FDR < 0.1 in at least one pairwise comparison were reported. A detailed list of the significant features for this locus is provided in Data S55 (for a description of each feature see Data S45-S46). For visualization purposes, outliers are not shown.

[0064] FIG. 44 shows boxplots illustrating the distributions of Cluster 1 wearable-derived static features across genotype groups at the lead variant for locus chr7: 1,789,321-1,791,353 (rsl 13525298. The number of individuals in each genotype group is: 2,294 (AA), 101 (AG), 15 (GG). A detailed list of the significant features for this locus is provided in Data S55 (for a description of each feature see Data S45-46). This is an analogous representation to FIG. 43 for locus chr7: 1,789,321-1,791,353.

[0065] FIG. 45 shows boxplots illustrating the distributions of Cluster 1 wearable- derived static features across genotype groups at the lead variant for locus chrl 6:79,283,253- 79,302,474 (rs8051625. The number of individuals in each genotype group is: 2,267 (GG), 132 (GA), 11 (AA). A detailed list of the significant features for this locus is provided in Data S55(for a description of each feature see Data S45-46). This is an analogous representation to FIG.43 for locus chrl6:79, 283, 253-79, 302, 474.

[0066] FIG. 46 shows boxplots illustrating the distributions of Cluster 2 wearable-derived static features across genotype groups at the lead variant for locus chr6: 1,102,252-1, 130,154 (rs742521 ). The number of individuals in each genotype group is: 1915 (AA), 463 (AG), 32 (GG). A detailed list of the significant features for this locus is provided in Data S55 (for a description of each feature see Data S45-46). This is an analogous representation to FIG. 43 for locus chr6: 1,102,252-1,130,154.

[0067] FIG. 47 shows plots illustrating enrichment of EBI-NHGRI GWAS catalog traits on the set of significant loci from the multivariate GWAS for behavioral traits. For each trait (description of EFO term; x axis) the proportion of SNPs from the EBI-NHGRI GWAS catalog (y axis) that overlap the multivariate continuous GWAS hits (colored dot) is reported. Terms are grouped into brain / neuropsychiatric-related traits and heart-related traits. The color of the dot corresponds to the percentile of the observed enrichment with respect to the null distribution (gray violin plot), msrmnt.: measurement.

[0068] FIG. 48 shows resulting Manhattan plot from the sex-stratified GW AS involving genetic variants on chromosome X. Variants with genome-wide significant (< 5 - 10‘8) and suggestive (< 10'5) p values are represented as blue and green dots, respectively. Three genomewide significant loci are reported from the multivariate continuous GWAS for behavioral traits conducted on the male cohort (see Data S56).

[0069] FIG. 49 shows a quantile-quantile (QQ) plot of the observed -logio p values (y axis) from the multivariate GWAS for behavioral traits against the expected -logio p values (x axis) for variants located on chromosome X. The red line indicates the distribution of p values under the null hypothesis. The QQ plots are shown only for the GWAS runs that reported genome-wide significant loci (see FIG. 48 and Data S56). Specifically, these correspond to GWAS runs for Cluster 1 and 7 conducted on the male cohort. For each run the genomic inflation factor Xx (Cluster 1: 1.01; Cluster 7: 1.13) was also computed.

[0070] FIG. 50 shows graphs illustrating box plot representing the AUROC / performance metric of baseline, intermediate, and full models trained on healthy control vs Parkinson’s (left), or prodromal vs diagnosed (right). Baseline model represents a machine learning model with no digital / wearable data. Intermediate model includes average acceleration from the device. Fullmodel includes all of the digital descriptors / features we have generated in feature engineering stage. Here we can see that using the full set of digital descriptors / features yields the best result.

[0071] FIG. 51 shows a graph illustrating selection of optimal k with simulated missingness.DETAILED DESCRIPTIONDefinitions

[0072] As used herein, each of the following terms has the meaning associated with it in this section. Unless defined otherwise, all technical and scientific terms used herein generally have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. Generally, the nomenclature used herein and the laboratory procedures in molecular biology, immunology, animal pharmacology, pharmaceutical science, peptide chemistry, and organic chemistry are those well-known and commonly employed in the art. It should be understood that the order of steps or order for performing certain actions is immaterial, so long as the present teachings remain operable Any use of section headings is intended to aid reading of the document and is not to be interpreted as limiting; information that is relevant to a section heading may occur within or outside of that particular section. All publications, patents, and patent documents referred to in this document are incorporated by reference herein in their entirety, as though individually incorporated by reference.

[0073] In the application, where an element or component is said to be included in and / or selected from a list of recited elements or components, it should be understood that the element or component can be any one of the recited elements or components and can be selected from a group consisting of two or more of the recited elements or components.

[0074] In the methods described herein, the acts can be carried out in any order, except when a temporal or operational sequence is explicitly recited. Furthermore, specified acts can be carried out concurrently unless explicit claim language recites that they be carried out separately. For example, a claimed act of doing X and a claimed act of doing Y can be conducted simultaneously within a single operation, and the resulting process will fall within the literal scope of the claimed process.

[0075] As used herein, the singular form “a,” “an,” and “the” include plural references unless the context clearly dictates otherwise.

[0076] Unless specifically stated or obvious from context, as used herein, the term “about” is understood as within a range of normal tolerance in the art, for example within 2 standard deviations of the mean. “About” can be understood as within 10%, 9%, 8%, 7%, 6%, 5%, 4%, 3%, 2%, 1%, 0.5%, 0.1%, 0.05%, or 0.01% of the stated value. Unless otherwise clear from context, all numerical values provided herein are modified by the term about.

[0077] As used herein, the terms “comprises,” “comprising,” “containing,” “having,” and the like can have the meaning ascribed to them in U. S. patent law and can mean “includes,” “including,” and the like.

[0078] Unless specifically stated or obvious from context, the term “or,” as used herein, is understood to be inclusive.

[0079] Ranges provided herein are understood to be shorthand for all of the values within the range. For example, a range of 1 to 50 is understood to include any number, combination of numbers, or sub-range from the group consisting 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, or 50 (as well as fractions thereof unless the context clearly dictates otherwise).

[0080] As used herein, the term “ratio” refers to a relationship between two numbers (e.g., scores, summations, and the like). Although, ratios can be expressed in a particular order e.g., a to b or a.b), one of ordinary skill in the art will recognize that the underlying relationship between the numbers can be expressed in any order without losing the significance of the underlying relationship, although observation and correlation of trends based on the ration may need to be reversed. For example, if the values of a over time are (4, 10) and the values of b over time are (2, 4), the ratio a.b will equal (2, 2.5), while the ratio b.a will be (0.5, 0.4).Although the values of a and b are the same in both ratios, the ratios a.b and b.a are inverse and increase and decrease, respectively, over the time period.Detailed Description

[0081] Provided herein are systems, articles, and methods for generating digital phenotypes In some embodiments, the method includes receiving digital information for a subject from one or more sensors, determining one or more digital descriptors of the subject based upon the digital information, and determining a digital phenotype of the subject based upon the digital descriptors. The one or more sensors include any suitable sensor or mechanismfor detecting, measuring, and / or receiving the digital information relating to the subject. Such sensors include, but are not limited to, biosensors, wearable sensors, non-invasive biosensing mechanisms or devices, and / or combinations thereof. For example, in some embodiments, the one or more sensors include a wearable non-invasive biosensing technology, such as, but not limited to, a smartwatch, a fitbit, an accelerometer, any other suitable sensor, or a combination thereof.

[0082] The one or more sensors can be selected and / or configured to measure and / or receive any suitable digital information relating to the subject. Suitable digital information includes, but is not limited to, data that reflects physical processes, physiological processes, and / or can be used to infer higher-order behavioral events and their temporal dynamics. For example, in some embodiments, the digital information includes data relating to physical and / or physiological processes, such as, but not limited to, movement, pulse, metabolic intake, biochemical amounts (e.g., glucose from CGM, heart rate variability (HRV), VO2, (skin) temperature, CO2 of surrounding environment, weight, any other suitable digital information, and / or combinations thereof. Additionally or alternatively, in some embodiments, the digital information includes data relating to inferred higher-order behavioral events, such as, but not limited to, sleep, exercise, activity, sleep quality / staging, focus / concentration level, food / alcohol intake, shortness of breath, temporal dynamics thereof, any other suitable higher-order behavioral events, and / or combinations thereof.Feature Engineering to Generate Digital Descriptors

[0083] The digital descriptors described herein are generated from the measured and / or received digital information, and include any suitable digital descriptors of the physiological status, behavioral status, sports / activity participation, fitness, mood, mental state, and / or other digital descriptor of the subject. In some embodiments, the digital descriptors include, but are not limited to, aggregated (static) and / or time-resolved (dynamic) digital descriptors. For example, in some embodiments, the digital descriptors include an array of both aggregated (static) and time-resolved (dynamic) digital descriptors of the physiological and behavioral status of the wearer.

[0084] In some embodiments, the digital descriptors are generated through feature engineering of input data. In such embodiments, the input data includes measured and / or raw data from the one or more sensors. This input data may be gathered and / or provided from anysuitable source, such as, but not limited to, the subject, one or more control individuals, one or more case indivuals, or a combination thereof.

[0085] In some embodiments, the input data is preprocessed and / or subject to a quality control (QC) process prior to implementing the feature engineering. In some embodiments, the preprocessingincludes combining two or more different data modalities collected by the one or more sensors for each individual into a single matrix. In some embodiments, the preprocessing includes combining different data modalities collected by the wearable device for each individual into a single matrix. Additionally or alternatively, in some embodiments, the QC process is implemented to address any sparsity and / or missingness in the input data. In some such embodiments, the quality control process includes first filtering the length of the time series to be included in downstream analyses. Next, the time window that offered the best alignment across individuals is identified - that is, the period that provided the least amount of missingness across the greatest number of individuals. This process facilitates finding of the preferred time window, considering both the power of the data captured and the need for high-quality inputs. In some embodiments, the method further includes setting a criterion stipulating that each day has to contain a minimum portion of valid measurements within the previously identified window. For example, in some embodiments, the minimum portion includes no less than 60% of the maximum number of measurements. In such embodiments, any individual or portion of measurement failing to meet this requirement may be filtered out from the input data.

[0086] In some embodiments, the method includes imputing missing values in the input data prior to implementing the feature engineering. For example, in some embodiments, to address missing values in the QC-controlled time windows, a two-fold imputation strategy is employed for both categorical and quantitative data. In some such embodiments, the imputation strategy for both categorical and quantitative data includes introducing a 'Not Recorded' category and using the 'drift' method from the sktime package, respectively. Additionally or alternatively, in some embodiments, to account for potential imputation inaccuracies, an indicator time series is concatenated to the imputed data, explicitly describing which time point was imputed and which was directly measured by the device. When used for training of an artificial intelligence (Al) model, this permits the model to adaptively learn how to treat imputed points during training based on the surrounding, non-imputed data.

[0087] Additionally or alternatively, in some embodiments, imputing missing data includes applying a ^-Nearest Neighbors (KNN) imputation approach. In some embodiments, the KNN is refined through simulation-based tuning. For example, in some such embodiments, the KNN imputation approach includes first standardizing all activity features using z-score normalization (e.g., via the StandardScaler from scikit-learn), which centers each feature to mean = 0 and scales to unit variance while automatically ignoring missing values. Next, to determine the optimal number of neighbors (k) for imputation, additional missingness is simulated on the scaled data. In some embodiments, simulating the additional missingness includes randomly masking any suitable amount of the originally non-missing entries as missing, while preserving the original missing structure. Suitable amounts include, but are not limited to, at least 5%, at least 10%, at least 20%, at least 30%, between 10% and 30%, 20%, or any suitable combination, sub-combination, range, or sub-range thereof. For each A; in a defined range (e.g., 1-10), a KNNImputer is applied to the simulated dataset. The imputed values are then compared against the true (pre-masked) values at the simulated missing positions using mean squared error (MSE) as the evaluation metric. Based upon this, the k value yielding the lowest MSE is selected as the parameter for imputation. Using the selected parameter, a final imputation is performed on the scaled dataset using the corresponding KNNImputer. The imputed features are then inverse-transformed back to the original measurement scale to facilitate interpretability.

[0088] Following any preprocessing, quality control, and / or imputing, feature engineering is performed on the input data to generate the digital descriptors. In some embodiments, two different feature engineering strategies are performed to generate static and dynamic features. Static features are time-invariant thanks to the use of summarization, which permits production of straightforward and efficient features that are commonly used in downstream modelling. In some such embodiments, descriptive statistics are applied to expand on summary-based features collected by the one or more sensors, generating an expanded total of static features for each individual. The data based features include any statistical measure / metric that can be applied over an arbitrary and / or fixed time window in order to aggregate or summarize the information. For example, in one embodiment, descriptive statistics are applied to expand on 75 summary-based features collected by the wearable device, generating a total of 258 static features for each individual. In contrast to static features, dynamic features preserve the time-varying nature of the original digital signatures, retaining sequential and temporal patternsof the data. As such, in some embodiments, to better account for the temporal properties of these digital signatures, dynamic features are also generated for each individual. In some embodiments, generating the dynamic features include any time-resolved or time series based (dynamic) features of any amount of channels that are collected from sensor data. In some embodiments, the dynamic features are processed using signal imputation and alignment, and then concatenated to form the multichannel time series. For example, in one embodiment, the dynamic features include 48 time-series channels taking into account signal imputation and processing.

[0089] In some embodiments, implementing the feature engineering described herein maximizes the utility of the input data (e.g., wearable-derived data) for downstream analyses. While existing methods have relied on wearable-derived “summary” or static features to enable a variety of predictions of health states, human health is inherently temporal and dynamic, influenced by numerous factors like environment, diet, and stress. However, generating wearable-derived time-resolved digital phenotypes that can capture the temporal and sequential changes in data and reflect real-time variations and patterns poses additional challenges, as wearable data is often collected at different times and for different durations across individuals and hence may contain missing information. The methods disclosed herein ensure a harmonized and consistent set of measurements by producing robust, unbiased dynamic features.Additionally, the methods disclosed herein generate data that is compatible with different downstream research tasks, such as predictive modelling and genetic studies, complementing the insights gained from static features.Digital Phenotype Model

[0090] In some embodiments, the systems, articles, and methods include an artificial intelligence (Al) model configured to generate a digital phenotype based upon the digital descriptors. In some embodiments, the methods include training the Al model to generate the digital phenotype based upon the digital descriptors. In some such embodiments, training the Al model to generate the digital phenotype includes the steps of providing a training set including digital descriptors and training the Al model using the training set.

[0091] The training set includes digital descriptors generated from any suitable input data for training the Al model. In some embodiments, the digital descriptors of the training set are generated based upon input data from control individuals and case individuals. In someembodiments, the input data includes digital information from the control individuals and the case individuals. In some embodiments, the input data also includes one or more of genotype data and demographica data (e.g., sex, age, BMI, smoking status, socioeconomic status, etc.).

[0092] In some embodiments, control individuals include healthy individuals that do not have a condition of interest. In some embodiments, case individuals include individuals with a condition of interest. In some embodiments, the case individuals may be further segmented based upon the condition of interest. For example, when the condition of interest is Parkinson’s disease, the case individuals may be segmented into prodromal cases (e.g., where data was collected before diagnosis) and diagnosed cases (e.g., where data was collected after diagnosis).

[0093] In some embodiments, providing the training set includes selecting a previously created training set. Alternatively, in some embodiments, providing the training set includes creating the training set. In such embodiments, creating the training set includes collecting and / or receiving input data from control and case individuals, and feature engineering the input data as described herein to generate the digital descriptors.

[0094] In some embodiments, the preprocessing in connection with the feature engineering includes filtering input data to select desired control and case individuals. For example, the preprocessing can include filtering the input data to exclude individuals with missing data that cannot be imputed, exclude individuals with a prior diagnosis of a condition, or any other suitable element for filtering the individuals.

[0095] Additionally or alternatively, in some embodiments, creating the training set includes inferring activity features based upon the input data. For example, in some embodiments, suitable activities inferred from accelerometer digital information include, but are not limited to, raw acceleration (milligravity) and activities inferred from acceleration, such as, but not limited to, metabolic equivalent of task (MET; kcal / kg / hour), bicycling, mixed activities, sitting / standing, sleep, vehicle time, walking, and device wear time (proportion of time per hour). Although described above primarily with respect to inferring certain activities from acceleration data, as will be appreciated by those skilled in the art, the disclosure is not so limited and expressly includes any other suitable activity that can be inferred from an accelerometer and / or any other suitable wearable device. As will further be understood by those skilled in the art, some activities will have more relevance to certain conditions of interest. As such, in someembodiments, the method further includes selecting one or more activities to infer from the collected data based upon the condition of interest.

[0096] After providing the training set, the Al model is trained to generate the digital phenotype by relating the digital descriptors to any suitable health status / condition of interest. Suitable health statuses / conditions of interest include any disease or health condition that can be analysed in terms of genotype, macrophenotype, and digital phenotype. In some embodiments, the disease or health condition includes any disease or health condition for which the digital descriptors described herein can be related to. In some embodiments, the disease or health condition is a mental health condition, such as, but not limited to, a psychiatric disorder, a mental disease, anxiety, ADHD, any other suitable mental health condition, or a combination thereof. For example, in some embodiments, the Al model is configured to classify individuals with and without two neuropsychiatric conditions, anxiety disorder and attention-deficit / hyperactivity disorder (ADHD), using input data from a cohort of US adolescents (e.g., Adolescent Brain Cognitive Development or ABCD Project). Additionally or alternatively, in some embodiments, the Al model is configured to classify individuals with and without neurocognitive and / or neurodegenerative diseases. In one embodiment, for example, the neurodegenerative disease includes Parkinson’s disease.

[0097] The Al model can be trained in any suitable manner for relating the digital descriptors to the health status / condition of interest in order to generate the digital phenotype. For example, in some embodiments, to effectively handle the complex time-series nature of digital phenotype features, the Al model is trained through a dual-architecture approach that employs machine learning for static features and deep learning for dynamic features.

[0098] The machine learning includes any classical machine learning model, such as, but not limited to, XGBoost. For example, in some embodiments, the phenotype classification includes a canonical machine learning task using the digital descriptors. In such embodiments, the macrophenotype of subjects is predicted by inputting a set of features into a model, such as the XGBoostRegressor. XGBoost refines model accuracy through successive iterations, and its strengths can be leveraged through curated features. In one embodiment, for example, the digital descriptors for the XGBoost model includes 258 time-invariant wearable features derived from summary statistics, organized into seven clusters, supplemented by demographic, psychiatric,cognitive, and behavioral covariates. This feature set enables a nuanced analysis of the relationship between wearable data and individual covariates.

[0099] The deep learning includes any deep learning method tailored to time series modeling and capable of accounting for richer / temporal data contexts. In some embodiments, the deep learning method Xception is used for multichannel time series classification (MCTSC). Specifically, an individual’s macrophenotype can be predicted based on multichannel wearable time series data Cross-entropy loss with label smoothing can be used to optimize the model, enhancing its generalization by preventing overconfidence in predictions. Following training, the digital phenotypes generated by the Al model can be further used in a variety of downstream applications (e.g., diagnostic, improving genetic study, etc.).Predictive ModelIn some embodiments, the Al model is further trained to predict a health status of a subject, forming an Al predictive model. In some embodiments, the predictive model is trained to predict the health status of the subject utilizing the digital phenotypes generated according to one or more of the embodiments disclosed herein For example, in some embodiments, training the Al predictive model includes providing input data from healthy individuals and individuals representative of a health status of interest, generating digital phenotypes for each of the individuals in the input data, creating a training set based upon the digital phenotypes, and training the Al predictive model using the training set.

[0100] The Al predictive model can be configured for any suitable application depending upon the input data collected and / or the condition being assessed. For example, in some embodiments, a first construct includes a predictive model configured for disease prediction (e.g., healthy vs. sick) and risk assessment for various health conditions. In some embodiments, a second construct includes a predictive model configured for disease progression (prodromal-diagnosed). In some embodiments, a third construct includes a predictive model configured for disease progression rate / velocity. In some embodiments, the predictive model is specifically trained according to the health status being assessed / predicted. For example, in some embodiments, the predictive model is trained differently for each of the first construct, the second construct, and the third construct.First Construct

[0101] In some embodiments, such as in the first construct described above, the predictive model is trained to output a score relating to a health condition of the subject. In some embodiments, the method includes determining whether the subject has a health condition based upon the score generated from the predictive model. For example, in some embodiments, the predictive model is configured to generate a wearable combination score which integrates the digital features using non-linear models to predict a macrophenotype (e.g., summarize the likelihood that an individual has a given health condition). Additionally or alternatively, in some embodiments, the predictive model is configured to generate a severity score (< >) for assessing disorder severity. In some embodiments, the score is generated by extracting the final layer of the deep learning model. In some embodiments, the final layer of the deep learning model includes softmax probability, which returns probabilities for each category in a multi-class problem that sum up to 1.

[0102] In some embodiments, training the predictive model for the first construct includes creating a 1:1 age-matched case-control dataset, which minimizes the confounding effects of age. For example, in some embodiments, age frequency matching includes using 5-year age bins to ensure comparable age distributions between cases and controls. In such embodiments, for each age bin, the minimum sample size between cases and controls is determined, and an equal number of participants are randomly sampled from each group using a fixed random seed to ensure reproducibility. The matched subsets across all age bins are then concatenated to form the final 1: 1 age-matched case-control dataset.

[0103] In some embodiments, training the predictive model for the first construct includes model parameter tuning. The model parameter tuning is employed to identify a preferred classifier (e.g., Random Forest) configuration for distinguishing case versus control participants in any suitable cohort, such as, but not limited to, the age-matched cohort. For example, in some embodiments, the parameter tuning includes performing a grid search with 5-fold cross-validation using a pipeline-based framework. In such embodiments, for each feature set, the features are standardized using z-score normalization with a pipeline to ensure consistent scaling across folds. One or more supervised tasks are then defined and out-of-fold predictions are generated to avoid leakage whenever model outputs are used as genetic traits. Each individual z randomly assigned to one of the folds, where each fold is approximately 1:1 agematched in sample size, with frequency matching in a 5 -year old interval. In some embodiments, for example, a case-control task for a disease (e.., any Parkinson’s disease) vs control labels ^(seventy)£where 0 is control and 1 is disease.

[0104] Next, a classifier (e.g., Random Forest classifier [RandomForestClassifier]) is refined using grid search over key hyperparameters. The classifier may be refined over any suitable number of hyperparameters, such as, but not limited to, two or more key hyperparameters (e.g., number of estimators (n_estimators) and maximum tree depth (max_depth)). The model tuning is then conducted, for example, using five-fold stratified cross-validation (StratifiedKFold, n = 5, shuffle = True, random_state = 0) to maintain balanced class representation in each fold. Model performance can be evaluated in any suitable manner, such as, but not limited to, the area under the receiver operating characteristic curve (AUROC) averaged across folds.

[0105] Following the training described herein, the predictive model is configured to produce a severity score (pi (case-control) equal to the probability p, for individual participant i (i.e., <pi = pi). In some embodiments, the trained model can be evaluated to interpret the contribution of features in each feature set to the model’s predictions. For example, in some embodiments, intrepting the contribution of features includes computing SHAP (SHapley Additive exPlanations) values for the full model. The calculated SHAP values can then be merged with participant metadata, providing SHAP-anchored digital phenotypes.Second Construct

[0106] In some embodiments, the wearable-derived features provide higher-resolution behavioral signals that offer valuable insights into the transition from the prodromal to diagnosed stages. Accordingly, in some embodiments, such as in the second construct described above, the predictive model is trained to output a score relating to disease progression in the subject. In some embodiments, the method includes determining a progression of a health condition based upon the score generated from the predictive model. For example, in some embodiments, the predictive model is configured to generate a progression score (TT for assessing health condition progression.

[0107] In some embodiments, training the predictive model for the second construct includes creating a 1: 1 age-matched prodromal-diagnosed dataset, which minimizes theconfounding effects of age. The age-matching of the prodromal -diagnosed dataset can be performed in the same manner as discussed above for the age-matching of the case-control dataset in construct 1. Similarly, the predictive model for the second construct can be trained in the same manner as described above for the first construct, only with the age-matched prodromal-diagnosed dataset (second construct) in place of the case-control dataset (first construct). Following such training, the trained predictive model of the second construct is configured to produce a progression score TT, (prodromal-diagnosed) equal to the probability p, for individual participant i disease progression (z.e.,= pf).

[0108] As with the first construct, the trained model of the second construct can be evaluated to interpret the contribution of features in each feature set to the model’s predictions. For example, in some embodiments, intrepting the contribution of features includes computing SHAP values for the full model. The calculated SHAP values can then be merged with participant metadata, providing SHAP-anchored digital phenotypes.Third Construct

[0109] In some embodiments, such as in the third construct described above, the predictive model is configured to output a score relating to a progression rate of a health condition. In some embodiments, the method includes determining a progression rate of a health condition based upon the score generated from the predictive model.. For example, in some embodiments, the predictive model is configured to generate a progression rate (or velocity) score (V ) for assessing a rate of progression of a health condition in the subject.

[0110] In some embodiments, training the predictive model for the third construct includes training the model to generate the progression score rr;as described for the second construct and then combine the progression score with a time to diagnosis (TTD) to produce the progression rate score in progression units per unit time The TTD can be in any suitable time frame, such as, but not limited to, days, months, or years. In some embodiments, computing the TTD includes defining a progression point relative to a disease course of the health condition. For example, in some embodiments, when the health condition is Parkinson’s disease, the progression point is defined as a point that is universally crossed during a patient’s PD disease course at the time of diagnosis, where TT = 0.5 and TTDt= 0. In such embodiments, n is centered at 0.5 because this is both the classifier’s decision boundary and the hypothesized valueat the clinical boundary (TTD = 0); centering makes deviations from this boundary explicit. Accordingly, with diagnosis date tx, wearable date fAear, and TTD;=| t(wear- t-lxl(days), the velocity (progression rate) can be defined as:Tij - 0.5v,1= - -TTDl

[0111] As will be appreciated by those skilled in the art, the predictive model for the third construct can be filtered in any suitable manner. For example, in some embodiments, to enforce non-negativity and reduce proxy-phenotype noise, the diagnosed-vs-prodromal subset is restricted to those that the classifier labeled correctly, including only individuals in which score polarity and clinical time aligned. Additionally or alternatively, in some embodimetns, to limit boundary leakage and extreme leverage, individuals with TTD in [-365, +730] days are excluded, TTD=0 are removed, and / or ±3 SD around the velocity mean are trimmed.

[0112] In some embodiments, one or more of the constructs described herein is standardized with ancestry before genetics. In some such embodiments, standardizing constructs with respect to ancestry before genetic analysis reduces confounding due to population stratification and improves the validity of downstream associations. This approach mitigates ancestry -related bias and enhances generalizability across diverse cohorts.Genome-Wide Association Studies (GWAS)

[0113] Further provided herein are methods of applying the digital phenotype generated according to the methods disclosed herein in multivariate genome-wide association studies (GWASs). In some embodiments, the method includes using the derived static features and / or Al model score as response variables for GWAS (e.g., regressing wearable-derived features on genetic variants). In some embodiments, the method includes multivariate wearable GWAS, where multiple wearable-derived measurements are modelled together as a multivariate response variable to identify genetic associations. In some embodiments, the method further includes uncovering genetic biomarkers that may be condition-specific (e.g., ADHD) or broadly related to specific physiological traits measured by the wearable device (e.g., heart rate). Additionally or alternatively, in some embodiments, the method provides greater statistical power than the traditional categorical diagnostic labels, even when applied to small cohorts of individuals (n < 1,500).

[0114] In some embodiments, the multivariate wearable GWAS involves a hierarchical two-step test of association. First, a multivariate test assesses associations between each clusters of correlated features and genetic variants, boosting association power due to the correlated nature of features within clusters, and to the less stringent study -wide significance threshold (5 10‘8 / q, instead of 5 10'8 / p, with q = number of clusters, p = total number of static features, and q « p). In the second step, post-hoc univariate tests (pairwise Wilcoxon signed-rank test among genotype groups) are used to identify specific features within significant clusters driving the associations, further refining the results.

[0115] For example, in one embodiment, the method includes grouping any suitable number of wearable-derived static features (e.., 258) into clusters based on their degree of correlation, and performing multivariate tests to assess associations between each cluster of features and genetic variants, adjusting for covariates. To evaluate the relevance of significant genetic associations to ADHD, an interaction term (Genotype: Disorder) can be introduced to explore whether the genetic effect on the wearable phenotype differs between individuals with ADHD and controls. Additionally, an association can be performed to assess the relationship between genetic variants and wearable features regardless of ADHD status. In some embodiments, the Multivariate Asymptotic Non-parametric Test of Association (MANTA) R package can be used for the GWAS analyses.

[0116] In some embodiments, the GWAS methods include linking genetic data to the output of the predictive model according to any of the embodiments disclosed herein. For example, in some embodiments, the GWAS methods include linking genetic data and disease-anchored digital phenotypes for the first construct and / or the second construct. Additionally or alternatively, in some embodiments, the GWAS methods include linking genetic data and velocity for the third construct.

[0117] In some embodiments, prior to linking, the method includes performing quality control on the genetic data. In some embodiments, the quality control includes passing the genetic data through any suitable imputation filter. In some embodiments, the quality control includes applying PLINK 2.0 QC with best-guess hard calls. In some embodiments, the quality control analyses use autosomes only. Any suitable covariates can be included in the GWAS, such as, but not limited to, models adjusted for age, sex, ten ancestry PCs (PC1-PC10) obtained directly from a biobank (e.g., UK Biobank), and pre-specified lifestyle / anthropometric factors(e.g., BMI, smoking, socioeconomic index). In some embodiments, covariates are variance-standardized prior to regression.

[0118] In some embodiments, linking genetic data and disease-anchored digital phenotypes for the first construct and / or the second construct includes employing a severity construct and a binary GWAS. In some embodiments, in the severity construct for the casecontrol model, the severity score for an individual z is defined as the OOF probability of the disease (e.g., Parkinson’s disease) class: <pf0F- P( Yi = case | ) ). TreeSHAP OOF explanations yield feature-level contributionsthat additively explain (pt(baseline excluded) These feature-level contribution can be used elsewhere for interpretation, grouping, and temporal analyses. Additionally or alternatively, in some embodiments, the binary GWAS includes running a standard case-control GWAS using PLINK 2.0 logistic regression (-glm, Firth fallback when needed) with the same covariates:, ( PK = 1) \M a rM T7Here Ytis disease status, Gtis effect-allele dosage, and Ztare covariates. In some embodiments, per-allele effect sizes [3gand p values with a 5 x 10-8genome-wide threshold are reported. The same method can be employed for the second construct by substituting TT for <p.

[0119] In some embodiments, linking genetic data and velocity for the third construct includes running genome-wide univariate linear regression using PLINK 2.0 (-glm) on the quality controlled autosomal imputed data, using vtas the outcome and the covariates above. Disentanglement

[0120] In some embodiments, the methods described herein further include disentangling genetic and environmental determinants of activity and temporal patterns associate with a health condition. For example, in some embodiments, the disentangling includes relating model attributions (SHAP) to an individual-level polygenic risk score (PRS) for the health condition to separate genetic liability from environment / context in the behaviors used by the classifier. Any suitable dataset can be used for the disentangling, such as, but not limited to, the age-matched, correctly classified subset as described elsewhere herein. In some embodiments, the PRSs are computed using PRSice2 and / or the classifier models are trained with k-fold cross-validation. In some such embodiments, within each fold, SHAP explanations are computed out-of-fold forheld-out subjects using that fold’s trained model and background distribution. Because SHAP decomposes the prediction relative to a fold-specific baseline bm= E [cr(X)], in some embodiments, only the contribution terms are analyzed and then concatenated across folds.Y = 4>i -fLet <pt denote the model’s health condition severity score (case probability) for individual i. Using SHAP,, is decomposed into feature-level contributionsthat additively explain the score (i.e., <pLr equals the SHAP contribution, baseline excluded). For feature f, feature importance is summarized as the mean absolute contribution across individuals:,mpf =i

[0121] To study interpretable behavior, features are grouped into activity categories (e.g, walking, sleep, sit-stand, cycling, vehicle, overall acceleration, demographics). For each activity group g, a per-individual group contribution is computed by averaging that individual’s SHAP contributions across the features in gf tgIn some embodiments, group importance is defined as the mean importance across the group’s constituent features:ImPg =wfegThen, Sigis regressed on PRS across individuals by ordinary least squares across individuals, Sig ~ Mg + (3g, PRS} + Sig,in order to quantify each group’s genetic coupling and visualized the regression association strength e.g., \(3g\, supported by RA2 and - log10(p)) on the y-axis against Impgon the x-axis. Groups in the upper-right quadrant are simultaneously important for classification and tightly coupled to polygenic liability (genetically weighted relevance), whereas high-importance groups with low PRS association (lower right quadrant) appear more environmentally / contextually driven.

[0122] To examine diurnal structure, this procedure can be repeated at the hour-of-day level within predefined daily contexts. In such embodiments, for hour h, feature contributions - belonging to that hour are averaged to obtain Sih= — • Eh <f>i f, hourly importance is lfllsummarized as Imph= —eflImp, and the same regression is fit:$ih ~ + Ph> + Ei hThe resulting temporal plot displays \ / 3h\ versus Imph, highlighting time windows where health-condition-relevant behavior is more genetically loaded versus environmentally modulated.

[0123] In both analyses, the x-axis (average SHAP importance) quantifies relevance to the health condition, and the y-axis (association with PRS) serves as a pragmatic proxy for heritability. In some embodiments, the PRS proxy can be replaced with direct SNP -based heritability estimates for each group to place the y-axis on an h2scale.

[0124] As will be appreciated by those skilled in the art, the disentanglement described herein can be applied to any suitable health condition. For example, in some embodiments, the disentanglement is applied to Parkinson’s disease.

[0125] As many wearable-derived features (e.g., heart rate, step count, calorie consumption) are inherently correlated since they are measured simultaneously in the same individual by the same device and influenced by shared environmental and behavioral factors, modelling these features together as a multivariate response provides a number of advantages. First, it enhances statistical power for identifying genetic biomarkers, since it allows to mitigate the burden of multiple testing corrections that would otherwise arise from testing each feature independently. Second, it provides a more comprehensive view of an individual’s health across different physiological systems, compared to using a single binary metric of the disease which may be defined by arbitrary cut-offs. Both advantages enhance the identification of potential genetic risk biomarkers, which could play a pivotal role in refining diagnosis and treatment options — a critical need in neuropsychiatric disorders, where the lack of objective and sensitive screening methods remains a major barrier to effective treatment.

[0126] Without wishing to be bound by theory, it is believed that as the digital phenotype, predictive diagnostic tool, and GWAS analyses disclosed herein can accurately predict a disease or medical condition (macrophenotype), they are able to capture a higher degree of heterogeneity within traditional constructs of disease. Therefore, the systems and methodsdisclosed herein can be more effective in studying the genetic and molecular underpinnings of the condition. Additionally, the embodiments disclosed herein provide uniform processing of wearable and genomic data and integration with Al modeling and GWAS; an Al framework that uses digital phenotypes (e.g., wearable digital phenotypes) to better predict health conditions (e.g., psychiatric disorders); univariate and multivariate digital phenotypes that can act as a continuous response for GWAS; and / or wearable GWAS that detects a larger number of loci compared to traditional case-control GWAS.Model Application

[0127] Also provided herein are methods of applying the model trained according to one or more of the embodiments disclosed herein. In some embodiments, the method includes providing input data from a subject, generating digital descriptors of the subject from the input data, and inputting the digital descriptors to the model trained according to one or more of the embodiments disclosed herein. The input data can be collected directly or previously collected, and includes data measured according to any of the embodiments disclosed herein. The digital descriptor can be generated from the input data according to any of the embodiments disclosed herein. In some embodiments, the digital descriptors are input into the Al model trained to generate a digital phenotype, resulting in output of a digital phenotype specific to the subject. In some embodiments, the digital descriptors or the digital phenotype of the subject are / is input into the Al predictive model trained to generate a predictive health status of the subject, resulting in output of a health status specific to the subject. In such embodiments, the health status can include a health status according to any of the embodiments disclosed herein.

[0128] Those skilled in the art will recognize, or be able to ascertain using no more than routine experimentation, numerous equivalents to the specific procedures, embodiments, claims, and examples described herein. Such equivalents were considered to be within the scope of this invention and covered by the claims appended hereto.

[0129] It is to be understood that wherever values and ranges are provided herein, all values and ranges encompassed by these values and ranges, are meant to be encompassed within the scope of the present invention. Moreover, all values that fall within these ranges, as well as the upper or lower limits of a range of values, are also contemplated by the present application.

[0130] The following examples further illustrate aspects of the present invention.However, they are in no way a limitation of the teachings or disclosure of the present invention as set forth herein.EXAMPLESExample 1 - Digital phenotyping from wearables using Al characterizes psychiatric disorders and identifies genetic associations.Summary

[0131] Psychiatric disorders are influenced by genetic and environmental factors.However, their study is hindered by limitations on precisely characterizing human behavior. New technologies such as wearable sensors show promise in surmounting these limitations in that they measure heterogeneous behavior in a quantitative and unbiased fashion. Here, we analyze wearable and genetic data from the Adolescent Brain Cognitive Development (ABCD) study. Leveraging >250 wearable-derived features as digital phenotypes, we show that an interpretable Al framework can objectively classify adolescents with psychiatric disorders more accurately than previously possible. To relate digital phenotypes to the underlying genetics, we show how they can be employed in univariate and multivariate GW AS. Doing so, we identify 16 significant genetic loci and 37 psychiatric-associated genes, including ELFN1 and ADORA3, demonstrating that continuous, wearable-derived features give greater detection power than traditional, casecontrol GWAS. Overall, we show how wearable technology can help uncover new linkages between behavior and genetics.Introduction

[0132] Psychiatric disorders of childhood and adolescence currently affect 1 in 7 youths in the United States and globally. Externalizing disorders such as attention-deficit / hyperactivity disorder (ADHD), and internalizing disorders such as anxiety, are among the most prevalent and represent a wide spectrum of dysfunctional behavior patterns. Treatment barriers are complex and multifaceted but major contributors include our limited understanding of psychiatric phenotypes and difficulty identifying youth individuals that experience these disorders.

[0133] Traditionally, psychiatric disorders have been conceptualized as categorical macrophenotypes based on clinical manifestations of a disease, which are defined according tothe number and type of symptoms and the presence of distress or impairment. While this has practical benefits in terms of reliability and ease of diagnosis, it poses several challenges to the research of these disorders and, consequently, to the development of treatments. Furthermore, given the high heritability of psychiatric disorders, dissecting their underlying genetic architecture is of interest to researchers. While cost-effective and accurate genotyping technologies in large cohorts of individuals have significantly advanced the field, barriers associated with missing heritability and the need for improved phenotyping strategies are still present. In fact, many psychiatric genome-wide association studies (GWASs) to date rely on subjective, dichotomized (i.e., binary) traits. However, psychiatric disorders are complex and often comorbid, and this high degree of heterogeneity is not always accurately translated into categorical diagnostic labels, which may be defined by arbitrary cut-offs. Digital phenotypes derived from biosensor data can address these challenges by more precisely capturing an underlying macrophenotype and better describing the heterogeneity potentially missed by existing diagnostic categories. Therefore, when compared to macrophenotypes, the quantitative nature of these digital phenotypes could enable improved dissection of the genetic architecture underlying psychiatric disorders.

[0134] To improve our understanding of psychiatric disorders, it is important that we identify digital phenotypes that not only offer a more comprehensive representation of an individual’s behavior with respect to the environment but also relate well with existing clinical definitions and aid in diagnosis. Once identified, these digital phenotypes can also be used to guide more comprehensive studies to identify genetic associations and biomarkers that may ultimately improve precision treatments.

[0135] To achieve this goal, it is important to leverage new emerging technologies that can quantitatively assess an individual’s behavioral patterns. Wearable sensors such as smartwatches collect data that reflect physical and physiological processes (e.g., movement, pulse, metabolic intake), and can be used to infer higher-order behavioral events (e.g., sleep, exercise) and their temporal dynamics. Because of the documented relationship between such higher-order behavioral events and mental health, and given their low cost and minimal invasiveness, wearable devices have emerged as promising tools for mental health monitoring and psychiatric evaluation. Indeed, there is a rich literature leveraging wearable sensors for thecharacterization, detection, and even treatment of childhood and adolescent disorders, such as common externalizing (e.g., ADHD) and internalizing (e.g., anxiety disorders) disorders.

[0136] Prior studies have demonstrated the utility of wearables for detecting key behavioral and physiological traits in youths with psychiatric disorders. For example, markers of physical fitness measured by these devices have been associated with psychopathology and may offer insights into treatment interventions. Consequently, wearable biosensors show promise for capturing digital phenotypes relevant to behavior and psychiatric disorders, ultimately enabling improved GWASs. However, significant computational challenges remain in using Al to generate digital phenotypes that leverage the temporal nature of the raw wearable data and describe the full spectrum of a given psychiatric disorder. Moreover, further curation of these digital phenotypes is necessary to identify genetic associations that have clinical and biological relevance.

[0137] To address these limitations, we developed an Al modeling framework that flexibly leverages data from wearable devices to generate digital phenotypes in the form of static and dynamic digital features. We establish the validity of these features as digital phenotypes by classifying externalizing and internalizing disorders with an accuracy beyond baseline expectation, and even surpassing the performance of some other gold-standard digital phenotypes, such as functional magnetic resonance imaging (fMRI) measurements.Interpretability modules in our Al framework enable us to identify key temporal and physiological insights between clinical diagnosis and wearable-derived features, further supporting their validity as digital phenotypes. We further curate these digital phenotypes and employ them in GWAS models to identify genetic associations and biomarkers that capture the continuous spectrum of psychiatric disorders and behavioral patterns. Finally, we identify 16 significant loci, several of which overlap previously reported genetic variants associated with behavioral traits and mental illnesses and are proximal to genes with a documented role in neurodevelopmental and psychiatric disorders.

[0138] In sum, this work shows how wearable devices can advance our understanding of psychiatric disorders by establishing a more objective and dimensional approach that can ultimately lead to improved treatments in precision psychiatry.ResultsLeveraging the Adolescent Brain Cognitive Development cohort

[0139] To improve our understanding of psychiatric disorders, we leveraged and analyzed a dataset from a cohort of US adolescents recruited by the NIH Adolescent Brain Cognitive Development Consortium (ABCD) project, consisting of clinical, wearable, and genetic data (FIG. 1, FIG. 6 (Panel A), and FIG. 11; STAR Methods “Dataset Description” section) (Volkow, N. D., Koob, G. F., Croyle, R. T., Bianchi, D. W., Gordon, J. A, Koroshetz, W. J., Perez-Stable, E. J., Riley, W. T., Bloch, M. H., Conway, K, et al. (2018). The conception of the ABCD study: From substance use to a broad NIH collaboration. Dev Cogn Neurosci 32, 4-7). The ABCD cohort consists of a total of 11,878 adolescents (5,682 males and 6,196 females), of age between nine and fourteen years and belonging to four different ethnicities. We identified nine categories of psychiatric phenotypes (Table SI), which were established using a gold standard parent diagnostic semi-structured interview (Kiddie Schedule for Affective Disorders and Schizophrenia- 5) (van Dijk, M. T., Murphy, E., Posner, J. E., Talati, A., and Weissman, M. M. (2021). Association of Multigenerational Family History of Depression With Lifetime Depressive and Other Psychiatric Disorders in Children: Results from the Adolescent Brain Cognitive Development (ABCD) Study. JAMA Psychiatry 78, 778-787). The healthy controls represented adolescents who did not meet the criteria for any of those nine psychiatric disorders We defined these clinical labels as the categorical macrophenotypes in the study (FIG. 1 (Panels A-B)). Our modeling framework also included extensive covariates typical of psychiatric studies, such as demographics, cognitive tests (e.g., NIH Toolbox), and behavioral checklists (FIG. 2 (Panel A), and FIGS. 12A-L, and Data S40-S44).Table SI - Psychiatric cohorts and sample sizes, related to FIG. 1 and STAR Methods “Dataset Description” section.Cohort Sample size Anxiety Disorders 1,232 Attention-Deficit / Hyperactivity Disorder (ADHD) 400 Obsessive / Compulsive Disorder (OCD) 341 Panic Disorder 180 Sleep-related Disorders 128 Bipolar / Psychotic Disorders 255Eating Disorders 96 Depressive Disorders 35 Post-Traumatic Stress Disorder (PTSD) 23Generating digital phenotypes from wearable-derived data

[0140] We processed data obtained from FitBit smartwatches, which comprise measurements of heart rate, calories, activity intensity, steps, metabolic equivalents (METs), sleep level and sleep intensity (FIG. 1 (Panel C) and FIG. 6 (Panel B) and Data S45-S46; STAR Methods “Dataset Description” section). These measurements quantify an individual’s physiological processes and their real-time changes in response to environmental stimuli, and can thus provide key information about an individual’s behavior.

[0141] To reconstruct the full spectrum of an individual’s behavioral functioning from these data, we applied two different feature engineering techniques, allowing us to generate wearable-derived dynamic and static features, which we consider as digital phenotypes. The dynamic features preserve the time-varying nature of the original data as a time series, enabling sequential and temporal patterns of the data to be retained. In contrast, the static features summarize patterns of the digital data and produce time-invariant, quantitative features that are commonly used in downstream modeling.

[0142] To generate dynamic features, we performed signal imputation and processing after filtering the individuals with sparse data, and obtained 48 channels of time series (FIG. 2 (Panel B), FIG. 7A, and FIGS. 13-14; STAR Methods “Dataset Description” section) Compared to the static features, this further processing allowed us to preserve both local and global temporal patterns potentially relevant to characterizing behavior and neurological response to stimuli.

[0143] To generate static features, we first collected a total of 49 FitBit summary-based features (Data S45-S46; STAR Methods “Machine Learning Classifier” section). We next applied descriptive statistics (e.g. mean, median, etc.) to each of these features and generated a total of 258 static features for each individual (FIG. 2 (Panel C) and Data S47). We then grouped these static features into seven main clusters, each of which summarizes different aspects of physiological and behavioral processes, such as heart rate, sleep duration and quality,metabolic intake or physical activity (FIG. 2 (Panel D) and Data S48-S49; STAR Methods “Machine Learning Classifier” section).

[0144] Altogether, static and dynamic features represent the physiological and behavioral profiles of the adolescents, and can be leveraged as digital phenotypes in a wide range of analyses to better characterize psychiatric phenotypes. For instance, we generated wearable combination scores, performed macrophenotype classification, and assessed model interpretability. In particular, the wearable combination scores integrate our digital features using non-linear models to predict the macrophenotype (FIG. 2 (Panel E); STAR Methods “Model Training and Evaluation” section). Practically, these scores summarize the likelihood that an individual has a given disorder and further enable biomarker identification via wearable GWAS Classifying psychiatric macrophenotypes using wearable-derived digital phenotypes

[0145] To demonstrate the validity of static and dynamic features as clinically relevant digital phenotypes and to evaluate their utility as a diagnostic tool, we employed these features in an array of classification tasks to identify individuals with either an externalizing (ADHD) or internalizing (anxiety) disorder from their typically developing peers. We selected ADHD and anxiety due to their high prevalence in adolescents, which is mirrored in the cohort (FIG. 1 (Panel B) and Table SI).

[0146] We applied a gradient boosting machine learning algorithm, XGBoost, for classification tasks using static features (FIG. 2 (Panels C and D); STAR Methods “Machine Learning Classifier” section). On the other hand, to fully leverage the time series nature of the dynamic features, we used a convolutional neural network for time series, featuring depthwise separable convolution, called Xception (FIG. 2 (Panel E) and FIG. 7B; STAR Methods “Multichannel Time Series Classifier” section). Variable convolutional filters and residual (skip) connections, coupled with efficient parametrization, allow short and long time series patterns of physiology and behavior to be optimally leveraged when performing downstream classification of psychiatric disorders. In both modeling approaches we also considered our full list of covariates (FIGS. 2 (Panel A), 7A, and 15-18, and Data S40-S44, S50). To assess the benefit, in terms of model performance, of including wearable-derived data, we also trained a baseline model using just the covariates, which served as a comparison to the models including static or dynamic features (Data S51-S52). In practice, this comparison allowed us to determinewhether wearable-derived features can improve diagnostic accuracy relative to that achievable using only a widely used broadband behavior rating scale.

[0147] After data filtering, we first used static features to classify 216 individuals with ADHD (an externalizing disorder) versus 1,737 of their typically developing peers (healthy controls) (FIG. 3 (Panel A), FIGS. 8A-B, and FIGS. 19-21; STAR Methods “Model Training and Evaluation” section). Using static features with XGBoost, we achieved an average area under the receiver operating characteristic curve (AUROC) of 0.87 and precision of 0.79. When using the dynamic features and Xception, we were able to achieve an average AUROC of 0.89 and precision of 0.83. The baseline model consisting of only the covariates achieved an average AUROC of 0.83, suggesting that the inclusion of wearable-derived features facilitates a clinically meaningful improvement in diagnostic accuracy. This improvement between the baseline and dynamic features model demonstrates statistical significance (one-sided / -test between baseline model and dynamic features model, / ? value = 0.0022).

[0148] Second, we evaluated the performance of our model using static or dynamic features in the classification of 666 individuals diagnosed with anxiety disorder (internalizing disorder) versus 1,737 of their typically developing peers (healthy controls) (FIG. 3 (Panel B) and FIGS. 21-22; STAR Methods “Model Training and Evaluation” section). Here, we again repeated the use of the same modeling framework, i.e., static features with XGBoost and dynamic features with Xception, and compared it to the baseline covariate model. We found that static and dynamic features achieve an average AUROC of 0.69 and 0.71 and precision of 0.64 and 0.68, respectively. In both models, the performance was greater than that of the baseline model (average AUROC of 0.67), with the dynamic features model showing the largest and most significant improvement in performance (one-sided / -test between baseline model and dynamic feature model, p value = 0.00016). Overall, the fact that the models using dynamic features achieved the highest performance suggests that the temporal patterns intrinsic to wearable-derived data are useful towards understanding human behavior.Interpreting wearable features prioritized by the deep learning model

[0149] Deep learning methods are typically characterized by complex internal structures that cannot be easily interpreted by humans. While maximizing the classification accuracy is one crucial aspect for characterizing complex phenotypes, it is also critical to understand which features are most important for the classification task. To this end, we utilized ablationtechniques to determine the relative contribution of each individual feature to model performance (FIG. 3 (Panels C-D) and FIGS. 23-27; STAR Methods “Model Interpretability” section) For the ADHD classification task, heart rate was the most important feature (largest change in AUROC), followed by other dynamic features (i e., sleep, steps, METs) as well as covariates such as demographics, family history, and cognitive scores from picture memory and stop-signal reaction time tests (FIG. 3 (Panel C) and FIGS. 23-26)

[0150] On the other hand, the ablation study for the anxiety classification task revealed a different set of important features. In this case, sleep quality and stage, calories, and step count were the most important dynamic features, whereas heart rate features, which were extremely important for classifying ADHD, were not prioritized in the anxiety model (FIG. 3 (Panel D) and FIG. 27). Additionally, while the anxiety model prioritized some covariates that were relevant also for the ADHD model (e.g., sex, family history, and family divorce), cognitive scores from tests such as picture memory did not appear to be important for the identification of individuals diagnosed with anxiety, consistent with theory-driven accounts of neurocognitive aspects of anxiety disorders.

[0151] Additionally, we assessed the importance that dynamic features at various timepoints of the day have on model performance. Specifically, we calculated the relative importance of each time point using gradient-weighted class activation mapping (Grad-CAM) and ablation techniques (FIG. 3 (Panels E-F) and FIGS. 28-29; STAR Methods “Model Interpretability” section). For ADHD, we observed enriched significance of the heart rate dynamic feature around the early afternoon, potentially suggesting stronger behavioral differences between adolescents with ADHD and their typically developing peers (healthy controls) during this time of day (FIG. 3 (Panel E) and FIG. 28). This is consistent with clinical research demonstrating time-of-day effects on ADHD symptom expression. In contrast, sleep-related dynamic features during the night are much more informative in classifying anxiety, consistent with clinical expectations (FIG. 3 (Panel F) and FIG. 29). Together, these results suggest a role for wearable-derived features to not only serve as digital phenotypes, but also to more closely reveal insights into the behavioral and physiological temporal patterns related to categorical macrophenotypes.Using wearable-derived features as digital phenotypes in GWAS for ADHD

[0152] Our Al modeling framework used wearable-derived features as predictors (independent variables) of psychiatric macrophenotypes. Our accurate predictions suggest that these quantitative features could be useful for studying other aspects of psychiatric disorders, such as their underlying genetic architecture. Therefore, we leveraged these features as digital phenotypes (dependent / response variable) in the following GWAS to identify genetic associations relevant to psychiatric conditions (FIG. 1 (Panel D), FIGS. 4A-C, and FIGS. SO-35; STAR Methods “Quality Control of Genetic Data” and “Covariates included in the GWAS” sections). We focused specifically on ADHD (STAR Methods “GWASs for ADHD” section), given the higher predictive power observed with our models (FIG. 3 (Panels A-B)) and its higher estimated heritability compared to anxiety (>75% vs. 30-60%). We selected 1,191 individuals (137 individuals with ADHD and 1,054 healthy control individuals) with genetic and wearable data available. We first treated the clusters of wearable-derived features as a digital phenotype (i.e., the response variable) and assessed whether genetic variants impact these features differently in ADHD vs. control individuals. Specifically, we performed a continuous multivariate GWAS regressing the vector of wearable features on the genotype, the covariates, and an interaction term between the genotype and an individual’s macrophenotype (ADHD or control) (FIG. 9 A). By including this interaction term, we were able to identify genetic variants that not only have a significant impact on the digital phenotype, but are also relevant to ADHD. We identified two genome-wide significant (p value < 5 - 10‘8) loci and six psychiatry-related genes (FIG. 4B and Table 1 and Table S2 and FIGS. 36-38). One of these loci is located within a cluster of genes relevant for ADHD (ELOVL5, FBX09, CILK1). Following the continuous multivariate GWAS, we performed a post-hoc test to determine which of the wearable features within a cluster drives the significant association (FIG. 9B and Data S53). We found that individuals with ADHD carrying the CC genotype at rsl86003 (chr6:53,320,326) reported lower amounts of sedentary time compared to ADHD individuals carrying the AA genotype (FIG. 4B).This difference was not observed among the three genotype groups in control individuals, suggesting that the effect of the genotype on the wearable feature is specific to the ADHD condition. Overall, this hierarchical testing strategy (i.e., multivariate GWAS followed by a post-hoc test) allowed us to reduce the multiple-testing correction burden, resulting in increased statistical power.Table 1 Results for the 16 genetic loci identified by the continuous univariate and multivariate GWASs.Locus Chr Start End Lead Position P Genes GWAS Phenotype Variant value Method1 1 111,372 111,482 rs114081 111,372,1 1.1 I E- TMIGD3,,165,359 965 66 08 ADORA3,RAP1A,CHI3L22 3 161,873 161,927 rs792032 161,909,2 3.89E- -,055,820 33 61 083 4 184,417 184,424 rs142555 184,421,9 2.15E- IRF2, CASP3,,766,056 1 04 08 PRIMPOL4 10 121,524 121,582 rs140794 121,524,6 2.31 E- FGFR2,611,200 722 12 095 11 38,982, 39,384, rs151239 39,273,49 3.80E- - ADHD 793 610 852 7 086 14 26,214, 26,216, rs149074 26,216,53 2.95E- NOVA1 Continuous683 532 469 2 08 Univar.7 14 62,903, 63,014, rs143225 63,014,79 3.89E- KCNH5, RHOJ677 798 169 8 088 17 7,101,6 7,101,6 rs116530 7,101,608 8.51 E- CLEC10A, DLG407 08 54 099 17 32,256, 32,283, rs650529 32,270,86 1.12E- RHBDL3,997 356 3 3 08 RHOT1,C17orf75,ZNF207,PSMD11,LRRC37B,CDK5R1,MYO1D10 19 4,495,6 4,495,6 rs150855 4,495,611 2.75E- - 10 11 276 0811 11 103,922 104,046 rs750926 104,046,7 1 55E- AMY1C,953,796 61 96 08 Continuous12 6 53,240, 53,356, rs186003 53,320,32 3.46E- GCLC, CILK1, Multivar429 412 6 08 ELOVL5,FBXO9, GCM113 6 1,102,2 1,130,1 rs742521 1,129,893 5.20E- - 52 54 0914 7 1,789,3 1,791,3 rs113525 1,791,353 5.09E- MAD1L1,21 53 298 09 ELFN1, PSMG3, Continuous Behavioral MAFK Multivar15 14 23,392, 23,418, rs365990 23,392,60 5.33E- MYH6, CMTM5,601 974 2 09! L25, BCL2L2,BCL2L2- PABPN116 16 79,283, 79,302, rs805162 79,288,21 6.04E- WWOX253 474 5 7 09For each locus we report the genomic coordinates in human assembly GRCh38, the lead variant rsID with corresponding genomic position and p value, the GWAS method (continuous multivariate or continuous univariate), and the phenotype (ADHD or behavioral traits) associated with the GWAS. Brain- orneuropsychiatry -related genes proximal to the locus are also listed in Tables S2-S4,Table S2 - Genome-wide significant loci identified by the continuous multivariate GWAS for ADHD, related to FIGS. 4A-C, Table 1, and STAR Methods “GWASs for ADHD” & “Statistical Significance and Functional Dissection of GWAS” sections.OverlapBrain- Brain- withLocus Lead GWAS related related Chr Start End Position P value ENCOnumber Variant run nearby nearby DE4genes eGenes cCREs103,922, 104,046, rs7509266 104,046, 1.54872 Cluster1 1 yes AMY1C - 953 796 1 796 E-08 3GCLC. C CILK1, ILK1, ELOVL5 53,240,4 53,356,4 53,320,3 3.46138 Cluster ELOVL5 2 6 rsl86003 yes29 12 26 E-08 5FBXO9, FBXO9, GCM1 GCM1 Expanded version of Table 1 for the results derived from the continuous multivariate GWAS for ADHD. For each locus we report genomic coordinates in assembly GRCh38, the lead variant rsID with corresponding genomic position and p value, and the cluster of wearable-derived static features used as multivariate response variable ('‘GWAS run”) Additionally, we annotate whether the locus overlaps an ENCODE4 candidate cis-regulatory element (cCRE), and its proximal brain or neuropsychiatry- related genes and eGenes (i.e., proximal genes with an eQTL overlapping the locus).

[0153] In addition to using clusters of wearable-derived features as a multivariate response variable for GWAS, we also conducted another type of GWAS using, as a response variable, the wearable combination scores derived from our Al framework. These scores combine wearable-derived features into a single continuous variable that summarizes an individual’s likelihood for ADHD. When using these scores in a continuous univariate GWAS, we identified 10 significant loci and 21 psychiatric or brain-related genes (FIG. 4C and Table 1 and Table S3 and FIGS. 39-40). Three of the identified genes (ADORA3, PSMD11 and DLG4) have been previously associated with ADHD, bolstering the overall functional significance of the results. Furthermore, several of these loci overlap with previously reported GWAS SNPs related to ADHD, neuroticism, sleep disruption and other clinically relevant traits (FIG. 4C, FIG. 41, and Data S54). In comparison to either of the above two GWASs, when performing a traditional case-control (i. e., binary univariate) GWAS for ADHD on the same set of individuals using the binary diagnostic label (presence / absence of disorder) as response variable, we did not identify any significant loci (FIG. 1 (Panel A) and FIG. 4A). This result is consistent with the higherstatistical power of continuous measurements over dichotomized (i.e., binary) traits (FIG.9C; STAR Methods “Statistical Power of Binary vs. Continuous Traits” section).Table S3 - Genome-wide significant loci identified by the continuous univariate GWAS for ADHD, related to FIGS. 4A-C, Table 1, and STAR Methods “GWASs for ADHD” & “Statistical Significance and Functional Dissection of GWAS” sections.OverlapBrain- Brain- withLocus Lead GWAS related related Chr Start End Position P value ENCOnumber Variant run nearby nearby DE4genes eGenes cCRESTMIGD TMIGD3,XGB + 3, 111,482, rsll408 111,372, 1.10916 ADORA1 1 111,372,165 CBCL ADORA 359 1965 166 E-08 yes 3,v2 3,RAP1A, CHI3L2 CHI3L2 XGB +161,927, rs79203 161,909, 3.891372 3 161,873,055 CBCL yes - - 820 233 261 E-08v2IRF2, 184,424, rsl4255 184,421, 2.14745 Xceptio CASP3, 3 4 184,417,766 yes IRF2056 51 904 E-08 n PRIMPOL XGB +121,582, rsl4079 121,524, 2.3052E4 10 121,524,611 CBCL yes FGFR2 FGFR2200 4722 612 -09v2XGB +39,384,6 rsl5123 39,273,4 3.804685 11 38,982,793 CBCL yes - - 10 9852 97 E-08v226,216,5 rsl4907 26,216,5 2.94681 XGB +6 14 26,214,683 no NOVAI - 32 4469 32 E-08 CBCL63,014,7 rsl4322 63,014,7 3.88941 Xceptio KCNH5, KCNH5, 7 14 62,903,677 yes98 5169 98 E-08 nv2 RHOJ RHOJXGB + CLEC10 7,101,60 rsll653 7,101,60 8.511379 17 7,101,607 CBCL no - 8 054 8 E-09 A,v2 DLG4RHBDL3, RHBDL RH0T1, 3, C17orf7 RHOT1, 5, C17orf7 ZNF207, 5, 32,283,3 rs65052 32,270,8 1.12122 CBCL8 17 32,256,997 yes PSMD1 PSMD156 93 63 E-08 ext1, 1, LRRC37 LRRC37 B, B, CDK5R CDK5R 1, 1 MYOID XGB +4,495,61 rsl5085 4,495,61 2.7450510 19 4,495,610 CBCL yes1 5276 1 E-08 - - v2Expanded version of Table 1 for the results derived from the continuous univariate GWAS for ADHD. For each locus we report genomic coordinates in assembly GRCh38, the lead variant rsID with corresponding genomic position and p value, and the wearable combination score used as variable response (“GWAS run”). Additionally, we annotate whether the locus overlaps an ENCODE4 candidate cis-regulatory element (cCRE), and its proximal brain or neuropsychiatry-related genes and eGenes (ie.proximal genes with an eQTL overlapping the locus).Employing digital phenotypes to detect genetic associations with behavioral traits

[0154] While the GWASs above focus on a particular disorder (ADHD), it is also possible to directly use the full set of wearable-derived features from the pooled set of individuals across all disorders and controls, as a way to represent the continuum of psychiatric disorders and mental states. In fact, these features can collectively capture behavioral patterns by measuring physiological processes and their real-time changes in response to environmental stimuli, and are not restricted to a specific cohort of individuals. Therefore, we next performed a multivariate GWAS where we regressed the vector of wearable features on the genotype of each genetic variant (FIGS. 5A and 10A; STAR Methods “Continuous multivariate GWAS for behavioral traits” section), employing a larger cohort spanning healthy controls and individuals with any psychiatric disorder (n = 2,410). Similar to the previous multivariate GWAS, we performed a post-hoc test to identify which features within a particular cluster drive the significant genetic association. In this case, we identified four significant loci and ten genes with a documented role in neurodevelopmental and psychiatric disorders (FIG. 5A and Table 1 and Table S4 and FIGS. 42-46, and Data S55; STAR Methods “Statistical Significance and Functional Dissection of GWAS” section). Many of these loci overlap with previously identified GWAS SNPs related to heart and brain traits (FIGS. 5A and 47). This aligns with theclose association between physiological functions, the central nervous system, and individual behavior.Table S4 - Genome-wide significant loci identified by the continuous multivariate GWAS for behavioral traits, related to FIGS. 5A-C, Table 1, and STAR Methods “Continuous multivariate GWAS for behavioral traits” & “Statistical Significance and Functional Dissection of GWAS” sections.OverlapBrain- Brain- withLocus Chr Start End Lead Position P value GWAS ENCO related related number Variant run nearby nearby DE4genes eGenes cCREs1,102,25 1,130,15 rs74252 1,129,89 5.2039E1 6 cluster 2 yes - - 2 4 1 3 -09MAD1L 1, 1,789,32 1,791,35 rsll352 1,791,35 5.094182 7 cluster 1 yes ELFNI, ELFNI 1 3 5298 3 E-09PSMG3, MAFK MYH6, CMTM5, IL25, MYH6, 23,392,6 23,418,9 rs36599 23,392,6 5.32594 heart BCL2L2 CMTM5 3 14 yes01 74 0 02 E-09 rate, IL25,PABPN BCL2L2 1, BCL2L2 79,283,2 79,302,4 rs80516 79,288,2 6.044284 16 cluster 1 yes WWOX53 74 25 17 E-09 - Expanded version of Table 1 for the results derived from the continuous multivariate GWAS for behavioral traits using the pooled set of individuals For each locus we report genomic coordinates in assembly GRCh38, the lead variant rsID with corresponding genomic position and p value, and the cluster of wearable-derived static features used as multivariate response vanable (“GWAS run”) Additionally, we annotate whether the locus overlaps an ENCODE4 candidate cA-regulatory element (cCRE), and its proximal brain or neuropsychiatry-related genes and eGenes (i e., proximal genes with an eQTL overlapping thelocus)

[0155] To further investigate the loci identified by the behavioral GWAS, we dissected the variants using a battery of publicly available genomic resources (Consortium, E. P., Moore, J. E., Purcaro, M. J., Pratt, H. E., Epstein, C. B., Shoresh, N., Adrian, J., Kawli, T., Davis, C. A., Dobin, A., et al. (2020). Expanded encyclopaedias of DNA elements in the human and mouse genomes. Nature 583, 699-710; Consortium, G. T. (2020). The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science 369, 1318-1330). Many of these loci overlapeither GTEx expression quantitative trait loci (eQTLs) or ENCODE candidate c is- regulatory elements (cCREs), suggesting a link between the biochemical activity of these variants and their functional impact on the macrophenotype (Table S4; STAR Methods “Statistical Significance and Functional Dissection of GWAS” section). We also explored the impact of these loci beyond behavioral traits and their relationship with clinical psychopathology. For example, behavioral traits significantly associated with a specific genetic variant may correlate with clinical symptoms of a specific psychiatric cohort. Indeed, in some cases we show that the genetic variant in question is also differentially enriched between that specific psychiatric cohort and healthy individuals. For instance, we found the minor allele (G) at rs365990 to be significantly associated with an increase in mean heart rate and a decrease in interday heart rate variation (FIG. 5B, left). The variant, missense for MYH6. had been previously linked to atrial fibrillation, ventricular tachycardia and resting heart rate, and the entire locus shows a significant enrichment of chromatin features in heart samples compared to other tissues (FIG. 10B). We also found the same allele to be enriched in the bipolar / psychotic disorder cohort compared to healthy controls (FIG. 5B, right). This cohort included youth meeting criteria for bipolar or unspecified psychotic spectrum disorder, and such severe pathology is known to be associated with characteristic irregularities in heart activity. SNP rs365990 is also a GTEx eQTL for the CMTM5 gene (FIGS. 10C-D), which is highly expressed in brain subregions and has been implicated in stress response and childhood adversity, further supporting the relevance of this locus for psychiatric conditions in addition to heart pathophysiology. In a similar fashion, we explored variant rsl 13525298. The minor allele at rsl 13525298 is associated with prolonged periods in bed and shorter vigorously active time during the day, and it appears at a lower frequency in the ADHD cohort compared to healthy individuals (FIG. 5C). This suggests a potential protective role of the allele against hyperactivity disorders, further supported by the proximity of the SNP to ELFN1, previously implicated in the pathophysiology of ADHD.

[0156] Overall, these results highlight how wearable-derived features can be leveraged as digital phenotypes in GWAS and enable the identification of genetic variants relevant to clinical psychiatry, with significant effects on exhibited behaviors among adolescents.Discussion

[0157] Psychiatric disorders have been traditionally described with diagnostic categories based on retrospective self-report of symptom sets. However, current efforts in the field areincreasingly leveraging novel technologies to transition from retrospective self-reporting and fixed symptom sets to more dimensional conceptualizations. This aims to capture the complex and heterogeneous nature of psychiatric disorders for more accurate research into their underlying structure. One approach to enhancing dimensional models is the use of quantitative phenotypes. Although quantitative phenotypes have been derived from cellular, tissue, and organ levels of information, computational strategies that generate useful quantitative phenotypes in the behavioral domain are currently limited. Wearable biosensors such as smartwatches offer a unique opportunity to objectively study psychiatric disorders in a non-invasive way by measuring their underlying physiological and behavioral foundations over time.

[0158] Towards this end, we used wearable data to generate static and dynamic features that were employed by our Al modeling framework as digital phenotypes to distinguish between adolescents with and without psychiatric disorders. Models utilizing these wearable-derived digital phenotypes performed comparably to those based on more expensive data sources such as fMRI measurements. To gain critical theoretical insights and inform treatment development efforts, we augmented the modeling framework with interpretability modules, allowing us to pinpoint temporal and functional regions of the time series that were highly correlated with overall diagnostic status. These interpretability modules have the potential to facilitate mechanistic studies that offer deeper insight into the underlying complexities of these disorders. For example, our interpretability modules revealed that heart rate time series held high importance in predicting ADHD. This finding aligns with the clinical manifestation of ADHD-affected children are characterized by episodes of heightened arousal that are often incongruent with environmental demands. Conversely, the interpretability modules identified sleep intensity and quality as key predictors in our anxiety disorder models, in line with disruptions in sleep patterns and circadian rhythms commonly observed in youth with anxiety disorders.

[0159] Wearable-derived digital phenotypes are not just effective for detecting the presence of psychiatric disorders in individuals, but they also serve as a valuable research tool for understanding the correspondence between behavior patterns and molecular attributes. This comprehensive approach helps to uncover the foundational elements of pathological behavior patterns. In this context, we focused on establishing links with genetics. Specifically, we showed that these digital phenotypes can serve as response variables in GWAS models. Their continuous nature enhances statistical power compared to categorical diagnostic labels. Furthermore, wetook advantage of the features’ correlated structure to create multivariate response variables and implemented a hierarchical testing strategy that increased the statistical power of our GWAS. This strategy is advantageous because it mitigates the multiple-testing correction burden that arises from independently evaluating numerous features. Conversely, from a biological standpoint, these wearable GWASs allowed us to explore triaxial associations encompassing genetic, physiological, and psychiatric factors. Utilizing our framework, we successfully identified a significant association between a missense variant of the MYH6 gene, which encodes the cardiac muscle myosin, and heart rate patterns. Heart activity receives complex inputs from the CNS, which implies behavioral influence and, in combination with our GWAS, supports the notion of a gene-behavior-disorder pathway. Building on this finding, we discovered enrichment of the same genetic variant among individuals with bipolar / psychotic disorders, psychiatric conditions known to be associated with characteristic irregularities in heart activity. While additional research is needed to confirm such associations, our findings resonate with the objectives of the RDoC initiative. Specifically, wearable-derived digital phenotypes serve as objective markers of behavior, bridging lower-level biological systems like genetics to broader psychiatric disorders.

[0160] Although we report several associations between genotype, digital phenotypes, and psychiatric macrophenotypes, this study did not consider causal relationships between digital phenotypes and psychiatric macrophenotypes. For example, in some cases the digital phenotype could be situated at an intermediate level in the causal chain between genotype and macrophenotype. However, in other cases the psychiatric macrophenotype could lead to secondary changes in the digital phenotype. Moreover, this relationship could involve bidirectional associations or positive feedback loops. Finally, other scenarios involving environmental or non-genetic exposures (e.g., medication) could also play a role in the interplay of digital phenotype and macrophenotype (FIG. 6). Further experimentation and analyses, such as Mendelian randomization, will be required to establish the structure of causal chains linking genotype, digital phenotype, and macrophenotype.

[0161] While we have employed these wearable-derived digital phenotypes in a targeted research context (i.e., to enhance a psychiatric GWAS), their broad applicability makes them promising for other domains of health research. For example, the scores generated by our AI-modeling framework could be used to assess disorder severity, and the genetic variants identifiedby our wearable GWAS could be employed to construct more comprehensive polygenic risk scores for behavioral and psychiatric disorders. Unlike other diseases (e.g., cancer) where objective biomarkers are common, psychiatry faces a significant barrier in treatment due to the lack of objective and sensitive screening methods. Therefore, these physiological and genetic features could be leveraged as objective biomarkers to subtype patients more accurately within diagnostic categories, which in turn could help move towards precision treatment delivery in psychiatry. This approach holds promise in identifying early markers of treatment efficacy, potentially offering a more dynamic assessment of therapeutic impact. Additionally, the integration of interpretability modules provides a novel avenue to correlate physiological and behavioral characteristics with higher-order pathological traits. Such correlations may reveal the physiological and behavioral foundations of specific psychiatric disorders, ultimately guiding the development of novel treatment strategies in research settings and personalized application of treatments in clinical practice.

[0162] We anticipate that further development of our Al modeling framework, coupled with an expanded array of wearable devices, could transform how psychiatric disorders are measured and understood in both research and clinical settings. This could lead to more nuanced digital phenotypes and open new avenues for the study of human behavior.Method DetailsDataset Description (related to “Leveraging the Adolescent Brain Cognitive Development cohort” and “Generating digital phenotypes from wearable-derived data” in the main text, FIGS. 1, 2, 6, 7A-B, and Table SI)ABCD Dataset

[0163] The Adolescent Brain Cognitive Development (ABCD) study is a comprehensive longitudinal project initiated in 2015 with the purpose of characterizing the neural, cognitive, and behavioral aspects of adolescent development. Commissioned by a consortium of U. S. federal agencies, ABCD investigators deeply phenotyped a large and representative sample of children aged 9-14 with plans to track their development into early adulthood. The ABCD dataset incorporated multimodal brain imaging data, substance use history, behavioral and psychological measures, genetic data, and an all-encompassing collection of demographic, physical health and activity, mental health, and environmental information, including data derived from wearable devices.Clinical Diagnoses

[0164] Clinical diagnoses were operationalized using the parent report version of the Kiddie Schedule for Affective Disorders and Schizophrenia (KSADS). The KSADS is a gold standard semi-structured diagnostic interview that is used to establish a broad range of clinical diagnoses in children and adolescents. It has been previously used to define clinical groups in case / control studies conducted with data from the ABCD study.Cohort Definitions

[0165] We identified several clinical groups of interest in order to evaluate our framework across different forms of psychopathology. The nonclinical comparison cohort was composed of youth who did not meet any current diagnostic criteria for any disorder on their most recent administration of the parent report KSADS. Similar diagnostic categories, based on ICD10 diagnostic codes, were combined to create cohorts with sufficient sample sizes for our modeling framework. Each clinical cohort was composed of the following diagnostic groups, based on the currently reported symptom sets: Anxiety Disorder, Attention-Deficit / Hyperactivity Disorder (ADHD), Obsessive / Compulsive Disorder (OCD), Panic Disorder, Sleep-related Disorders, Bipolar / Psychotic Disorders, Eating Disorders, Depressive Disorders, Post-Traumatic Stress Disorder (PTSD) (FIG. 1 (Panel B), Table SI, and FIGS. 12A-L)Preprocessing and Quality Control of Wearable Device Data

[0166] We commenced by combining, into a single dataframe, data from seven distinct wearable-derived modalities (heart rate, calories, intensity, steps, METs, sleep level, and sleep intensity) collected for 5,339 individuals, resulting in highly sparse data structures (FIG. 1 (Panel C) and FIG. 6 (Panel B)). We excluded individuals with at least one missing wearable modality, leaving us with 3,538 participants. To address the impact of missing values on further analysis, we implemented a two-stage quality control procedure. In the initial phase, for each data modality we examined all potential time windows for two selected days in each week. Our objective was to balance the maximization of data inclusion with the assurance of its quality. We pinpointed the time window that offered the best alignment — that is, the period which had the highest number of valid measurements across all modalities. This procedure enabled us to determine the optimal time window for downstream analysis, ensuring both a sufficient sample size and a high quality of data. In the next stage, we established a criterion that each day must have at least 60% valid measurements within the identified optimal window for a givenindividual. We based this decision on thresholding approaches consistent with literature utilizing wearables in health research and with the recommendations of the ABCD wearable working group. Participants who did not meet this standard were removed from our dataset. We provide a visual representation of the data processing and QC steps in FIGS. 7A, 11, and 13-14.Imputation

[0167] Missing values are still existent in the resulting QC-controlled time windows. To handle the data missingness, we devised an imputation strategy for categorical and quantitative data modalities, respectively. For the categorical data, we introduced a 'Not Recorded' category into the frame for imputation and subsequently applied label encoding. For the quantitative data, we used the 'drift' method from the sktime package (vO.19.1) with its default settings (Loning, M, Bagnall, A., Ganesh, S., Kazakov, V., Lines, J., andKiraly, F. J. (2019). sktime: A Unified Interface for Machine Learning with Time Series. 33rd Conference on Neural Information Processing Systems (NeurlPS 2019)). Recognizing that these imputation strategies may not be adept at capturing non-polynomial dynamics, we further included an indicator time series for each data modality:indicator (0 = 1(HO = NA)where l(-) is the indicator function and T(i) = NA indicates the data at time step i is missed.

[0168] We concatenated the indicator time series with the imputed time series along the channel dimension. The indicator time series serves as a mask that shows where imputations have been made, while the imputed time series contains both the actual and imputed data. By including this additional indicator time series, we are effectively providing the model with the flexibility to learn an adaptive imputation strategy, where the model can learn how to treat imputed data points based on the surrounding, non-imputed data.

[0169] The indicator time series are referred to as ‘flag’ channels in the model. Every time series modality generates a corresponding binary ‘ _flag’ channel after imputation. For example, ‘heart_rate_flag’ indicates the imputation status of the heart rate data, where 1 indicates that the data at that time step is imputed whereas 0 denotes an actual recorded measurement. Since data missingness could be related to behavioral patterns, missingness indicators can help in capturing this relationship. By encompassing *_tlag channels, the model could recognize andweigh imputed values during training and enables recurrent neural networks to effectively model and handle missing data in clinical time series.Machine Learning Classifier (related to “Generating digital phenotypes from wearable-derived data” and “Classifying psychiatric macrophenotypes using wearable-derived digital phenotypes” in the main text, FIGS. 2, 3, 7A-B, and 8A-B)Problem Formulation

[0170] We first formulated the phenotype classification as a canonical machine learning task with manually engineered eatures, which is outlined as follows. Given an input for a set of features, X, machine learning classification (MLC) targets an output value y which represents the macrophenotype of the subject:X€ RNxdHere, N is the number of individuals and d is the number of features. Specifically, we chose the curated XGBoostRegessor model implemented in the xgboost package (vl.7.5) as our backbone ML models, i.e.:XGBoostRegressor(X) ■— y

[0171] XGBoost (extreme Gradient Boosting) has emerged as an effective machine learning framework, noted for its optimized speed, scalability, and robustness. As a variant of gradient boosted decision trees, XGBoost is tailored for efficiency and demonstrates consistent performance across diverse machine learning applications. Central to XGBoost is its adeptness at engineering trees which pinpoint and rectify residuals from prior iterations, continually refining model accuracy. In this work, we take advantage of the strengths that XGBoost offers, guided by a carefully crafted set of features.Feature Engineering

[0172] Our feature engineering for the XGBoost model is elaborated below. Specifically, the time-invariant wearable features Xwwere primarily derived from summary statistics of the time-series wearable data. We identified seven clusters of time-invariant wearable features from a total of 258 features. We further included curated covariates Xcovas additional features to supplement the time-invariant wearables features. Covariates used for machine learning model include demographic background (sex, race, age at second-year follow-up, divorced parents, parents’ level of education, parent income, adoption) family history of psychiatric illness (bipolardisorder, schizophrenia, antisocial behavior, nervous breakdown, psychiatric treatment, hospital admission, suicide), cognitive scores (flanker test, piememory, process speed, reading score, stop reaction time, etc.), and child behavioral checklist (CBCL internalizing and externalizing scores) (FIG. 2 (Panels A-C), and FIGS. 11-12 and 15). We also considered wear time and sports participation in a subset of our analyses (FIGS. 16-17). Our complete feature set encompasses both covariates and wearable-derived static features:X = Concat({Xcov, vl y2,,. yjpwhere Concat(-) denotes concatenation on the feature dimension. This enabled us to characterize a nuanced interplay of wearable features with individual covariates, substantially accentuating the power of our model.Clustering of Wearable-derived Static Features

[0173] We considered the 258 wearable static features in a subset of 2,410 ABCD individuals with complete genetic, wearable and covariate information. We computed Pearson’s r correlation coefficients between all possible pairs of features and used these correlation values as distance measures to perform hierarchical clustering (R function “hclusf ’ & clustering method “Complete”). We also performed k-means clustering of the correlation matrix by varying the number of clusters from two to twenty (R function “kmeans”, with nstart = 10), and chose an optimal number of seven clusters based on the elbow curve of the total within-cluster sum of squares. A heatmap representation of the seven clusters is shown in FIG. 2 (Panel D), and the list of static features for each cluster is provided in Data S48-S49.Class Balancing

[0174] Imbalanced training labels, where one class substantially outnumbers the other (e.g. 1,737 control individuals versus 216 ADHD individuals), pose a substantial impact on the model performance. To address this issue and ensure a more robust model, we implemented stochastic downsampling techniques on classes with higher representation in each run of the model. To formalize this, we assume two classes, A and B, where |4| and |B| represent the number of instances in each class. Assuming 1^41 > \B |, we calculate the ratio r:r — —|A|We then randomly select a subset A' from A such that:|A'| = r x |J4|The downsampled dataset will then consist of A' and B:Downsampled Dataset = A' U BMultichannel Time Series Classifier (related to “Classifying psychiatric macrophenotypes using wearable-derived digital phenotypes” in the main text, FIGS. 2, 3, 7A-B, and 8A-B) Problem Formulation

[0175] We formulate the phenotype classification as a multichannel time series classification problem which is described as follows. Given an input multichannel time series X:X = Concat{^,ndicator, Xcov},where XwE 5?WxcwxLjx^dlcatorg {o, 1}WXCM / XL, Xcov6WxcwxLHere, N is the number of samples, cwthe number of wearable modalities, L the number of measurements, Xwthe multichannel wearables data, %dlcatorthe multichannel indicator data (See section Imputation), and Xcovcovariates (detail in next section). The multi-channel time series classification (MCTSC) targets an output value y which represents the macrophenotype of the subject:where C - 2 X cw+ cC0Vis the number of input time-series channels. We further define a parameterized model which maps X to the output y:feWwhere f represents the mapping function, which is parameterized by 0. To optimize 0, we employed cross-entropy loss as the objective function, which is defined as:CE(y,y):= -Sfc=iyfclog(fe)where K denotes the number of classes. We employed a label smoothing regularizer to the ground truth label:yfc = yfc(l - a) + JHere, a is a smoothing parameter (we chose 0.1). This label smoothing technique helps to prevent the model from becoming too confident about the class labels, which could potentially bolster its generalization ability.Covariate Integration

[0176] In order to integrate both covariates and time-series data for classification, we adapted the same covariates used for XGBoost feature engineering into a time-series format. Essentially, we transformed these variables into time-invariant sequences, where the value for each covariate remains the same at every time step. The transformed time-series covariates were then merged with wearable sensor data along the channel dimension. This approach allows the model to capture potential interactions between covariates and wearable measures, wherein the model can adjust its weights accordingly if a certain covariate influences the interpretation of the wearable data.Xception Encoder

[0177] The XceptionTime encoder harnesses the power of one-dimensional convolutional neural networks (Id-CNNs) as its underlying architecture. The model is structured with convolutional filters of various sizes, which are sequentially followed by MaxPooling, Batch Normalization, and ReLU activation functions, which form residual connections.Formally:^bottleneck = ConvlDCH^1) ^MaxConvPooi = ConvlD^MaxPooKH'-1)) = ©{©fc{XceptionConvk(H^ottleneck)}, ^MaxConvPool}AH1- BatchNorm(AH*) AH' = ReLU(AH')Hl= + hHlHere, ConvlD(') denotes Id convolution, XceptionConv(-) represents depthwise separable convolution, BatchNorm(-) represents Batch Normalization, ReLU(-) represents ReLU activation function, and © aggregates feature maps from convolution filters of different sizes. A visual representation of the model could be found in FIGS. 2 and 7B. In summary, the input feature maps are first projected to a bottleneck features map where the number of input channels is much larger than the number of output channels. A sequential operation of max pooling and Id convolution is then performed on the input features maps to increase the expressivity of the model. The variation in the size of Xception convolutional filters gives rise to multi-level receptive fields, allowing the model to aggregate and process information at different levels ofgranularity or resolution. Such a property is particularly advantageous when dealing with data from wearable devices, as wearable data often exhibits both local patterns (i.e., minute-by-minute changes) and global trends (i.e., hourly or daily rhythms).

[0178] The XceptionTime encoder introduces a modification to the vanilla Id convolution model by substituting the Id convolution with a Id depth-wise separable convolution. The operation can be broken down into two steps:XceptionConv(Hz x):= PointwiseConv^DepthwiseConv(H( r))

[0179] In contrast to the traditional convolution operation, the depth-wise separable convolution first applies a convolutional filter to each channel individually. This is followed by a 1x1 pointwise convolution module, which performs a linear combination of the outputs across channels. This process reduces the computational complexity of the model while still allowing for complex feature extraction. These steps are described in detail below.

[0180] Depthwise Convolution. This applies a single filter to each input channel which can be expressed as:Fd= Hc* Kdwhere Fdis the output feature map for channel c after the depthwise convolution, Hcis the input feature map for channel c, and Kdis the depthwise filter (or kernel) for channel c. * denotes the convolution operation.

[0181] Pointwise Convolution'. This operation combines the outputs from the depthwise convolution across channels:Fp= ®cYcd* Kpwhere Fpis the output feature map after the pointwise convolution, Ydis the input feature map for channel c, Kpis the pointwise filter, which has a spatial dimension of 1x1 and operates across all channels, and ® is used to denote the aggregation of feature maps from all channels.Model Training and Evaluation (related to “Generating digital phenotypes from wearable-derived data” and “Classifying psychiatric macrophenotypes using wearable-derived digital phenotypes” in the main text, FIGS. 2, 3, 7A-B, and 8A-B)Training Details

[0182] We split the dataset into 70% training set and 30% test set. We ran different iterations with 10 random seeds, and the final results were calculated as the mean of the 10 runs This helps to mitigate the risk of overfitting on a specific split and provides a more robust estimate of the model performance. We used ADAM as the optimizer for training, with 1x1 O'3as the initial learning rate. The neural network model is trained on an NVIDIA VI 0032GB graphical processing unit using the PyTorch and tsai deep learning libraries.Wearable Combination Scores

[0183] In our study, we computed wearable combination scores RS E 7? / VxlJfl by extracting the final layer of the deep learning model, specifically the softmax probability For the XGBoost model, we leverage the pred_prob method implemented in the XGBoost library. Specifically:efjwhere f E HNXKis either the XceptionTime logits in the XceptionTime model, or the sum of outputs from all trees in the XGBoostRegressor model. The softmax function, used in the final layer of the deep learning model, returns probabilities for each category in a multi -class problem that sum up to 1. Similarly, XGBoost's predict_proba method generates class probabilities as output. These scores can serve as a measure of the likelihood associated with each class or outcome. We utilized these scores as the response variable in our subsequent GWAS study (see Methods section “Continuous and binary univariate GWASs for ADHD”). This approach not only bridged the gap between deep learning modeling and GWAS but also significantly enhanced the power of our GWAS.Model Interpretability (related to “Interpreting wearable features prioritized by the deep learning model” in the main text, FIGS. 3 and 8A-B)Ablation Method for Step and Feature Importance

[0184] The ablation method we present was used to measure the importance of features in a dataset. Ablation methods are based on randomly rearranging the values of a feature or agroup of features across all subjects in the dataset, and then calculating an importance score based on the decrease in a chosen metric. In our case, we utilized the Area Under the Receiver Operating Characteristic Curve (AUROC) as the metric to calculate this score. The rationale behind this method is that if a feature is important for model predictions, shuffling the values of that feature should disrupt the model’s ability to make accurate predictions, leading to a decrease in the chosen performance metric. The larger the decrease, the more important the feature is considered to be.

[0185] The ablation importance score can be applied to calculate both feature importance and step importance. For step importance, the implementation is slightly different. Instead of shuffling individual features, we shuffled the values within selected windows of the time series. The time series was divided into windows of a chosen length (in our case, 1 hour), and these windows were then shuffled across all subjects, allowing us to assess the importance of information at different time steps or periods. If the model performance significantly decreases when the values within a certain time window are shuffled, the information within that time window is important for the model predictions. These analyses are shown in FIGS. 23-29. Grad-CAM

[0186] The weighted class activation mapping (CAM) method is a well-established technique for examining how a trained model makes its predictions. In the context of time-series data, it can highlight which time steps are particularly influential in the model's decision-making process.

[0187] We first computed the gradient of the score for the predicted class y with respect to the feature map of the first layer activations ACQ1of a convolutional layer. This gradient, denoted as -^77, provides a measure of how a small change in the activation of the convolutional layer could affect the final prediction of the model. To convert these gradients into a measure of importance for each channel (indexed by c), we employed a global average pooling, which calculates the average of all gradients across the sequence length (indexed by i). This resulted in a set of channel-wise gradient averages, denoted as ac. Mathematically, this is expressed as:where Z is a normalization constant, typically the total number of elements in the layer, and L is the length of the sequence.

[0188] We next generated the Gradient-weighted Class Activation Mapping (Grad- CAM). This is a visual representation of the importance of each time step for the model's predictions. The Grad-CAM, denoted as LlGM, is defined as:£tCM= ReLU(ScMoi)where the ReLU (Rectified Linear Unit) function is used to ensure that only features with a positive influence on the class of interest result in high activation. Essentially, this means that only the time steps that positively contribute to the model's decision will have high importance scores.

[0189] Finally, for each time step, we computed the average Grad-CAM scores across the entire test set. This allowed us to determine which time steps in the input data were most influential in the model predictions.Quality Control of Genetic Data (related to “Using wearable-derived features as digital phenotypes in GWAS for ADHD” and “Employing digital phenotypes to detect genetic associations with behavioral traits” in the main text, FIGS. 1, 4A-C, 5A-C, 9A-C, and I0A-D)

[0190] We obtained genotyped and imputed genetic data for 11,099 individuals as part of the ABCD Data Release 3 (abcdstudy.org). We used the genotyped data to infer population stratification and the imputed data to perform the different GWASs described below (see Methods sections “GWASs for ADHD” and “Continuous Multivariate GWAS for behavioral traits”).

[0191] We applied a quality control (QC) protocol on the genotyped data. Specifically, we performed the QC steps described ingithub.com / MareesAT / GW A_tutorial / blob / master / l_QC_GWAS.zip (file “l_Main_script_QC_GWAS.txt”) using PLINK. Briefly, of the initial set of 516,598 variants, we kept those with a missingness rate across individuals < 0.02 (n = 481,920). Of the initial set of 11,099 individuals, we kept those with a missingness rate across variants < 0.02 (n = 10,660) Next, we considered only variants located on autosomal chromosomes (n = 470,076), those with Minor Allele Frequency (MAF) > 0.01 (n = 427,704), and those that did not deviate from Hardy-Weinberg equilibrium (p value > 10'10; n = 370,002). These variants were pruned to a final set of156,556 variants (window size = 50; number of variants to shift the window at each step = 5; multiple correlation coefficient 0.2). We computed the heterozygosity rate for each individual using the pruned set of variants and kept individuals with a heterozygosity rate deviating < 3 standard deviations from the mean (n = 10,467). We also used pruned variants to assess cryptic relatedness by identifying groups of individuals with Proportion Identity-By-Descent (pi_hat) > 0.2. For every group of related individuals, we then selected the individual with the lowest variant missingness rate, leaving a total of 8,816 individuals. We used PLINK to perform a Principal Component Analysis (PCA) on the 156,556 pruned genotyped variants from the 8,816 selected individuals. We report the PCA results in FIG. 1 (Panel D), where each individual (dot) is colored based on the ethnicity score group information provided by the ABCD metadata, available for 8,791 individuals (see also FIGS.30-33).

[0192] We filtered the imputed genetic variants for MAF > 0.01 and estimated imputation accuracy (R2) > 0.3, and we obtained a final set of 11,954,686 variants for the GWAS analysis (FIG.34). We also computed distributions of R2for all (genotyped and imputed) variants, and of empirical leave-one-out R2(ER2) for genotyped variants (FIG.35).Covariates included in the GWAS (related to “Using wearable-derived features as digital phenotypes in GWAS for ADHD” and “Employing digital phenotypes to detect genetic associations with behavioral traits” in the main text, FIGS. 1, 4A-C, 5A-C, 9A-C, and 10A-D)

[0193] We considered five different groups of covariates: basic (sex, age at second-year follow-up, first five genotype PCs), behavioral (CBCL internalizing and externalizing scores, DSM internalizing and externalizing scores), family history of psychiatric illness (bipolar disorder, schizophrenia, antisocial behavior, nervous breakdown, psychiatric treatment, hospital admission, suicide), family situation (divorced parents, parents’ level of education, family income, adoption), and other (ACS raked propensity score, DNA extraction batch). 3,579 of the previously selected 8,791 individuals reported complete information for these 24 covariates. GWASs for ADHD (related to “Using wearable-derived features as digital phenotypes in GWAS for ADHD” in the main text, FIGS. 4A-C, Table 1, FIGS. 9A-C, and Tables S2-S3)

[0194] For the three GWASs for ADHD described below (continuous multivariate, continuous univariate, and binary univariate), we focused on a subset of 1,191 individuals that were either diagnosed with ADHD (n = 137) or belonged to the non-clinical control group (n =1,054). We used only the set of basic covariates (sex, age, first five population structure PCs), as these were also the covariates used in previous GWASs for ADHD.Continuous multivariate GWAS for ADHD

[0195] In this case we performed a battery of GWAS runs using each of the 7 clusters of 258 wearable-derived static features as a multivariate response variable (i.e., multivariate digital phenotype; FIG. 4B; see Methods section “Clustering of Wearable-derived Static Features”). Since the 258 features are not germane to ADHD, in order to assess the relevance of the significant genetic associations to ADHD, we specifically introduced an interaction term Genotype: Disorder as described in the formula below:Multivariate Digital Phenotype ~ Covariates + Disorder + Genotype + Genotype: Disorder where “Genotype” ( ) corresponds to the genotype group of an individual at a particular genetic variant, and “Disorder” (m) corresponds to the status of the individual (0 = control individual; 1 = individual with ADHD) (FIG. 9A). In this context, a significant interaction (gx m) refers to a genetic effect on the multivariate digital phenotype (J") that differs between individuals with ADHD and those in the control group. Besides using each of the seven clusters of features as a multivariate response variable, we also conducted a GWAS using the 14 heart-related static features as the multivariate response variable (features: InterdayCV, InterdaySD, IntradayCV nean, IntradayCV_median, IntradayCV_sd, IntradayMean nean, IntradayMean nedian, IntradayMean_sd, Intraday SD mean, IntradaySD_median, Intraday SD sd, Mean, Median, STD). Therefore, we ran a total of eight multivariate GWASs (one for heart rate features and one for each of the seven clusters of features).

[0196] We used the Multivariate Asymptotic Non-parametric Test of Association R package (MANTA, https: / / github.com / dgarrimar / manta) to test for association between genetic variants and the multivariate wearable trait. We performed all the analyses within a containerized Nextflow pipeline, available at https: / / github.com / dgarrimar / mvgwas-nf. We employed the option -interaction of the mvgwas-nf pipeline to encode the interaction term in the GWAS. Since MANTA is a non-parametric method, normalization of the multivariate wearable trait was not required. We considered variants that reported a genome-wide significant p value (p < 5 • 1 O’8) for the Genotype: Disorder interaction term. More specifically, for these variants we required both the ADHD and control groups to have at least ten individuals in at least two of the threegenotype groups (in the case of variants located on chromosome X and tested in the male cohort, we required at least ten individuals to be present in both the “0” and “2” genotype groups). Additionally, variants on autosomal chromosomes (or on chromosome X when tested in the female cohort) that reported less than ten individuals in one of the three genotype groups either in the ADHD or control cohorts were tested again for association after excluding the genotype group with less than ten individuals. These variants were reported in the final list of significant loci only if they also reached a Genotype: Disorder p value < 5 10'8in this second association test. After performing the different GWAS runs, we used FUMA for loci definition (reference panel population: “1000GPhase3 ALL”). Results related to the continuous multivariate GWAS for ADHD are shown in FIG. 4B, Table 1 (GWAS Method: “Continuous multivar.” and Phenotype: “ADHD”), FIG. 36, and Table S2. As MANTA p values do not come from a normal distribution, we employed Ax (instead of the commonly used AG) to estimate the genomic inflation factor.

[0197] Previous studies have reported heteroscedasticity as a potential source of p value inflation in GWASs testing SNP-environment interactions compared to the marginal SNP tests. To rule out this possibility in our analyses, we evaluated the heteroscedastic behavior of the two clusters of wearable-derived static features for which we reported genome-wide significant loci (i.e., Clusters 3 and 5; FIG. 36 and Table S2). Specifically, we considered all the genetic variants that reported a GWAS association p value < 10'2for the interaction term Genotype: Disorder. For each of these variants we then performed a permutation test of multivariate homogeneity of variances considering the six Genotype: Disorder groups (i.e., three genotype groups X two disorder categories), using the R package vegan. First, for a given variant and for the six Genotype: Disorder groups of individuals, we computed the distances to the group centroids defined in the space of the covariate-adjusted (i.e., residualized) features (R function betadisper). Next, we performed a permutation-based test to evaluate if one or more groups is more variable than the others (R function permutest, n permutations = 10,000). We found that only 2.4% and 0.83% of variants reported a p value < 10'3(i.e., rejected the null hypothesis of homogeneity of variances) for Clusters 3 and 5, respectively. This suggests that heteroscedasticity is unlikely to significantly bias the association p values for the Genotype: Disorder interaction term in our multivariate GWAS for ADHD.

[0198] Since the multivariate GWAS does not indicate which specific features within a given cluster are significantly associated with the genetic variant, we complemented these results with a post-hoc univariate test (FIG. 9C). For a given genome-wide significant locus, we considered each feature in the significantly-associated cluster and compared the distributions of the feature’s residualized value among the three genotype groups (two-sided Wilcoxon Rank-Sum test). We performed pairwise comparisons among the three genotype groups independently for individuals with ADHD and control individuals. For each of the two genome-wide significant loci, we provide in Data S53 the final list of features that reported a Benjamini -Hochberg adjusted p value < 0.1 in at least one of the three pairwise comparisons performed in the ADHD cohort (see also FIGS. 37-38).Continuous and binary univariate GWASs for ADHD

[0199] We obtained, for each individual, ten different ADHD wearable combination scores based on the XGBoost and Xception predictive models (see Methods section “Wearable Combination Scores”). Specifically, we used combination scores from the following six models: baseline model using CBCL externalizing score (“CBCL ext.”); baseline model using CBCL internalizing score (“CBCL int.”); XGBoost model using wearable features (“XGB”); XGBoost model using wearable features and CBCL scores (“XGB + CBCL”); Xception model using wearable features (“Xception”); and Xception model using wearable features and CBCL scores (“Xception + CBCL”). For models “XGB”, “XGB + CBCL”, “Xception” and “Xception + CBCL”, we also implemented the “liability-CC” trait methodology. This methodology consists of converting the predictive modeling combination score of the cases (i.e., individuals with ADHD) to a value of 1, while keeping the original combination scores for the controls. These four additional types of scores are labeled as “XGB v2”, “XGB + CBCL v2”, “Xception v2” and “Xception + CBCL v2”. We performed ten GWAS runs to test for associations between genetic variants and each of these ten combination scores (continuous univariate GWAS; FIG. 4C). For each run, we defined a model that included the combination score dSCore) as a univariate response variable, and the genotype and covariates as independent variables, as described by the formula below (see also FIG. 9A):Wearable Combination Score ~ Covariates + Genotype

[0200] We also performed a GW AS testing for association between genetic variants and ADHD diagnosis, encoded as a binary outcome (ADHD = 1, control = 0; univariate binary GWAS; FIGS. 4A and 9A), as described by the formula below:ADHD Diagnosis ~ Covariates + Genotype

[0201] We used PLINK to perform both the continuous univariate and the binary univariate GWASs, and the FUMA platform for loci definition (reference panel population: “1000G Phase3 ALL”). Results related to the continuous univariate GWAS for ADHD are shown in FIG. 4C, Table 1 (GWAS Method: “Continuous univar ”), FIGS.39-40, and Table S3Statistical Power of Binary vs. Continuous Traits (related to “Using wearable-derived features as digital phenotypes in GWAS for ADHD” in the main text and FIGS. 9A-C)

[0202] To compare the statistical power of genetic association testing using binary and continuous traits, we simulated a cohort of 1,500 individuals. In each individual i, we generated biallelic SNPs with a binomial model (i.e., the genotype at each SNP followed a binomial distribution, with the number of trials equal to 2 and probability of success on each trial equal to a given MAF). We chose the cohort size to approximate the number of individuals (n = 1,191) in the univariate GWAS for ADHD described above (Methods section “Continuous and binary univariate GWASs for ADHD”). For each individual / , we then simulated a continuous trait (Ci) as the sum of the genotype effect (b) at a given SNP with genotype xt (0, 1, or 2) and random noise (ez):= xrb + efwhereb~N(0,1)e~N(0,1)We also simulated a binary trait (B,) for each individual z, followingBj = l if C[ > median(C), otherwise 0where C is the vector of simulated continuous traits for the entire cohort

[0203] For a particular genotype effect b, we ran 10,000 simulations. Under this scenario, we estimated the power of the simulated continuous and binary traits as the fraction ofsignificant (i.e. Benjamini-Hochberg adjusted p value < 0.05) linear and logistic regression tests, respectively. We employed linear and logistic regression as implemented in the R functions “Im” (library “stats”) and “glm” (family = binomial; library “MASS”), respectively. Overall, we simulated 50 different values of b across six different MAFs (FIG.9C).Continuous multivariate GWAS for behavioral traits (related to “Employing digital phenotypes to detect genetic associations with behavioral traits” in the main text, FIGS. 5A-C, Table 1, FIGS. 10A-D, and Table S4)

[0204] This second type of GWAS consists in testing the association between genetic variants and the multivariate digital phenotype, which we consider a proxy for behavioral patterns in the general population. Therefore, for this GWAS we pooled all the individuals with complete genetic, wearable, and covariate data independently of their diagnosis. Since a more heterogeneous group of individuals was employed in this analysis, we included the full set of 24 covariates in the GWAS model (see Methods section “Covariates included in the GWAS”).

[0205] As in the case of the multivariate GWAS for ADHD, we conducted a battery of eight GWAS runs using as multivariate digital phenotype either the 14 heart-rate related features (available for 3,256 individuals), or the 7 clusters of 258 static features (available for 2,410 individuals). For each run, we defined a model that included the cluster of wearable features as multivariate response variable (cF), and the genotype ( ) and covariates (c) as independent variables, as described by the formula below (FIG. 10A):Multivariate Digital Phenotype ~ Covariates + Genotype

[0206] We performed the GWAS runs with the containerized Nextflow pipeline https: / / github.com / dgarrimar / mvgwas-nf requiring for a given genetic variant a minimum number of ten individuals per genotype group. After performing the different GWAS runs, we used FUMA for loci definition (reference panel population: “1000G Phase3 ALL”). These results are shown in FIG. 5A, Table 1 (GWAS Method: “Continuous multivar.”, Phenotype: “Behavior”), FIG. 42, and Table S4. We complemented the results from the multivariate GWAS with post-hoc univariate tests on the four genome-wide significant loci, as described for the continuous multivariate GWAS for ADHD (FIGS. 5B-C, 9B, and 43-46, and Data S55).Statistical Significance and Functional Dissection of GWAS (related to “Using wearable-derived features as digital phenotypes in GWAS for ADHD” and “Employing digital phenotypes to detect genetic associations with behavioral traits” in the main text, FIGS. 4A-C, 5A-C, 9A-C, andlOA-D)Genome-wide vs. Study-wide Significance

[0207] We selected the conventional genome-wide significant p value threshold of 5- 10'8to identify significant loci from all GWAS runs. However, in line with previous GWAS studies, we also considered a study-wide significance threshold to account for the fact that multiple GWAS runs were performed. In our case, the study -wide significant thresholds are 5 - 1 O'9(5 - 10'81 10 GWAS runs) for the continuous univariate GWAS for ADHD, and 6.25 • 10'9(5 • 10’8 / 8 GWAS runs) for the continuous multivariate GWASs for ADHD and behavioral traits. Based on these thresholds, one locus from the continuous univariate GWAS for ADHD and all four loci from the multivariate GWAS for behavioral traits would pass the study -wide significance threshold. Similar to other GWASs, we also considered a suggestive p value threshold of 1 - 1 O'5(FIGS. 4A-C, 5A, and 39).

[0208] Overall, our hierarchical two-step test of association strategy (i.e., a multivariate GWAS followed by post-hoc univariate tests) offers a number of advantages compared to running multiple univariate GWASs on individual features First, it helps alleviate some of the multiple testing correction burden. If we conducted univariate GWASs for each of the 258 features, the study-wide significance threshold would be 1.94- 10'10(similar to what is reported in a recent study about the gut microbiome that performed univariate GWAS runs on each of 257 metagenomic features). This threshold would be at least one-order of magnitude more stringent than the current study-wide threshold (6.25 10'9; FIG. 9B) Second, the multivariate test also allows for increased association power by leveraging the correlated structure of the wearable features within each cluster.Chromosome X

[0209] In all our analyses the genotypes of variants on chromosome X are encoded for both males and females on a 0...2 scale, with males encoded as either 0 or 2, and females encoded as either 0, 1, or 2, to account for X chromosome inactivation. This is the chromosome X encoding implemented by default in the PLINK2 tool.

[0210] We performed a sex-stratified analysis on chromosome X across all four types of GWASs Specifically, variants located on chromosome X were tested for significant associationsindependently in the male and female cohorts. We employed the same GWAS tools and the same set of covariates as described for autosomal chromosomes in the above sections (“GWASs for ADHD” and “Multivariate GWAS for behavioral traits”), with the exception of sex which was not included as a covariate. Results related to chromosome X GWASs are shown in FIGS. 48-49 and Data S56Neuropsychiatry-related Proximal Genes and eGenes

[0211] For each genome-wide significant locus, we retrieved the ten closest genes when considering a window of ± 250 Kb from the center of the locus, using the GENCODE human genome annotation version 41. Next, we labeled as “neuropsychiatric-related” those proximal genes that are associated with psychiatric disorders according to OpenTargets (https: / / platform.opentargets.org / ). We further intersected our catalog of genome-wide significant loci with previous eQTL catalogs using bedtools intersect (v2.30.0), and identified a subset of proximal neuropsychiatric-related genes with eQTLs overlapping our list of loci. We labeled these genes as “neuropsychiatric-related proximal eGenes” (Table S2-S4).Chromatin Dissection of locus chr 14:23392601-23418974

[0212] We first performed an exploratory analysis by intersecting our two lists of significant loci with the ENC0DE4 registry of candidate cis-regulatory elements (cCREs) (https: / / www.encodeproject.org / search / ?type=Annotation&encyclopedia_version=current&annot ation_type=candidate+Cis- Regulatory+Elements&annotation_type=chromatin+state&annotation_type=representative+DNa se+hypersensitivity+sites&status=released&encyclopedia_version=ENCODE+v4) (Tables S2-S4). Given the documented role of locus chrl4:23392601 -23418974 in heart-related traits and diseases, we next evaluated the enrichment of heart-specific epigenetic features (nucleosome positioning, histone modifications, and transcription factor (TF) binding) at this locus. We downloaded peak calling files for DNase-seq, ATAC-seq, ChlP-seq (histone marks & TFs) and Mint-ChIP for histone marks available for human biosamples from the ENCODE portal (https: / / www.encodeproject.org / metadata / ?control_type%21=%2A&status=released&perturbed= false&assay_title=Histone+ChIP-seq&assay_title=TF+ChIP-seq&assay_title=DNase-seq&assay_title=ATAC-seq&assay_title=Mint-ChIP-seq&files.file_type=bigBed+narrowPeak&replicates. library. biosample. donor, organism. scientific _name=Homo+sapiens&type=Experiment&files.analyses.status=released&files.preferred_default=true; access date: 09 / 27 / 2022) We then grouped human biosamples based on their “biosample ontology organ slim” (https: / / www.encodeproject.org / report / 9type=Experiment&control_type!=*&status=released&pe rturbed=false&assay_title=TF+ChIP-seq&assay_title=Histone+ChIP-seq&assay_title=DNase-seq&assay_title=ATAC-seq&assay_title=Mint-ChIP-seq&replicates.library. biosample, donor, organism. scientific_name=Homo+sapiens&field=biosam ple_ontology.organ_slims&field=biosample_ontology.cell_slims&field=biosample_ontology.sys tem_slims&field=%40id«&;field=biosample_ontology.term_name). To test the tissue-specific enrichment of chromatin features in a particular organ, we computed the number of times any of the five significant variants at the locus overlapped a peak from experiments in that organ compared to all other organs (two-sided Fisher’s exact test, Benjamini -Hochberg adjusted p value < 0.1) (FIG. 10B). For this analysis, we counted only once those overlaps involving variants that are < 100 bp apart.Exploring the Genetic-behavioral-psychiatric Axis

[0213] For each of the two examples shown in FIGS. 5B-C (left panels), we evaluated the enrichment of the minor allele (in both cases the G allele) in individuals within a specific psychiatric cohort vs. non-clinical control individuals (two-sided Fisher exact test; FIGS. 5B-C right panels). Given the reduced number of individuals with GG genotype for SNP rsl 13525298 n = 15), the enrichment of the minor allele was tested by merging individuals with AG and GG genotypes For SNP rs365990, the enrichment was computed only in individuals with GG genotype. For all tests, we required at least one individual to be present in every cell of the 2x2 matrix employed for the Fisher’s exact test (a = n individuals with minor allele AND part of the psychiatric cohort; b = n individuals without minor allele AND part of the psychiatric cohort; c = n individuals with minor allele AND part of the healthy controls; d= n individuals without minor allele AND not part of the healthy controls).Intersection of genome-wide significant loci with the GWAS Catalog

[0214] To assess the clinical relevance of our GWAS loci, we intersected them with variants from the NHGRI-EBI GWAS catalog (https: / / www.ebi.ac.uk / gwas / ; access date:05 / 16 / 2023). For the loci identified by the GWASs for ADHD (both continuous multivariate and continuous univariate), we considered overlaps with brain- or neuropsychiatric-related GWAS hits (FIGS. 4B-C and 41). For the loci identified by the multivariate GWAS for behavioral traits,we considered any overlaps with heart-, sleep-, metabolism- or physical activity-related GWAS hits, since our wearable-derived features are mostly related to measurements of heart rate, sleep quality, metabolic intake and physical movement. Additionally, given the presence of individuals with psychiatric disorders in the ABCD cohort, we also considered intersections with brain- or neuropsychiatric-related GWAS hits (FIGS.5A and 47). We acknowledge that colocalization analysis would be the most appropriate way to compute these intersections, and we performed this analysis for loci identified by the continuous univariate GWAS for ADHD (see Methods section “Colocalization analysis”). However, MANTA does not provide estimates of variant effect sizes that can be directly employed in co-localization analysis. For this reason, we evaluated the strength of these intersections against a null distribution. Specifically, we computed the proportion of GWAS variants associated with a particular trait that overlap our significant loci, and compared it to the proportions observed across 10,000 random sets of genomic loci with the same size and chromosome location. We report the percentile of our GWAS enrichments compared to the null distribution in FIG. 41 (univariate GWAS for ADHD) and FIG. 47 (multivariate GWAS for behavioral traits).Colocalization Analysis

[0215] We performed colocalization analysis using the R package coloc (function coloc.abf, default parameters) on the results obtained from the continuous univariate GWAS for ADHD. Specifically, we focused on two of the seven overlapping brain-related traits with available GWAS summary statistics (FIG. 4C and Data S54), and tested the hypothesis of signal co-localization between our ADHD wearable combination scores at the intersecting loci. Locus chrl7:32256997:32283356 reported a posterior probability of 0.99 for signal co-localization with a locus previously associated with chronotype measurement. We also tested locus chr7:68219282:68338849 (suggestive association at p value < 10'5) for co-localization with a previously reported locus for ADHD. In this case, given that the two traits being tested are the same, we set all three parameters pl, p2 and p 12 equal to 1 • 10'5, and reported a posterior probability of 0.25.Additional DataData S40. Summary of various covariates used in downstream modeling of behavioral phenotypes, related to Figure 2 and STAR Methods “Machine Learning Classifier” Section Cavariate Category More DetailsSex M / FRace / Ethnicity Categorical Race / EthnicityAge Numeric AgeParent Divorce Status Divorce StatusParent Grade Parent GradeFamily Income Numeric incomeBipolarSchizophreniaAntisocialFamily History NervesTreatmentHospitalSuicideAdopted AdoptedstopsigreactiontimestopsignolgortstandevgortpicvocabCognitive Test ScoresflankerprecessspeedpiememoryreadingscoreChild Behavior Checklist, CBCL CBCL score

[0216] This table presents an array of covariates utilized in the downstream modeling of behavioral phenotypes. It includes demographics, family history of psychopathology, cognitive assessment data and CBCL scores which are utilized to assess internalizing and externalizing psychopathology in youth through a standardized self-report measure known as the Child Behavior Checklist (CBCL).Data S41. Description of cognitive scores, related to Figure 2 and STAR Methods “Machine Learning Classifier” SectionLabe! Test Outcome stopsigreactiontime Stop signal reaction time stopsignolgort Stop Signal Task Go reaction time standevgort Standard deviation reaction time picvocab NIH Toolbox Picture Vocabulary Test Age corrected score flanker NfH Toolbox Flanker Inhibitory Control arid Attention Test Age corrected score precessspeed NIH Toolbox Pattern Comparison Processing Speed Age corrected score piememory NIH Toolbox Picture Sequence Memory Test Age corrected scorereadingscore NIH Toolbox Oral Reading Recognition Test Age corrected score

[0217] This table shows the cognitive scores considered in our study. It details the specific test from which each score is generated.Data S42. Sports / Activity participation mapping table, related to Figure 2 and STAR Methods “Machine Learning Classifier” SectionIndex SportsName1 Ballet, Dance2 Baseball, Softball3 Basketbail4 Climbing5 Field Hockey6 Football7 Gymnastics8 ice Hockey9 Horseback Riding, Polo10 Ice or Inline Skating11 Martial Arts12 Lacrosse13 Rugby14 Skateboarding15 Skiing, Snowboarding16 Soccer17 Surfing18 Swimming, Water Polo19 Tennis20 Track, Running, Cross-country21 Mixed Martial Arts22 Volleyball23 Yoga, Tai Chi24 Musical Instrument (Singing, Choir, Guitar, Piano, Drums, Violin, Flute, Band, Rock Band, Orchestra) 25 Drawing, Painting, Graphic Art, Photography, Pottery, Sculpting26 Theater, Acting, Film27 Crafts like Knitting, Building Model Cars or Airplanes28 Competitive Games like Chess, Cards, or Darts29 Hobbies like collecting stamps or coins

[0218] This table provides a mapping of encoded labels to corresponding sports / activity utilized in our study.Data S43. Mann-Whitney U Test Results for Wear Time Comparison Among Different Groups, related to Figure 2 and STAR Methods “Machine Learning Classifier” Section Comparison Groups U-statistic p-valueBipalar / Psychotic vs Control 60550.50 0.52Anxiety v$ Control 323598.50 0.41Depression vs Control 5719.00 0.15Eating vs Control 24894.50 0.99OCD vs Control 79254.00 0.16PTSD vs Control 2638.50 0.36Sleep vs Control 26560.00 0.92ADHD vs Control 111752,00 0.39Panic vs Control 44771,50 0.42

[0219] This table shows the results of the Mann-Whitney U test comparing wear time of the Fitbit device among various phenotypic groups and the control group. The U-statistic and p values indicate that there is no statistically significant difference between the control group and any phenotypic groups. We also performed a Kruskal -Wallis test comparing the wear time across all groups, which resulted in no significant differences (p value: 0.538).Data S44. Mann-Whitney U Test Results for Sports and Activities Participation Comparison Among Different Groups, related to Figure 2 and STAR Methods “Machine Learning Classifier”SectionComparison Groups U -statistic p-valueBipolar / Psychotic vs Control 108837.50 0.751Anxiety vs Control 305061.00 0293Depression vs Control 36046.00 0.053Eating vs Control 78947.00 0.140OCD vs Control: 25763.50 0.836PTSD vs Control 61360.00 0.382Sleep vs Control 27191.50 0.341ADHD vs Control 6815.50 0.550Panic vs Control 3486.00 0.975

[0220] This table shows the results of the Mann-Whitney U test comparing sports and activities participation (the number of different sports / activities an individual participated in) of individuals among various phenotypic groups and the control group. The U-statistic and p valuesindicate that there is no statistically significant difference between the control group and any of the psychiatric groups. We also performed a Kruskal-Wallis test comparing sports and activities participation across all groups, which indicated no significant differences (p value: 0.405).Data S45. Fitbit Daily Physical Activity Features Definitions, related to Figure 1, Figure 2, and STAR Methods “Dataset Description” SectionFeature Name Feature Descriptionfit_ss_protocol_startdate First day of (expected) wear of 3 week protocolfit_ss_protocol_wear Is this day during the 3 week protocol period?fit_ss_wear_date Date of Participant WearTotal number of steps observed in all minutes (no exclusions) from midnight (00:00) to fit_ss_day_total_steps_no_el11: 59PM (23:59) regardless of sleep categorizationNumber of minutes with heart rate value from midnight (00:00) to 11:59P (23:59) not fit_ss_day_minidentified as sleep (at 60 second “classic” level) or exclusion rulesNumber of minutes with Heart Rate value from midnight (00:00) to 11: 59P (23: 59) that fit_ss_night_minARE identified as sleep (at 60 second “classic” level) or exclusion rulesNumber of minutes with Heart Rate value from midnight (00:00) to 11: 59P (23: 59) that ARE identified as sleep (at 60 second “classic” level) or exclusion rules from the first fit_ss_sleep_minminute of sleep on the day in question to the first minute of awake on the next day (i e. crosses midnight)fit_ss_30_second_data_existed Does 30 second sleep data exist for this dayNumber of minutes from midnight (00:00) to 11: 59P (23:59) not identified as sleep (at 60 fit_ss_excl_day_minsecond “classic” level) excluded for any reasonNumber of minutes from midnight (00:00) to 11:59P (23:59) not identified as sleep (at 60 fit_ss_excl_day _min_hr50second “classic” level) excluded because HR was lower than 50 bpmNumber of minutes from midnight (00:00) to 11:59P (23:59) not identified as sleep (at 60 fit_ss_excl_day_min_nohrsecond “classic” level) excluded because there was no HR value for the given minute Number of minutes from midnight (00:00) to 11:59P (23:59) not identified as sleep (at 60 fit_ss_excl_day_min_hr_rept second “classic” level) excluded because there were identical HR values repeated for 10+ instancesNumber of minutes from midnight (00:00) to 11:59P (23:59) identified as sleep (at 60 fit_ss_excl_night_minsecond “classic” level) excluded for any reasonNumber of minutes from midnight (00:00) to 11:59P (23:59) identified as sleep (at 60 fit_ss_excl_night_min_hr50second “classic” level) excluded because HR was lower than 50 bpmNumber of minutes from midnight (00: 00) to 11: 59P (23:59) identified as sleep (at 60 fit_ss_excl_night_min_nohrsecond “classic” level) excluded because there was no HR value for the given minute Number of minutes from midnight (00:00) to 11:59P (23:59) identified as sleep (at 60 fit_ss_excl_night_min_hr_rept second “classic” level) excluded because there were identical HR values repeated for 10+ mstancesNumber of minutes from the first minute of sleep on the day in question to the first minute fit_ss_excl_sleep_minof awake on the next day excluded for any reason (these data cross midnight) Number of minutes from the first minute of sleep on the day in question to the first minute fit_ss_excl_sleep_min_hr50 of awake on the next day excluded because HR was lower than 50 bpm (these data cross midnight)Number of minutes from the first minute of sleep on the day in question to the first minute fit_ss_excl_sleepjnin_nohr of awake on the next day excluded because there was no HR value for the given minute (these data cross midnight)Number of minutes from the first minute of sleep on the day in question to the first minute fit_ss_excl_sleep_min_rept of awake on the next day excluded because there were identical HR values repeated for 10+ instances (these data cross midnight)fit_ss_day_min_gt_600 Does this day have >599 minutes of non-sleep wear after all exclusionsDoes this day have >299 minutes of sleep wear (based on 60 second data from fitabase) fit_ss_sleep_min_gt_300after all exclusionsfit_ss_first_hr_date First day that HR appears in fitabase recordfit_ss_weekday Day of the weekfit_ss_wkno Week number since start of protocolfit_ss_weekend_md Is this day a weekend (Saturday or Sunday)Total number of valid minutes after all exclusions from midnight (00:00) to 11:59 PM fit_ss_total_min(23:59) regardless of sleep statusTotal number of steps observed after all exclusions from midnight (00:00) to 11:59 PM fit_ss_total_step(23:59) regardless of sleep statusAverage METS / minute of all valid minutes after all exclusions from midnight (00:00) to fit_ss_total_ave_met11:59 PM (23:59) regardless of sleep statusNumber of minutes of sedentary (<1 5 METS) time observed in all valid minutes after all fit_ss_total_sedentary_mmexclusions from midnight (00: 00) to 11:59 PM (23: 59) regardless of sleep status Number of minutes of lightly active time (1.5-29 METS) observed in all valid minutes fit_ss_total_light_active_minafter all exclusions from midnight (00:00) to 11:59 PM (23:59) regardless of sleep status Number of minutes of moderately active (3-5.9 METS) time observed in all valid minutes fit_ss_total_fairly_active_minafter all exclusions from midnight (00:00) to 11:59 PM (23:59) regardless of sleep status Number of minutes of vigorously active (>6 METS) time observed in all valid minutes fit_ss_total_very_active_minafter all exclusions from midnight (00:00) to 11:59 PM (23:59) regardless of sleep status fit_ss_fitbit_totalsteps Fitbit based number of steps for the day FROM DAILY LEVEL SUMMARY Fitbit based number of minutes spent in sedentary (<1.5 METS) time for the day FROM fit_ss_fitbit_sedentaryminDAILY LEVEL SUMMARYFitbit based number of minutes spent in light activity (1 5-29 METS) for the day FROM fit_ss_fitbit_lightlyactiveminDAILY LEVEL SUMMARYFitbit based number of minutes spent in moderate activity (3-59 METS) for the day FROM fit_ss_fitbit_fairlyactivemmDAILY LEVEL SUMMARYFitbit based number of minutes spent in vigorous (>6 METS) for the day FROM DAILY fit_ss_fitbit_veryactiveminLEVEL SUMMARYfit_ss_fitbit_restingheartrate Fitbit based number restmg heart rate for the day FROM DAILY LEVEL SUMMARY Is the QC’d step value after all exclusions 80%+ the value of the daily level as reported by fit_ss_mstep_lt_8O_dailystepfitbitfit_ss_dayt_total_steps Total number of steps observed during non-sleep (night) valid minutes fit_ss_dayt_ave_met_value Average METS / minute during non-sleep (night) valid minutesNumber of minutes of sedentary (<1 5 METS) time observed during non-sleep (night) valid fit_ss_dayt_sedentary_minminutesNumber of minutes of lightly active time (1.5-29 METS) observed during non-sleep fit_ss_dayt_light_active_min(night) valid minutesNumber of minutes of moderately active (3-5.9 METS) time observed during non-sleep fit_ss_dayt_fairly_active_min(night) valid minutesNumber of minutes of vigorously active (>6 METS) time observed during non-sleep fit_ss_dayt_very_active_min(night) valid minutes

[0221] The table above provides descriptions of the features collected by the FitBit device. This table contains physical activity metrics, heart rate data, time-specific measurements, and quality control features. It also shows data reflecting subjects’ daily routines, such as periods of sleep and activity and the intensity of activities is classified into sedentary, light, moderate, and vigorous. The metadata including parameters or identifiers related to data collection are also included.

[0222] While many of the technical details of the algorithms used to derive behavioral and physiological features from raw sensor data are proprietary and not publicly available, academic researchers have conducted validation studies to assess the concordance between metrics derived from consumer-grade wearables and established gold-standard assessment techniques thought to measure similar constructs. Systematic reviews of this literature generally conclude that consumer-grade wearables show acceptable, if not clinically interpretable, accuracy versus gold-standard measurement methods for specific measurement categories (i.e., step count, heart rate), although performance varies across devices.

[0223] METs are a standardized measure used to express the energy cost / intensity of physical activities (PAs) as multiples of a standard resting metabolic rate (RMR). One MET is defined as the energy cost of a person at rest, which is calculated by the amount of oxygen consumed (3.5 mL 02 / kg / min) and it is equivalent to a caloric consumption of 1 kcal / kg / hour. METs provide a quantitative way to conveniently compare energy expenditure across various activities, ranging from sleeping (approximately 0.9 METs) to vigorous activities like running (over 6 METs). METs could be broadly categorized into light, moderate, and vigorous intensity groups, or used continuously to provide a more precise measure of activity intensity. Although limitations of METs as a direct measure of energy expenditure during physical activity have been noted in the literature2, it is frequently used as an index of physical activity level useful in the context of health research involving wearables.

[0224] Despite potential limitations to the criterion-related validity of consumer-grade wearable devices in terms of their alignment with gold-standard assessment methods, there isample research demonstrating the potential of wearables and similar, consumer-grade devices, to capture nuanced patterns of physiology and behavior in the context of mental health research.Data S46. Fitbit Daily Sleep Summary Features Definitions, related to Figure 1, Figure 2, and STAR Methods “Dataset Description” SectionFeature Name Feature Descriptionfit ss sleepdate Date of the night (PM portion of sleep period) of sleep being considered fit ss wkno Week number since start of protocolfit ss weekday Day of the weekfit ss weekend ind Is this day a weekend (Saturday or Sunday)fit_ss_protocol_date First day of (expected) wear by 3 week protocolfit_ss_protocol_wear Is this day during the 3 week protocol period?First Minute that the participant was identified as being in-bed (but not necessarily fit ss first inbed minutesasleep)fit ss first sleep minutes First Minute that the participant was identified as being asleepFirst Minute that the participant was identified as being out-bed (but may have fit ss outbed minuteswoken up prior to this)fit ss wakeup minutes First Minute that the participant was identified as being awakeTotal time within the sleep window (including wake periods) as calculated by 30 fit ss total sleep minutessecond sleep recordfit ss sleepperiod minutes Total time in sleep (light plus deep plus REM)Total number of Minutes (in 30 second increments) identified as being awake fit ss wake minutesduring the sleep windowTotal number of Minutes (in 30 second increments) identified as being in “light” fit ss light minutessleep stage during the sleep windowTotal number of Minutes (in 30 second increments) identified as being in “deep” fit ss deep minutessleep stage during the sleep windowTotal number of Minutes (in 30 second increments) identified as being in “REM” fit ss rem minutessleep stage during the sleep windowTotal number of unique awakenings during the sleep period (excludes final fit ss wake countawakening)fit ss avg hr wake Average heart rate during Minutes that were classified as awake fit ss avg hr light Average heart rate during Minutes that were classified as light sleep stage fit ss avg hr deep Average heart rate during Minutes that were classified as deep sleep stage fit ss avg hr rem Average heart rate during Minutes that were classified as REM sleep stage

[0225] This table provides descriptions of features related to sleep collected from the wearable Fitbit devices used in the study.Data S47. 258 Wearable-derived Static Features, related to Figure 2 and STAR Methods “Machine Learning Classifier” SectionIndividual-level Features Population-level Summary StatisticsMean SD Media Min Max Range Ske Kurtos SE n w is fit ss avg hr deep max 81.23 10.73 81.00 50.00 139.00 89.00 0.42 0.68 0.17 fit ss avg hr deep mean 72.30 8.58 72.11 47.75 104.67 56.92 0.18 0.00 0.14 fit ss avg hr deep median 71.90 8.65 72.00 47.00 106.00 59.00 0.19 -0.01 0.14 fit ss avg hr deep min 65.42 8.35 65.00 42.00 101.00 59.00 0.27 0.09 0.13 fit ss avg hr deep sd 4.80 2.11 444 0.00 19.80 19.80 1.70 6.03 0.03 fit ss avg hr light max 76.55 10.03 76.00 48.00 120.00 72.00 0.57 0.82 0.16 fit ss avg hr light mean 69.44 7.78 69.15 46.88 99.33 52.46 0.22 0.00 0.13 fit ss avg hr light median 69.09 7.76 69.00 47.00 101.00 54.00 0.23 0.02 0.13 fit ss avg hr light min 64.18 7.49 64.00 43.00 98.00 55.00 0.29 0.11 0.12 fit ss avg hr light sd 3.74 1.89 3 33 0.00 22.63 22.63 2.26 9.84 0.03 fit ss avg hr rem max 79.21 9.45 79.00 52.00 124.00 72.00 0.51 0.78 0.15 fit ss avg hr rem mean 72.35 7.58 72.33 48.45 106.64 58.19 0.07 -0.02 0.12 fit ss avg hr rem median 72.07 7.58 72.00 48.00 105.50 57.50 0.06 -0.06 0.12 fit ss avg hr rem min 67.03 7.64 67.00 42.00 100.00 58.00 0.10 0.01 0.12 fit ss avg hr rem sd 3.66 1.80 3 27 0.00 20.51 20.51 2.41 11.34 0.03 fit ss avg hr wake max 81.09 10.24 80.00 52.00 129.00 77.00 0.69 1.23 0.17 fit ss avg hr wake mean 73.41 7.63 73.24 50.82 104.07 53.25 0.18 -0.03 0.12 fit ss avg hr wake median 73.03 7.63 73.00 51.00 105.00 54.00 0.20 0.02 0.12 fit_ss_avg_hr_wake_min 67.85 7.34 68.00 48.00 98.00 50.00 0.24 0.00 0.12 fit ss avg hr wake sd 3.96 2.01 3 52 0.00 21.92 21.92 2.27 9.30 0.03fit_s s_day _min_max 1027.1 115.4 1010.0 371.0 1440.0 1069.0 0.85 3.20 1.863 0 0 0 0 0 fit ss day min mean 801.21 83.20 814.50 225.5 1009.4 783.94 4.01 1.340 4 1.41fit ss day min median 839,29 89,73 855,50 158,0 1082,0 924,00 7.65 1.450 0 2.12 fit ss day min min 388.14 181.1 380.00 1.00 924.00 923.00 0.20 -0.54 2.929fit_ss_day_min_sd 177.35 62.94 173.11 7.78 463.15 455.38 0.54 0.64 1.02 fit_ss_day_total_steps_no_el_ ###### 5583. ##### 1651. ##### ##### 0.80 1.23 90.0 max # 22 ## 00 ## ## 4 fit_ss_day_total_steps_no_el_ 8907.7 3028. 8567.7 1095. ##### ##### 0.67 0.84 48.8 mean 8 71 5 25 ## ## 4 fit_ss_day_total_steps_no_el_ 8833.2 3232. 8543.0 447.0 ##### ##### 0.60 0.83 52.1 median 4 20 0 0 ## ## 3 fit ss day total steps no el 2645.4 2126. 2095.0 0.00 ##### ##### 1.76 5.23 34.2 min 1 04 0 ## ## 9 fit_ss_day_total_steps_no_el_s 3878.8 1479. 3663.1 112.4 ##### ##### 1.02 2.10 23.8 d 6 15 0 3 ## ## 5 fit_ss_dayt_ave_met_value_ma 2.66 0.46 262 1.32 5.72 4.40 0.86 1.65 0.01Xfit_ss_dayt_ave_met_value_me 2.04 0.26 202 1.28 4.06 2.78 0.66 1.61 0.00 anfit_ss_dayt_ave_met_value_me 2.02 0.27 200 1.29 3.99 2.71 0.62 1.40 0.00 dianfit_ss_dayt_ave_met_value_mi 1.53 0.24 151 1.00 3.90 2.90 1.00 4.16 0.00 nfit_ss_dayt_ave_met_value_sd 0.32 0.12 030 0.01 1.36 1.35 1.36 4.27 0.00 fit_ss_dayt_farily_active_min_ 77.45 50.57 67.00 0.00 384.00 384.00 1.33 2.82 0.82 maxfit_ss_dayt_farily_active_min_ 23.69 19.39 19.00 0.00 175.50 175.50 1.71 4.56 0.31 meanfit_ss_dayt_farily_active_min_ 17.59 19.78 12.00 0.00 175.50 175.50 1.96 5.96 0.32 medianfit_ss_dayt_farily_active_min_ 1.02 5.51 000 0.00 120.00 120.00 9.98 139.6 0.09 min 2 fit_ss_dayt_farily_active_min_ 23.22 14.37 20.43 0.00 112.84 112.84 1.31 2.98 0.23 sdfit ss dayt light active min 399,51 85,19 395,00 91,00 1018,0 927,00 0,36 1.10 1,37 max 0 fit_ss_dayt_light_active_min_ 262.09 57.85 261.61 54.25 558.50 504.25 0.09 0.24 0.93 meanfit_ss_dayt_light_active_min_ 268.20 62.50 269.00 24.50 587.00 562.50 0.00 0.27 1.01 medianfit_ss_dayt_light_active_min_ 102.14 62.36 94.00 0.00 513.00 513.00 0.79 0.92 1.01 minfit_ss_dayt_light_active_min_s 84.24 26.09 81.85 2.12 229.10 226.98 0.84 2.10 0.42 dfit ss dayt sedentary min ma 728.33 149.9 711.00 216.0 1440.0 1224.0 1.18 3.35 2.42X 9 0 0 0 fit_ss_dayt_sedentary_min_me 505.65 83.71 506.14 159.0 819.33 660.33 0.59 1.35 an 0 0.09 fit ss dayt sedentary min me 516.97 90.86 520.00 127.5 885.00 757.50 0.68 1.47 dian 0 0.19 fit_ss_dayt_sedentary_min_mi 236.52 109.5 228.00 0.00 717.00 717.00 0.37 -0.07 1.77 n 0fit_ss_dayt_sedentary_min_sd 135.89 44.88 130.32 4.95 388.82 383.87 0.94 2.11 0.72 fit ss dayt total steps max ###### 5576. ##### 1651. ##### ##### 0.81 1.26 89.9# 56 ## 00 ## ## 3 fit ss dayt total steps mean 8835.9 3014. 8499.7 1063. ##### ##### 0.67 0.87 48.60 11 1 25 ## ## 1 fit ss dayt total steps median 8764.1 3220. 8460.0 402.0 ##### ##### 0.61 0.84 51.95 75 0 0 ## ## 4 fit ss dayt total steps min 2564.1 2101. 2024.0 0.00 ##### ##### 1.78 5.39 33.85 28 0 ## ## 9 fit ss dayt total steps sd 3886.3 1481. 3664.0 107.4 ##### ##### 1.02 2.08 23.97 90 8 8 ## ## 0fit_ss_dayt_very_active_min_ 38.15 38.95 26.00 0.00 345.00 345.00 2.02 5.88 0.63 maxfit_ss_dayt_very_active_min_ 9.78 12.56 524 0.00 146.33 146.33 2.68 10.84 0.20 meanfit ss dayt very active min 6,11 11,33 1,00 0,00 154,00 154,00 3.49 18.75 0,18 medianfit_ss_dayt_very_active_min_ 0.27 2.45 000 0.00 83.00 83.00 18.3 466.0 0.04 min 4 3 fit_ss_dayt_very_active_min_s 11.39 11.51 790 0.00 99.89 99.89 2.06 6.37 0.19 dfit ss deep minutes max 125.64 21.54 125.00 27.00 249.50 222.50 0.13 1.60 0.35 fit ss deep minutes mean 88.98 14.73 89.28 24.25 135.33 111.08 0.29 0.240.23 fit_ss_deep_minutes_median 89.09 15.61 89.50 24.25 145.50 121.25 0.29 0.250.20fit ss deep minutes min 51.77 19.08 52.00 2.00 119.50 117.50 0.02 -0.17 0.31 fit ss deep minutes sd 22.50 6.92 21.94 0.00 87.58 87.58 1.16 6.30 0.11 fit_ss_excl_day_min_hr50_ma 1.91 10.02 000 0.00 231.00 231.00 10.2 143.3 0.16X 3 7 fit_ss_excl_day_min_hr50_me 0.30 2.13 000 0.00 52.33 52.33 13.5 237.8 0.03 an 6 5 fit_ss_excl_day_min_hr50_me 0.11 1.31 000 0.00 43.50 43.50 20.2 511.4 0.02 dian 0 3 fit ss excl day min hr50 min 0.00 0.18 000 0.00 10.00 10.00 49.2 2668. 0.004 65 fit_ss_excl_day_min_hr50_sd 0.55 2.96 000 0.00 50.83 50.83 10.0 125.0 0.052 9 fit_ss_excl_day_min_hr_rept_ 364.89 323.4 262.00 0.00 1400.0 1400.0 0.90 -0.10 5.22 max 3 0 0 fit_ss_excl_day_min_hr_rept_ 58.43 76.45 32.11 0.00 718.00 718.00 2.82 11.26 1.23 meanfit_ss_excl_day_min_hr_rept_ 15.20 56.81 000 0.00 718.00 718.00 5.96 42.09 0.92 medianfit_ss_excl_day_min_hr_rept_ 1.23 16.87 000 0.00 453.00 453.00 19.1 418.4 0.27 min 5 5 fit_ss_excl_day_min_hr_rept_s 109.29 105.0 73.27 0.00 753.78 753.78 1.31 1.69 1.69 d 9fit_ss_excl_day_min_max 871,43 234,5 892,00 0,00 1400,0 1400,0 1.13 3,784 0 0 0.84 fit ss excl day min mean 199.83 133.9 163.17 0.00 918.00 918.00 1.70 3.50 2.168fit_ss_excl_day_min_median 105.43 158.0 42.00 0.00 1028.0 1028.0 2.46 6.16 2.550 0 0 fit ss excl day min min 11.47 59.41 000 0.00 735.00 735.00 7.71 66.47 0.96 fit_ss_excl_day_min_nohr_ma 776.26 255.2 803.00 0.00 1392.0 1392.0 0.45 4.12X 9 0 0 0.68 fit_ss_excl_day_min_nohr_me 141.10 92.11 117.28 0.00 829.00 829.00 2.04 6.31 1.49 anfit ss excl day min nohr me 49.98 94.98 21.00 0.00 834.00 834.00 4.07 19.42 1.53 dianfit_ss_excl_day_min_nohr_min 3.47 28.23 0.00 0.00 652.00 652.00 14.7 257.8 0.463 0 fit_ss_excl_day_min_nohr_sd 229.51 92.94 224.82 0.00 813.17 813.17 0.62 2.49 1.50 fit ss excl day min sd 262.10 94.37 255.33 0.00 773.57 773.57 0.54 1.85 1.52 fit_ss_excl_night_min_hr50_m 17.57 57.79 000 0.00 618.00 618.00 4.72 25.74 0.93 axfit ss excl night min hr50 m 4.72 22.27 000 0.00 378.32 378.32 8.07 84.41 0.36 eanfit_ss_excl_night_min_hr50_m 3.12 20.84 000 0.00 410.50 410.50 10.2 131.5 0.34 edian 6 7 fit_ss_excl_night_niin_hr5O_m 0.19 3.92 000 0.00 175.00 175.00 31.8 1214. 0.06 in 7 03 fit_ss_excl_night_min_hr5O_sd 5.24 17.75 000 0.00 199.10 199.10 4.86 27.11 0.29 fit_ss_excl_night_min_hr_rept 77.97 73.21 58.00 0.00 759.00 759.00 2.35 8.39 1.18 maxfit_ss_excl_night_min_hr_rept 15.06 15.98 10.64 0.00 239.00 239.00 3.22 20.38 0.26 meanfit_ss_excl_night_min_hr_rept 5.98 13.31 000 0.00 239.00 239.00 5.25 50.23 0.21 _medianfit ss excl night min hr repl 0,05 0,95 0.00 0.00 32.00 32,00 23.0 605.9 0,02 min 3 0 fit_ss_excl_night_min_hr_rept 23.26 21.81 17.48 0.00 302.64 302.64 3.05 18.21 0.35_sdfit_ss_excl_night_min_max 110.61 99.42 79.00 0.00 759.00 759.00 1.82 3.77 1.60 fit ss excl night min mean 24.41 28.59 16.59 0.00 398.08 398.08 4.23 30.13 0.46 fit_ss_excl_night_min_median 13.24 27.84 3 00 0.00 443.50 443.50 5.73 50.56 0.45 fit ss excl night min min 0.39 5.04 0.00 0.00 175.00 175.00 22.8 631.7 0.089 0 fit ss excl night min nohr m 41.19 72.22 20.00 0.00 595.00 595.00 3.85 16.88 1.16 axfit ss excl night min nohr m 4.82 7.52 2.19 0.00 121.53 121.53 4.08 31.35 0.12 eanfit ss excl night min nohr m 0.59 2.75 000 0.00 65.00 65.00 9.09 129.3 0.04 edian 8 fit ss excl night min nohr m 0.01 0.31 000 0.00 19.00 19.00 61.5 3802. 0.00 in 1 78 fit ss excl night min nohr sd 11.22 18.89 5 68 0.00 186.39 186.39 3.95 19.48 0.30 fit ss excl night min sd 32.52 28.43 24.19 0.00 302.64 302.64 2.15 7.34 0.46 fit_ss_excl_sleep_min_hr50_m 19.85 64.02 000 0.00 810.00 810.00 4.65 25.80 1.03 axfit_ss_excl_sleep_min_hr50_m 5.67 26.54 000 0.00 445.78 445.78 8.05 84.00 0.43 eanfit_ss_excl_sleep_min_hr50_m 3.96 25.13 000 0.00 457.00 457.00 9.83 119.9 0.41 edian 5 fit_ss_excl_sleep_min_hr50_m 0.59 8.89 0.00 0.00 359.00 359.00 24.6 794.0 0.14 in 0 6 fit_ss_excl_sleep_min_hr50_sd 5.76 18.90 000 0.00 199.19 199.19 4.71 25.83 0.30fit_ss_excl_sleep_min_max 131.33 126.8 88.00 0.00 886.00 886.00 2.00 4.37 2.053fit ss excl sleep min mean 29.47 34.19 19.93 0.00 460.75 460.75 4.17 29.40 0.55 fit ss excl sleep min median 17.15 32.57 900 0.00 473.00 473.00 5.50 46.67 0.53 fit_ss_excl_sleep_min_min 1.49 11.46 000 0.00 404.00 404.00 17.8 470.2 0.182 4 fit ss excl sleep min nohr m 56.05 105.6 25.00 0.00 885.00 885.00 3.63 14.19 1.70 ax 1fit ss excl sleep min nohr m 6.38 10.36 2 88 0.00 155.13 155.13 4.40 34.53 0.17 eanfit_ss_excl_sleep_min_nohr_m 0.88 3.45 000 0.00 65.00 65.00 7.43 85.19 0.06 edianfit ss excl sleep min nohr m 0.02 0.48 000 0.00 20.00 20.00 33.8 1265. 0.01 in 3 60 fit_ss_excl_sleep_min_nohr_sd 15.31 28.13 6.95 0.00 277.31 277.31 3.88 17.62 0.45 fit_ss_excl_sleep_min_rept_ma 86.50 85.82 62.00 0.00 759.00 759.00 2.73 10.13 1.38Xfit_ss_excl_sleep_min_rept_me 17.58 18.62 12.32 0.00 223.31 223.31 3.02 14.93 0.30 anfit_ss_excl_sleep_min_rept_me 8.01 15.62 000 0.00 214.00 214.00 4.26 29.63 0.25 dianfit_ss_excl_sleep_min_rept_mi 0.27 2.37 000 0.00 62.00 62.00 12.6 212.2 0.04 n 0 0 fit ss excl sleep min rept sd 25.70 25.34 18.91 0.00 266.59 266.59 3.23 15.95 0.41 fit ss excl sleep min sd 37.70 35.13 26.27 0.00 277.81 277.81 2.24 6.44 0.57 fit ss first inbed minutes ma 796.23 139.4 767.00 512.0 1438.0 926.00 1.38 2.75 2.25X 8 0 0 fit ss first inbed minutes me 660.13 75.37 648.60 417.9 1287.7 869.85 1.19 3.45 1.22 an 0 5 fit ss first inbed minutes me 651.80 77.76 639.00 394.5 1279.5 885.00 1.33 3.97 1.25 dian 0 0 fit ss first inbed minutes sd 71.94 52.52 58.08 0.71 910.75 910.05 3.52 28.34 0.85fit_ss_first_sleep_minutes_max 803.39 137.7 775.00 521.0 1439.0 918.00 1.36 2.68 2.222 0 0 fit ss first sleep minutes mea 666.87 75.20 655.73 422.5 1292.0 869.50 1.20 3.48 1.21 n 0 0fit ss first sleep minutes med 658,82 77,74 645,50 397,0 1284,0 887,00 1,33 3.96 1,25 ian 0 0 fit ss first sleep minutes min 563.95 104.0 574.00 0.00 1211.0 1211.0 8.36 1.680 0 0 1.96 fit_ss_first_sleep_minutes_sd 72.43 52.31 58.66 0.71 907.93 907.22 3.53 28.38 0.84 fit ss fitbit fairlyactivemin m 77.94 51.87 68.00 0.00 675.00 675.00 1.73 7.80 0.84 axfit ss fitbit fairlyactivemin m 23.85 19.50 19.18 0.00 175.50 175.50 1.70 4.49 0.31 eanfit_ss_fitbit_fairlyactivemin_m 17.72 19.85 12.00 0.00 175.50 175.50 1.94 5.85 0.32 edianfit ss fitbit fairlyactivemin m 1.04 5.54 000 0.00 120.00 120.00 9.85 136.5 0.09 in 7 fit ss fitbit fairlyactivemin sd 23.37 14.67 20.51 0.00 155.20 155.20 1.53 5.08 0.24 fit_ss_fitbit_lightlyactivemin_ 404.93 88.86 401.00 93.00 1193.0 1100.0 0.73 3.65 1.43 max 0 0 fit_ss_fitbit_lightlyactivemin_ 266.71 58.43 266.36 57.25 562.50 505.25 0.09 0.21 0.94 meanfit_ss_fitbit_lightlyactivemin_ 272.43 62.62 272.50 26.50 589.00 562.50 0.00 0.26 1.01 medianfit_ss_fitbit_lightlyactivemin_ 107.08 63.26 99.00 0.00 519.00 519.00 0.77 0.83 1.02 minfit ss fitbit lightlyactivemin s 84.07 26.64 81.69 2.83 279.17 276.35 1.05 3.34 0.43 dfit ss fitbit restingheartrate m 76.40 8.73 76.00 51.00 111.00 60.00 0.20 0.00 0.14 axfit ss fitbit restingheartrate m 72.17 7.91 72.00 48.92 100.08 51.16 0.16 -0.09 0.13 eanfit ss fitbit restingheartrate m 72.08 7.92 72.00 49.00 99.00 50.00 0.16 -0.08 0.13 edianfit_ss_fitbit_restingheartrate_m 68.42 7.75 68.00 47.00 98.00 51.00 0.22 -0.03 0.12 infit ss fitbit restingheartrate sd 2.36 1.12 2 12 0.00 11.39 11.39 1.70 5.38 0.02 fit ss fitbit sedentary min max 1248.7 145.9 1283.0 403.0 1440.0 1037.0 4.63 2.352 0 0 0 0 0 1.91 fit_ss_fitbit_sedentaryrmin_mea 703.25 128.8 687.43 204.5 1322.5 1117.9 0.78 1.36 2.08 n 1 7 0 3 fit ss fitbit sedentary niin mcd 659.04 146.7 631.50 0.00 1334.5 1334.5 1.48 3.44 2.37 ian 7 0 0 fit ss fitbit sedentarymin min 386.80 162.4 400.00 0.00 1212.0 1212.0 0.06 2.32 2.626 0 0 fit ss fitbit sedentaiymin sd 240.79 71.55 236.12 2.83 645.59 642.76 0.59 2.11 1.15 fit ss fitbit totalsteps max ###### 5597. ##### 1651. ##### ##### 0.81 1.22 90.2# 68 ## 00 ## ## 7 fit ss fitbit totalsteps mean 8952.3 3044. 8607.6 1103. ##### ##### 0.68 0.90 49.11 29 3 25 ## ## 0 fit ss fitbit totalsteps median 8865.5 3249. 8561.5 454.5 ##### ##### 0.63 1.02 52.44 45 0 0 ## ## 0 fit ss fitbit totalsteps min 2744.7 2141. 2208.0 4.00 ##### ##### 1.72 4.97 34.55 52 0 ## ## 4 fit ss fitbit totalsteps sd 3860.9 1473. 3640.9 112.4 ##### ##### 1.01 2.02 23.72 32 8 3 ## ## 6 fit ss fitbit veryactivemin ma 38.94 39.50 27.00 0.00 345.00 345.00 2.01 5.72 0.64Xfit ss fitbit veryactivemin me 9.97 12.70 5 36 0.00 146.33 146.33 2.63 10.37 0.20 anfit ss fitbit veryactivemin me 6.23 11.47 1.00 0.00 154.00 154.00 3.45 18.18 0.18 dianfit_ss_fitbit_veiyactivemin_mi 0.27 2.46 000 0.00 83.00 83.00 18.2 462.0 0.04 n 4 0 fit ss fitbit veryactivemin sd 11.62 11.68 8.10 0.00 99.89 99.89 2.05 6.18 0.19 fit ss light minutes max 337.20 51.91 333.00 110.0 701.00 591.00 0.90 3.75 0.840fit_ss_light_minutes_mean 259.89 32.06 260.32 92.00 398.00 306.00 0.62 0.520.05 fit ss light minutes median 259.36 33.75 259.50 92.00 425.50 333.50 0.00 0.87 0.54 fit ss light minutes min 184.55 44.64 187.50 19.00 378.00 359.00 -0.08 0.720.15 fit_ss_light_minutes_sd 45.98 18.50 42.96 0.00 225.36 225.36 1.83 8.24 0.30 fit ss night min max 611.33 89.53 615.00 0.00 1300.0 1300.0 8.75 1.440 0 1.03 fit ss night min mean 414.25 93.69 433.25 0.00 600.77 600.77 1.25 1.511.05 fit_ss_night_min_median 447.15 114.1 477.00 0.00 656.00 656.00 3.81 1.848 1.93fit ss night min min 67.28 88.10 41.00 0.00 522.00 522.00 2.23 6.10 1.42 fit ss night min sd 162.57 51.95 160.35 0.00 497.10 497.10 0.34 0.86 0.84 fit ss outbed minutes max 1276.8 85.47 1275.0 328.0 1439.0 1111.0 8.01 1.385 0 0 0 0 0.87 fit ss outbed minutes mean 1146.4 92.72 1154.9 129.0 1412.0 1283.0 25.72 1.507 5 0 0 0 3.77 fit ss outbed minutes median 1148.4 105.2 1146.0 51.00 1434.0 1383.0 51.47 1.704 5 0 0 0 5.65 fit ss outbed minutes min 978.30 294.6 1078.0 0.00 1395.0 1395.0 5.35 4.756 0 0 0 2.56 fit ss outbed minutes sd 97.24 118.9 62.64 0.00 869.74 869.74 3.15 10.05 1.921fit ss rem minutes max 140.67 26.17 140.50 38.50 293.00 254.50 0.17 1.28 0.42 fit ss rem minutes mean 97.20 17.85 97.70 31.75 154.90 123.15 0.26 0.290.18 fit ss rem minutes median 97.14 19.08 98.00 27.50 175.50 148.00 0.21 0.310.19 fit ss rem minutes min 54.17 22.71 54.00 1.00 146.00 145.00 0.17 -0.28 0.37 fit ss rem minutes sd 26.29 8.59 25.31 0.00 105.08 105.08 1.04 4.24 0.14fit_ss_sleep_min_max 628.18 89.37 618.00 166.0 1481.0 1315.0 1.52 7.29 1.440 0 0 fit ss sleep min mean 481.88 63.03 491.61 84.33 649.94 565.60 2.83 1.021.17fit ss sleep min median 491,51 62,97 500,00 21,00 736,00 715,00 6.29 1,021.61 fit ss sleep min min 274.69 154.2 316.00 0.00 590.00 590.00 -1.02 2.495 0.48 fit_ss_sleep_min_sd 96.78 51.16 85.03 2.83 469.16 466.33 1.31 2.77 0.83 fit ss sleeppenod ininutes ma 547.56 68.78 540.50 216.0 1071.5 855.50 1.34 6.47 1.11X 0 0 fit ss sleepperiod minutes me 446.07 40.55 450.20 192.5 589.42 396.92 1.96 0.65 an 0 0.78 fit_ss_sleepperiod_minutes_me 448.27 42.95 452.00 192.5 716.75 524.25 2.66 0.69 dian 0 0.79 fit ss sleepperiod minutes mi 331.19 77.32 342.50 93.50 525.00 431.50 -0.55 1.25 n 0.41 fit ss sleepperiod minutes sd 64.36 31.96 57.87 0.00 417.89 417.89 1.84 8.28 0.52 fit ss total ave met max 2.22 0.39 2.15 1.30 4.36 3.06 1.17 2.29 0.01 fit ss total ave met mean 1.71 0.20 1.69 1.23 3.72 2.49 1.18 4.68 0.00 fit ss total ave met median 1.69 0.20 1.67 1.21 3.90 2.69 1.18 5.46 0.00 fit ss total ave met min 1.33 0.18 1.31 1.00 3.26 2.26 1.44 7.04 0.00 fit ss total ave met sd 0.25 0.11 023 0.01 1.54 1.53 1.98 9.51 0.00 fit_ss_total_fairly_active_min_ 77.65 51.43 67.00 0.00 663.00 663.00 1.64 6.82 0.83 maxfit_ss_total_fairly_active_min_ 23.74 19.42 19.06 0.00 175.50 175.50 1.71 4.58 0.31 meanfit ss total fairly _acti\ c_inin_ 17.63 19.79 12.00 0.00 175.50 175.50 1.95 5.94 0.32 medianfit_ss_total_fairly_active_min_ 1.03 5.52 000 0.00 120.00 120.00 9.93 138.3 0.09 min 4fit_ss_total_fairly_active_min_ 23.27 14.49 20.47 0.00 132.44 132.44 1.40 3.68 0.23 sdfit_ss_total_light_active_min_ 404.35 88.91 400.00 92.00 1193.0 1101.0 0.73 3.65 1.43 max 0 0fit ss total light active min 265,42 58,27 264,92 56,75 562,50 505,75 0,09 0.22 0,94 meanfit_ss_total_light_active_min_ 271.39 62.74 272.00 25.00 589.00 564.00 0.00 0.27 1.01 medianfit_ss_total_light_active_min_ 104.54 62.86 97.00 0.00 519.00 519.00 0.80 0.91 1.01 minfit_ss_total_light_active_min_s 84.75 26.86 82.22 2.83 279.17 276.35 1.05 3.34 0.43 dfit ss total min max 1421.0 66.02 1440.0 632.0 1440.0 808.00 49.66 1.067 0 0 0 6.50 fit ss total min mean 1215.4 137.8 1252.1 506.7 1432.1 925.42 2.97 2.227 2 5 5 7 1.58 fit_ss_total_min_median 1303.6 162.6 1368.0 395.5 1440.0 1044.5 5.32 2.628 1 0 0 0 0 2.26 fit ss total min min 554.07 230.8 532.00 40.00 1425.0 1385.0 0.81 0.96 3.727 0 0 fit ss total min sd 261.41 92.49 253.93 0.71 773.57 772.87 0.58 1.82 1.49 fit_ss_total_sedentary_min_ma 1173.5 112.5 1186.0 391.0 1440.0 1049.0 5.33 1.82X 1 6 0 0 0 0 1.47 fit ss total sedentary min me 916.47 131.8 938.87 280.5 1246.8 966.33 1.48 2.13 an 5 0 3 1.01 fit ss total sedentary min me 966.62 151.4 1001.5 233.0 1344.5 1111.5 2.92 2.44 dian 2 0 0 0 0 1.51 fit ss total sedentary jnin mi 387.84 169.2 364.00 6.00 1108.0 1102.0 1.00 1.54 2.73 n 4 0 0 fit ss total sedentary min sd 224.58 71.49 217.20 14.14 669.63 655.49 0.60 1.49 1.15 fit ss total step max ###### 5580. ##### 1651. ##### ##### 0.81 1.26 90.0# 46 ## 00 ## ## 0 fit ss total step mean 8866.4 3017. 8536.6 1088. ##### ##### 0.67 0.87 48.65 91 0 25 ## ## 7fit_ss_total_step_median 8796.2 3225. 8493.0 406.5 ##### ##### 0.60 0.84 52.00 19 0 0 ## ## 1 fit ss total step min 2589.7 2104. 2045.0 0.00 ##### ##### 1.78 5.40 33.98 90 0 ## ## 5 fit ss total step sd 3887,2 1482, 3664,5 112,4 ##### ##### 1,02 2.07 23,98 13 4 3 ## ## 0 fit_ss_total_very_active_min_ 38.44 39.14 26.00 0.00 345.00 345.00 2.01 5.78 0.63 maxfit_ss_total_very_active_min_ 9.84 12.59 5 27 0.00 146.33 146.33 2.67 10.71 0.20 meanfit_ss_total_very_active_min_ 6.15 11.35 1.00 0.00 154.00 154.00 3.47 18.56 0.18 medianfit_ss_total_veiy_active_min_ 0.27 2.45 0,00 0.00 83.00 83.00 18.3 464.3 0.04 min 0 1 fit_ss_total_very_active_min_s 11.48 11.57 7 98 0.00 99.89 99.89 2.05 6.28 0.19 dfit_ss_wake_count_max 40.05 7.75 40.00 11.00 84.00 73.00 0.32 1.01 0.12 fit ss wake count mean 29.47 5.40 29.33 9.00 49.33 40.33 0.09 0.26 0.09 fit ss wake count median 29.39 5.59 29.00 9.00 50.00 41.00 0.07 0.21 0.09 fit ss wake count min 19.13 6.04 19.00 2.00 47.00 45.00 0.18 -0.01 0.10 fit ss wake count sd 6.33 2.25 604 0.00 27.79 27.79 1.47 6.90 0.04 fit ss wake minutes max 83.56 18.01 83.00 18.00 194.50 176.50 0.39 1.36 0.29 fit ss wake minutes mean 57.00 11.12 56.78 12.75 101.00 88.25 0.12 0.39 0.18 fit ss wake minutes median 56.32 11.76 56.00 12.75 105.00 92.25 0.15 0.40 0.19 fit ss wake minutes min 33.86 12.12 33.50 1.00 89.00 88.00 0.26 0.07 0.20 fit ss wake minutes sd 15.03 5.37 14.37 0.00 79.39 79.39 1.58 9.19 0.09 fit ss wakeup minutes max 1273.3 86.57 1270.0 296.0 1439.0 1143.0 8.09 1.406 0 0 0 0 0.83 fit ss wakeup minutes mean 1143.1 91.01 1151.0 122.6 1409.5 1286.8 26.70 1.475 6 7 0 3 3.75fit_ss_wakeup_niinutes_media 1145.7 101.6 1143.0 52.00 1434.0 1382.0 52.09 1.64 n 6 1 0 0 0 5.54 fit ss wakeup minutes min 975.12 285.8 1070.0 0.00 1390.0 1390.0 5.77 4.615 0 0 0 2.62fit ss wakeup minutes sd 96,44 116,2 62,96 0,00 869,74 869,74 3,23 10,74 1,871Interday CV 18.29 3.39 18.03 0.00 39.52 39.52 0.48 1.35 0.05 Interday SD 15.36 2.35 15.26 0.00 35.99 35.99 0.47 3.79 0.04 IntradaSD sd 4.91 1.97 4.51 0.00 14.47 14.47 0.78 0.47 0.03 IntradayCV mean 15.46 3.73 15.42 0.00 28.86 28.86 0.46 0.060.08 IntradayCV median 15.94 4.08 16.03 0.00 29.66 29.66 2.18 0.070.65 IntradayCV sd 4.91 1.97 451 0.00 1447 14.47 0.78 0.47 0.03 IntradayMean mean 84.63 7.14 84.79 57.86 122.31 64.45 0.05 0.36 0.12 IntradayMean median 84.32 7.33 84.44 57.75 132.00 74.25 0.11 0.65 0.12 IntradayMean sd 6.09 2.47 562 0.00 31.33 31.33 1.74 7.19 0.04 IntradaySD mean 12.99 2.73 13.18 0.00 22.58 22.58 0.98 0.040.46 IntradaySD median 13.29 3.04 13.55 0.00 23.02 23.02 3.87 0.051.23Mean 84.71 7.18 84.84 57.92 122.31 64.39 0.06 0.37 0.12 Median 82.39 8.40 82.14 54.49 132.00 77.51 0.20 0.45 0.14 STD 15.38 2.35 15.28 0.00 36.01 36.01 0.49 3.76 0.04

[0226] This table shows the static features derived from the Fitbit data. The individuallevel features listed in the first column are calculated directly from the data of individual participants based on summarization of individual-level measurements. In contrast, the population-level summary statistics are compiled from data across all cohorts’ participants. Itcontains the mean, standard deviation (sd), median, minimum (min), maximum (max), range, skewness, kurtosis, and standard error (se) of each feature.Data S48. 258 Wearable-derived Static Features Grouped into 7 Clusters, related to Figure 2 and STAR Methods “Machine Learning Classifier” SectionCluster 1 Cluster 2 Cluster 3 fit_ss_outbed_minutes_min fit_ss_excl_sleep_niin_max fiLss_t°'aLave_niet_maxfit_ss_wakeup_minutes_mm fit_ss_excl_sleep_min_sd fit_ss_total_ave_met_sd fit_ss_wakejninutes_min lit_ss_e\cl_niglitjrinjna\ fit_ss_dayt_ave_met_value_sd fit_ss_wake_couut_min fit_ss_excl_night_min_sd fit_ss_fitbit_veryactivemin_max fit_ss_sleepperiod_minutes_min fit_ss_excl_night_min_hr_rept_maxfit_ss_t°tal_very_active_min_maxfit_ss_light_minutes_min fit_ss_excl_sleep_min_rept_max fit_ss_dayt_very_active_mm_max fit_ss_outbed_minutes_median fit_ss_excl_mght_min_hr_rept_sd fit-SS-fitbi^veryactivem^sd fit_ss_wakeup_mmutes_median fit_ss_excl_sleep_min_rept_sd fit_ss_total_very_active_mni_sd fit_ss_outbed_minutes_mean lit_ss cxcl_night_min_hr_re]>t_meLin fit_ss_dayt_very_active_mm_sd fit_ss_wakeup_minutes_mean fit_ss_excl_sleep_min_rept_mean fit_ss_fitbit_veryactivemin_mean fit_ss_deep_minutes_min fit_ss_excl_mght_min_hr_rept_median fit_ss_t°tal_very_active_mm_inean fit_ss_rem_minutes_niin fit_ss_excl_sleep_min_rept_median fit_ss_dayt_very_active_mm_mean fit_ss_deep_minutes_mean lit_ss_e\cl_da\_min_hr50_niedian fit_ss_fitbit_veryactivemin_median fit_ss_deep_minutes_median l'it_ss_e\cl_da\_min_ln50_niean fit-BS-total-Wry-aetwe-mn-Uiediati fit_ss_remjninutes_niean Iit_ss_e\cl_da\_min_hr5!)_nia\ fit_ss_dayt_very_active_mm_median fit_ss_rem_minutes_median fit_ss_excl_day_min_hr50_sd fit_ss_fitbit_totalsteps_sd fit_ss_dayt_light_active_min_max fit_ss_excl_night_min_min fit_ss_day_total_steps_no_el_sd fit_ss_total_light_active_min_max lit_ss_e\cl_niglU_nan_ln50_aim fit-SS-total-Step-Sd fit_ss_fitbit_lightlyactivemin_max fit_ss_excl_sleep_min_miii fit_ss_dayt_total_steps_sd fit_ss_fitbit_lightlyactivemin_median fit_ss_excl_sleep_min_hr50_min fit_ss_fitbit_fairlyactivemm_median fit_ss_total_light_active_min_median lit_ss_excl_niglitjrinjnean fit_ss_total_fairly_active_min_median fit_ssjiaytJight_activejnin_niedian fit_ss_excl_sleep_mm_mean fit-BS-dayt-fairly-active^i^median nt_ss_cayt_liglH_active_niin_mean fit_ss_e\cl_niglitjrin_median fit-SS-dayt-fairly-actiye-inm-in631! fit_ss_total_light_active_min_mean fit_ss_excl_sleep_min_median fit_ss_total_fairly_active_min_mean fit_ss_fitbit_lightlyactivemm_mean fit_ss_excl_night_min_hr50_sd fit_ss_fitbit_fairlyactivemm_mean fit_ss_dayt_fairly_active_mm_min fit_ss_excl_sleep_min_hr50_sd fit_ss_dayt_ave_met_value_maxfit_ss_total_fairly_active_mm_min lit_ss_e\cljiiglitjrinJir5()_nax fit_ss_dayt_fairly_active_min_sd fit_ss_fitbit_fairlyactivemin_min fit_ss_excl_sleep_min_hr50_max fit_ss_total_fairly_active_min_sd fit_ss_dayt_very_active_min_min fit_ss_excl_iiight_niin_hr50_mean fit_ss_fitbit_fairlyactivemm_sd fit_ss_total_very_active_mm_min fit-SS-Cxcl-Sleep-miU-hrSOjnean fit_ss_dayt_fauly_active_min_maxfit_ss_fitbit_veryactivemin_niin fit-SS-excl-mght-mm-hrSOjnedian fit_ss_t°tal_fairly_active_mm_maxfit_ss_total_ave_met_miii fit_ss_excl_sleepjnin_lir50_median fit_ss_fitbit_fairlyactivemin_max fit_ss_cavt_ave_nct_valiic_min IntradayCV_mean fit_ss_fitbit_lightlyactivemin_min IntradayCV_median fit_ss_total_Ught_active_min_min Intraday SDjnean fit_ss_dayt_light_active_min_niin Intraday SD_median fit_ss_fitbit_totalsteps_min InterdayCV fit_ss_day_total_steps_no_el_min STDfit_ss_total_step_min Interday SD fit ss dayt total steps min fit_ss_dayt_ave_met_value_mean fit_ss_dayt_ave_met_value_median fit_ss_t°tal_ave_met_mean fit_ss_total_ave_met_median fit ss total step max fit_ss_dayt_total_steps_max fit_ss_day_total_steps_n°_el_max fit_ss_fitbit_totalsteps_max fit_ss_t°tal_step_median fit-SS-dayt-total-Stepsjnedian fit_ss_day_total_steps_n°_el_median fit_ss_fitbit_totalsteps_median fit_ss_fitbit_totalsteps_meaii fit_ss_day_total_steps_no_el_mean fit-SS-total-Step-mean fit_ss_dayt_total_steps_meanCluster 4fit_ss_fitbit_sedentarymin_minfit_ss_fitbit_sedentarymin_meanfit_ss_fitbit_sedentarymin_medianlit_ss_e\cl_da\_min_nohr_ineaiifit_ss_excl_day_min_meanfit_ss_excl_day_min_medianlit_ss_e\cl_da\_min_hr_rept_ma\lit_ss_e\cl_da\_min_hr_rcpt_sdfit_ss_excl_day_min_hr_rept_meanfit_ss_excl_day_min_hr_rept_medianCluster 5 Cluster 6 Cluster 7fit_ss_sleep_min_sd fit_ss_avg_hr_light_max fit_ss_dayt_sedentary_min_maxfit_ss_first_inbed_minutes_niean fit_ss_avg_hr_rem_maxfit_ss_dayt_sedentaiy_min_mean lit_ss_lirsl_sleep_niiniites_mean fit-SS-a^-hr-Wakejnax fit_ss_dayt_sedentaiy_min_median fit_ss_first_inbed_minutes_median fit_ss_avg_hr_deep_max fit_ss_night_min_max fit_ss_first_sleep_minutes_median Median fit_ss_t°tal_min_max fit_ss_sleepperiod_minutes_sd IntradayMean_median fit_ss_t°tal_sedentary_min_maxfit_ss_light_minutes_sd Mean fit_ss_t°tal_sedentaiy_min_mean fit_ss_outbed_minutes_sd IntradayMean_mean fit_ss_t°tal_sedentary_min_median fit_ss_wakeup_minutes_sd fit ss fitbit restingheartrate min fit ss light minutes max fit_ss_first_inbed_minutes_sd fit_ss_avg_hr_deep_mm fit_ss_wake_minutes_max fit_ss_first_sleep_minutes_sd fit_ss_avg_hr_rem_min fit_ss_wake_count_max fit_ss_fiist_inbed_minutes_max fit_ss_avg_hr_wake_min fit_ss_sleepperiod_minutes_max fit ss first sleep minutes max fit ss avg hr light min fit ss sleep min max nt_ss_litbit_rcstmgheartrate_sd fit-SS-fitbit-restmgheartratejnax fit_ss_deep_minutes_max fit_ss_avg_hr_deep_sd fit_ss_avg_hr_wake_niean fit-SS-remjninutesjnax fit_ss_avg_hr_rem_sd fit_ss_avg_hr_wake_median fit_ss_sleepperiod_minutes_mean fit_ss_avg_hr_wake_sd fit-SS-avgJiT-deepjnean fit_ss_sleepperi°d_mmutes_median fit_ss_avg_hr_light_sd fit_ss_avg_hr_deep_median fit_ss_light_mmutes_meanfit_ss_day_min_max fit-SS-ayg-hiMight-inean fit_ss_light_minutes_median fit_ss_outbed_minutes_max fit-SS-avgJiT-light-inedian fit_ss_wake_count_mean fit_ss_wakeup_minutes_max fit_ss_fitbit_restingheartrate_mean fit_ss_wake_count_median fit_ss_wake_minutes_sd fit_ss_fitbit_restingheartrate_median fit_ss_wake_minutes_mean fit_ss_wake_coimt_sd fit_ss_avg_hr_rein_mean fit_ss_wake_minutes_median fit_ss_deep_mmutes_sd fit_ss_avg_hr_rein_median fit_ss_sleep_min_min fit_ss_rem_minutes_sd fit_ss_sleep_niin_inean fit_ss_excl_sleep_min_nohr_mean fit_ss_sleep_min_inedian fit_ss_excl_sleep_min_nohr_maxfit_ss_mght_min_inean fit_ss_excl_sleep_min_nohr_sd fit_ss_night_min_inedian fit_ss_excl_night_min_nohr_mean fit_ss_total_mm_mean fit_ss_excl_night_min_nohr_max fit_ss_total_min_median fit_ss_excl_night_min_nohr_sd fit_ss_day_min_mean fit_ss_first_sleep_mmutes_mm fit_ss_day_rnin_median fit_ss_excl_day_min_hr50_niin fit_ss_day_rnin_min fit_ss_excl_sleep_min_nohr_min fit_ss_dayt_sedentaiy_miii_min fit_ss_excl_night_min_hr_rept_min fit_ss_night_min_min fit_ss_excl_sleep_min_rept_min fit_ss_total_min_min fit_ss_excl_night_min_nohr_min fit_ss_total_sedentary_min_min fit_ss_excljughtjninjiohr_medianfit_ss_excl_sleep_min_nohr_median IntradayMean_sd IntradayCV_sd IntradaySD_sd fit_ss_excl_day_min_min fit_ss_excl_day_mm_nohr_median fit_ss_excl_day_min_nohr_min fit_ss_excl_day_mm_hi_rept_inin fit_ss_dayt_light_active_min_sd fit_ss_total_hght_active_min_sd fit_ss_fitbit_lightlyactivemin_sd fit_ss_excl_day_mm_nohr_sd fit ss total sedentary min sd fit_ss_excl_day_mm_sd fit_ss_total_min_sd fit_ss_night_min_sd fit_ss_fitbit_sedentarymin_sd fit_ss_excl_day_mm_nohr_max fit_ss_excl_day_mm_max fit_ss_fitbit_sedentarymin_max fit_ss_day_min_sd fit_ss_dayt_sedentary_min_sd

[0227] This table shows the 258 unsupervised static features grouped into seven clusters, where each cluster signifies a set of features emphasizing different physiological and behavioral aspects of the data obtained from the Fitbit devices. The clustering pattern (Data S49) provides insight into potential associations among the features within each cluster.Data S49 Description of the Seven Clusters of Static Features, related to Figure 2 and STAR Methods “Machine Learning Classifier” SectionCluster Index Cluster DescriptionCluster 1 focuses on sleep and light physical activity' metrics. It includes various metrics of sleep stages (such as deep and REM sleep), wakefulness, and out-of-bed times, indicating variations in sleep quality and duration. Additionally, it includes metrics related to light and fairly active minutes. The cluster reflects an integration of sleep quality with low-intensity Cluster 1 physical activities, focusing on rest periods an dlight daily movements.Cluster 2 focuses on metrics of sleep exclusion and heart rate variability during sleep and Cluster 2waking hours. This cluster focuses on the quality of sleep and resting periods by excludingcertain time frames from analysis. It also evaluates heart rate paterns that may reflect overall heart health during sleep and awake states.Cluster 3 focuses on metrics of physical activity intensity and metabolic rates. It contains metrics for very active minutes, steps taken, and average metabolic equivalent tasks (METs), both daily and during active periods. It observes the physical activity patterns to evaluate Cluster 3 higher-intensity physical activities and their impact on metabolic health.Cluster 4 focuses on sedentary behavior and average daily heart rate metrics. This cluster Cluster 4 emphasizes the importance of monitoring sedentaiy time and its potential health implications.Cluster 5 focuses on the variability and extremities of sleep, sedentary' behavior, and heart rate metrics. This cluster highlights the range and variability in sleep quality, initiation times, and Cluster 5 restfulness. It also captures the scale of sedentary behavior and resting heart rate fluctuations.Cluster 6 focuses on various metrics reflecting heart rates during different sleep stages and while awake. This cluster highlights variations in heart rate that could reflect stress levels, sleep Cluster 6 quality.Cluster 7 focuses on sleep quality metrics, sedentary minutes, and total daily minutes, both active and inactive. This cluster aims to assess both rest and activity patterns, offering insights Cluster 7 into sleep depth, duration, and overall daily activity levels, including sedentary' behaviors.

[0228] This table provides descriptions of the seven clusters identified from the 258 unsupervised features. It outlines the qualitative nature of each cluster, offering context and interpretation to the grouping established in the previous analysis.Data S50. ANOVA on wearable features and covariates across different psychiatric disorder groups, related to Figure 2 and STAR Methods “Machine Learning Classifier” Section R- F- p- Disorder Fitbit Feature Covariate square statisti sum sq df valu d c e anxiety disord fit ss sleepperiod minutes medi family inco 0.05 7.37 12221.7 1.0 0.01 er an me 4 0 anxiety_disord fit_ss_sleepperiod_minutes_sd parent_grade 0.06 15.07 13443.8 1.0 0.00 er 1 0 severe_disorde fit ss first inbed minutes mean family inco 0.06 8.13 43288.0 1.0 0.00 r me 0 0 severe disorde fit ss first sleep minutes mean family inco 0.06 7.54 40187.2 1.0 0.01 r me 6 0 severe disorde fit ss avg hr wake mean sex 0.06 5.22 280.82 1.0 0.02 r 0 severe disorde fit ss avg hr light mean sex 0.05 5.01 284.52 1.0 0.03 r 0severe disorde fit ss avg hr rem mean sex 0.05 5.02 266.44 1.0 0.03 r 0 severe_disorde fit ss first inbed minutes max family inco 0.05 4.43 75255.7 1.0 0.04 r me 4 0 severe disorde fit ss first sleep minutes max family inco 0.05 4.24 70147.1 1.0 0.04 r me 1 0 severe_disorde fit_ss_first_inbed_minutes_medi family _inco 0.06 8.73 50177.4 1.0 0.00 r an me 3 0 severe disorde fit ss first sleep minutes media parent grade 0.06 3.85 22303.8 1.0 0.05 r n 5 0 severe disorde fit ss first sleep minutes media family inco 0.06 8.67 49962.9 1.0 0.00 r n me 3 0 severe disorde fit ss avg hr wake median sex 0.06 4.73 253.21 1.0 0.03 r 0 severe disorde fit ss avg hr light median sex 0.05 5.07 286.83 1.0 0.02 r 0 severe disorde fit ss avg hr rem median sex 0.05 5.15 272.79 1.0 0.02 r 0 severe disorde fit ss first inbed minutes sd parent grade 0.05 6.24 14373.5 1.0 0.01 r 6 0 severe_disorde fit_ss_first_sleep_minutes_sd parent_grade 0.05 5.71 12934.8 1.0 0.02 r 1 0 severe_disorde frt ss sleepperiod minutes sd parent grade 0.06 8.64 7718.76 1.0 0.00 r 0 severe disorde frt ss sleepperiod minutes sd family inco 0.07 10.75 9438.75 1.0 0.00 r me 0 severe disorde fit ss total ave met sd sex 0.05 4.10 0.05 1.0 0.04 r 0 bipolar group fit ss first inbed minutes sd parent div c 0.05 5.23 12048.7 1.0 0.02 at 4 0 bipolar group fit ss first inbed minutes sd parent grade 0.06 15.69 35933.2 1.0 0.007 0 bipolar group fit ss first sleep minutes sd parent div c 0.05 5.72 12952.4 1.0 0.02 at 7 0 bipolar_group fit_ss_first_sleep_minutes_sd parent_grade 0.06 15.30 34455.9 1.0 0.007 0 bipolar group frt ss sleepperiod minutes sd parent grade 0.06 6.07 5419.95 1.0 0.010eating group fit ss first inbed minutes mean parent grade 0.06 4.90 26166.4 1.0 0.039 0 eating group fit ss first sleep minutes mean parent grade 0.06 5.42 28992.2 1.0 0.020 0 eating group fit ss first sleep minutes max family inco 0.05 4 13 68248.3 1.0 0.04 me 6 0 eating_group fit_ss_first_inbed_minutes_medi parent_grade 0.06 10.13 58335.3 1.0 0.00 an 4 0 eating group fit ss first sleep minutes media parent grade 0.06 10.39 59949.0 1.0 0.00 n 0 0 eating group fit ss first inbed minutes sd parent div c 0.06 15.08 34550.9 1.0 0.00 at 5 0 eating group fit ss first inbed minutes sd parent grade 0.06 10.58 24286.0 1.0 0.009 0 eating group fit ss first inbed minutes sd family inco 0.07 4.65 10492.1 1.0 0.03 me 8 0 eating group fit ss first sleep minutes sd parent div c 0.06 14.96 33710.2 1.0 0.00 at 8 0 eating group fit ss first sleep minutes sd parent grade 0.05 11.22 25296.2 1.0 0.004 0 eating_group fit_ss_first_sleep_minutes_sd family _inco 0.07 4.84 10742.8 1.0 0.03 me 8 0 eating group fit_ss_total_ave_met_mean sex 0.13 7.18 0.25 1.0 0.010 eating group fit_ss_total_very_active_min_me sex 0.13 7.71 1088.00 1.0 0.01 an 0 eating group fit ss fitbit veiyactivemin mean sex 0.13 7.55 1085.48 1.0 0.010 eating group fit_ss_dayt_ave_met_value_mea sex 0.13 7.61 0.46 1.0 0.01 n 0 eating group fit_ss_dayt_very_active_min_me sex 0.13 7.73 1086.88 1.0 0.01 an 0 eating group fit ss total ave met max sex 0.09 6.14 0.88 1.0 0.010 eating_group fit_ss_total_very_active_min_ma sex 0.10 10.29 14264.7 1.0 0.00X 3 0 eating group fit_ss_fitbit_veryactivemin_max sex 0.10 10.12 14307.9 1.0 0.004 0eating group fit_ss_dayt_ave_met_value_max sex 0.10 9.26 1.73 1.0 0.000 eating group fit_ss_dayt_very_active_min_ma sex 0.10 10.44 14254.1 1.0 0.00X 0 0 eating group fit ss total ave met median sex 0.12 4.96 0.18 1.0 0.030 eating_group fit_ss_total_very_active_min_me sex 0.11 3.89 460.64 1.0 0.05 dian 0 eating group fit ss fitbit veryactivemin medi sex 0.11 4.10 495.17 1.0 0.04 an 0 eating group fit_ss_dayt_ave_met_value_medi sex 0.12 6.29 0.42 1.0 0.01 an 0 eating group fit_ss_total_very_active_min_sd sex 0.11 9.32 1083.27 1.0 0.000 eating group fit ss fitbit veryactivemin sd sex 0.11 8.98 1068.14 1.0 0.000 eating group fit ss dayt very active min sd sex 0.11 9.38 1081.13 1.0 0.000 ocd group fit ss first inbed minutes mean parent grade 0.05 4.73 25396.7 1.0 0.033 0 ocd_group fit_ss_first_sleep_minutes_mean parent_grade 0.05 4.73 25400.2 1.0 0.034 0 ocd_group fit ss first sleep minutes max family inco 0.05 4.21 69802.2 1.0 0.04 me 7 0 ocd group fit ss first inbed minutes medi parent grade 0.06 6.42 37155.8 1.0 0.01 an 8 0 ocd group fit ss first sleep minutes media parent grade 0.06 6.03 34966.8 1.0 0.01 n 3 0 ocd group fit ss sleepperiod minutes sd parent grade 0.05 7.65 6840.71 1.0 0.010 pstd group fit_ss_sleep_min_sd family inco 0.05 5.11 12656.7 1.0 0.02 me 5 0

[0229] This table shows the result of an ANOVA test, focusing on wearable features and covariates across different psychiatric disorder groups. The table shows results that are statistically significant, suggesting the importance of using covariates in our downstream modeling analyses. Only comparisons reporting p value < 0.05 and R-square > 0.05 are shown.Data S51. Description of Features Used in Models, related to Figure 2 and STAR Methods “Machine Learning Classifier” SectionFeature Evaluation DescriptionAbbreviation TypeD demographic information Objective FH family history Objective CS cognitive score Objective COV covariates (D, FH, CS) Objective CBCL childhood behavior check list score Subjective static Static Wearable Features (All 7 Clusters) Objective W static Static Wearable Features - Cluster 1 Objective.,. C2W state Static Wearable Features - Cluster 2 Objective >A / C3W state Static Wearable Features - Cluster 3 Objective... C4W state Static Wearable Features - Cluster 4 Objective C5W static Static Wearable Features - Cluster 5 Objective WC6state Static Wearable Features - Cluster 6 Objective W state Static Wearable Features - Cluster 7 Objective WtsTime series Wearable Features Objective

[0230] This table is a compilation of features used in the predictive models. They are described and abbreviated for ease of understanding. Each feature is categorized as objective (objectively measured) or subjective (assessed by medical professionals) based on its source. The common variables across all models include demographic information (abbreviated as D), family history (FH), and cognitive scores (CS). For simplification, in our models we have bundled these three under the abbreviation COV.Data S52. Description of Models, related to Figure 2 and STAR Methods “Machine Learning Classifier” SectionIndex Architecture Type Covariates Evaluation Type 1 XGBoost GOV Objective2 XGBoost COV + WC1,K, Objective3 XGBoost cov + wC2aa8eObjective4 XGBoost cov + wC3aa8eObjective5 XGBoost cov + w04,^ Objective6 XGBoost cov + w“.s„ Objective7 XGBoost COV + WC6staticObjective8 XGBoost COV + WC7staticObjective9 XGBoost COV + WA"aajcObjective10 XGBoost COV + CBCL Subjective and Objective 11 XGBoost COV + CBCL + WG1staticSubjective and Objective 12 XGBoost COV + CBCL + w“„,icSubjective and Objective 13 XGBoost COV + CBCL + WC3sta!icSubjective and Objective 14 XGBoost COV + CBCL + W04^ Subjective and Objective 15 XGBoost COV + CBCL + WC5stafeSubjective and Objective 16 XGBoost COV + CBCL + WC6s1aficSubjective and Objective 17 XGBoost COV + CBCL + WC7staSicSubjective and Objective 18 XGBoost COV + CBCL + WA!UScSubjective and Objective 19 Xception COV + WteObjective20 Xception COV + CBCL + Wt5Subjective and Objective 21 Xception COV + CBCL Subjective and Objective

[0231] This table outlines the various model architectures used. Each model design is given a model index, referred to in FIGS. 19, 20, and 22.Data S53. Significant features identified by the post-hoc test for the continuous multivariate GW AS hits for ADHD, related to FIG. 4, Table 1, FIG. 14, Table S2, and STAR Methods “GWASs for ADHD” Sectionfeature genotype genotype p value FDR value disease group 1 group 2 groupSTD GG GA 0.20851208 0.33729896 Ctrl 9 8STD GG AA 0.22553500 0.35441214 Ctrl 2 6STD GA AA 0.15172730 0.29227863 Ctrl 6 4 InterdayCV GG GA 0.10377115 0.25086135 Ctrl 4InterdayCV GG AA 0.35860030 0.50571837 Ctrl 3 6 InterdayCV GA AA 0.21908325 0.35095860 Ctrl 2 7 InterdaySD GG GA 0.20623094 0.33691193 Ctrl 1 3 InterdaySD GG AA 0.22510748 0.35441214 Ctrl 7 6 InterdaySD GA AA 0.14940497 0.29227863 Ctrl 3 4 IntradayCV mean GG GA 0.59895352 0.70372670 Ctrl 6 3 IntradayCV mean GG AA 0.19237790 0.32062984 Ctrl 7 4 IntradayCV mean GA AA 0.15645430 0.29227863 Ctrl 3 4 IntradayCV median GG GA 0.15247054 0.29227863 Ctrl 5 4 IntradayCV median GG AA 0.23464704 0.36525247 Ctrl 4 4 IntradayCV median GA AA 0.15526221 0.29227863 Ctrl 1 4 IntradaySD mean GG GA 0.96586767 0.97771881 Ctrl 6 3 IntradaySD_mean GG AA 0.19085741 0.32062984 Ctrl 2 4 IntradaySD mean GA AA 0.16251909 0.29467748 Ctrl 7 4IntradaySD median GG GA 0.65275038 0.73280452 Ctrl 3IntradaySD median GG AA 0.14489750 0.29227863 Ctrl 6 4 IntradaySD median GA AA 0.12162955 0.26758501 Ctrl 1 3 fit_ss_day_total_steps_no_el_mean GG GA 0.05342806 0.19640608 Ctrl 6 6 fit_ss_day_total_steps_no_el_mean GG AA 0.97459241 0.98053505 Ctrl 5 1 fit_ss_day_total_steps_no_el_mean GA AA 0.52480515 0.64621530 Ctrl 5 3 fit_ss_total_step_mean GG GA 0.05475563 0.19640608 Ctrl 6 6 fit_ss_total_step_mean GG AA 0.94076014 0.95818162 Ctrl 3 7 fit_ss_total_step_mean GA AA 0.51156379 0.63945474 Ctrl 3 1 fit_ss_total_ave_met_mean GG GA 0.11186370 0.25996495 Ctrl 6 1 fit_ss_total_ave_met_mean GG AA 0.18972286 0.32062984 Ctrl 4 fit_ss_total_ave_met_mean GA AA 0.0678393 0.21791414 Ctrl 9 fit_ss_total_fairly_active_min_mea GG GA 0.00537374 0.07880345 Ctrl n 9 fit_ss_total_fairly_active_min_mea GG AA 0.43636473 0.56693056 Ctrl n 6 2 fit_ss_total_fairly_active_min_mea GA AA 0.11488802 0.26039557 Ctrl n 8 6 fit ss fitbit totalsteps mean GG GA 0.05462802 0.19640608 Ctrl 7 6 fit_ss_fitbit_totalsteps_mean GG AA 0.98974636 0.98974636 Ctrl 5 5 fit ss fitbit totalsteps mean GA AA 0.55454140 0.67581376 Ctrl 5 9fit ss fitbit fairlyactivemin mean GG GA 0.00625023 0.07880345 Ctrl fit_ss_fitbit_fairlyactivemin_mean GG AA 0.38065637 0.52780085 Ctrl 4 5 fit ss fitbit fairlyactivemin mean GA AA 0.09889925 0.24934566 Ctrl 9 4fit ss dayt total steps mean GG GA 0.05660123 0.19870646 Ctrl 7 9 fit ss dayt total steps mean GG AA 0.93453771 0.95775604 Ctrl 2 fit ss dayt total steps mean GA AA 0.52214291 0.64621530 Ctrl 1 3 fit ss dayt ave met value mean GG GA 0.19985142 0.32975484 Ctrl 3 8 fit ss dayt ave met value mean GG AA 0.15147935 0.29227863 Ctrl 8 4 fit ss dayt ave met value mean GA AA 0.07829989 0.23070505 Ctrl 6 1 fit_ss_dayt_farily_active_min_mea GG GA 0.00527804 0.07880345 Ctrl n 9 fit_ss_dayt_farily_active_min_mea GG AA 0.42460257 0.56304595 Ctrl n 6 9 fit_ss_dayt_farily_active_min_mea GA AA 0.11678347 0.26039557 Ctrl n 6 fit_ss_day_total_steps_no_el_max GG GA 0.00711716 0.07880345 Ctrl 2 fit_ss_day_total_steps_no_el_max GG AA 0.58436379 0.69366925 Ctrl 9 7 fit_ss_day_total_steps_no_el_max GA AA 0.85879831 0.89120580 Ctrl 7 1 fit_ss_total_step_max GG GA 0.00420006 0.07880345 Ctrl 4 fit ss total step max GG AA 0.60136645 0.70372670 Ctrl 5 3 fit ss total step max GA AA 0.80457561 0.85099344 Ctrl 9 3fit_ss_total_fairly_active_min_max GG GA 0.00591263 0.07880345 Ctrl 6 fit_ss_total_fairly_active_min_max GG AA 0.3185807 0.45709404 Ctrl 8 fit_ss_total_fairly_active_min_max GA AA 0.07624427 0.23070505 Ctrl 7 1 fit ss fitbit totalsteps max GG GA 0.00596773 0.07880345 Ctrl 5 fit ss fitbit totalsteps max GG AA 0.55703437 0.67581376 Ctrl 9 9 fit_ss_fitbit_totalsteps_max GA AA 0.88133598 0.90887773 Ctrl 9 9 fit_ss_fitbit_fairlyactivemin_max GG GA 0.00615475 0.07880345 Ctrl 5 fit_ss_fitbit_fairlyactivemin_max GG AA 0.18821786 0.32062984 Ctrl 5 4 fit_ss_fitbit_fairlyactivemin_max GA AA 0.03681437 0.19640608 Ctrl 3 6 fit_ss_dayt_total_steps_max GG GA 0.00425068 0.07880345 Ctrl 1 fit_ss_dayt_total_steps_max GG AA 0.60604200 0.70420373 Ctrl 5 8 fit_ss_dayt_total_steps_max GA AA 0.80457561 0.85099344 Ctrl 9 3 fit_ss_dayt_farily_active_min_max GG GA 0.00671706 0.07880345 Ctrl 2 fit_ss_dayt_farily_active_min_max GG AA 0.32568072 0.46325275 Ctrl 4 4 fit ss dayt farily active min max GA AA 0.07692453 0.23070505 Ctrl 6 1 fit_ss_day_total_steps_no_el_medi GG GA 0.09615089 0.24934566 Ctrl an 6 4 fit_ss_day_total_steps_no_el_medi GG AA 0.79845177 0.85099344 Ctrl an 2 3 fit_ss_day_total_steps_no_el_medi GA AA 0.46537787 0.59525077 Ctrl an 6 2fit ss total step median GG GA 0.10158428 0.25017026 Ctrl 8 2 fit ss total step median GG AA 0.78639374 0.84807169 Ctrl 9 fit ss total step median GA AA 0.44806292 0.57758111 Ctrl 5 5 fit_ss_total_fairly_active_min_med GG GA 0.04167229 0.19640608 Ctrl ian 9 6 fit_ss_total_fairly_active_min_med GG AA 0.24811178 0.37905967 Ctrl ian 9 8 fit_ss_total_fairly_active_min_med GA AA 0.09478839 0.24934566 Ctrl ian 4 4 fit_ss_fitbit_totalsteps_median GG GA 0.09148635 0.24934566 Ctrl 3 4 fit_ss_fitbit_totalsteps_median GG AA 0.82793097 0.86461145 Ctrl 3 9 fit_ss_fitbit_totalsteps_median GA AA 0.47543674 0.60343894 Ctrl fit ss fitbit fairlyactivemin media GG GA 0.04645878 0.19640608 Ctrl n 1 6 fit ss fitbit fairlyactivemin media GG AA 0.25132985 0.38045345 Ctrl n 7 3 fit ss fitbit fairlyactivemin media GA AA 0.09973826 0.24934566 Ctrl n 6 4 fit_ss_dayt_total_steps_median GG GA 0.10490565 0.25086135 Ctrl 6 fit ss dayt total steps median GG AA 0.77866892 0.84526561 Ctrl 9 4 fit ss dayt total steps median GA AA 0.43351738 0.56693056 Ctrl 5 2 fit_ss_dayt_farily_active_min_med GG GA 0.04441092 0.19640608 Ctrl ian 9 6 fit_ss_dayt_farily_active_min_med GG AA 0.24583089 0.37905967 Ctrl ian 3 8fit ss dayt farily active min med GA AA 0.09559942 0.24934566 Ctrl ian 5 4fit ss day total steps _no_el_sd GG GA 0.00676336 0.07880345 Ctrl 3 fit_ss_day_total_steps_no_el_sd GG AA 0.66174468 0.73280452 Ctrl 8 fit_ss_day_total_steps_no_el_sd GA AA 0.67507052 0.74257758 Ctrl 9 2 fit_ss_total_step_sd GG GA 0.00475046 0.07880345 Ctrl 9fit_ss_total_step_sd GG AA 0.65205025 0.73280452 Ctrl 5fit_ss_total_step_sd GA AA 0.63666558 0.72951265 Ctrl 9 5 fit_ss_total_fairly_active_min_sd GG GA 0.00760133 0.07880345 Ctrl 5 fit_ss_total_fairly_active_min_sd GG AA 0.57747322 0.69045711 Ctrl 4 6 fit_ss_total_fairly_active_min_sd GA AA 0.15765332 0.29227863 Ctrl 4 4 fit_ss_fitbit_totalsteps_sd GG GA 0.00706869 0.07880345 Ctrl fit_ss_fitbit_totalsteps_sd GG AA 0.68951726 0.75344602 Ctrl 7 1 fit_ss_fitbit_totalsteps_sd GA AA 0.66019208 0.73280452 Ctrl 4 fit_ss_fitbit_fairlyactivemin_sd GG GA 0.00884116 0.08581128 Ctrl 3 3 fit ss fitbit fairlyactivemin sd GG AA 0.42654996 0.56304595 Ctrl 9 9 fit ss fitbit fairlyactivemin sd GA AA 0.09318287 0.24934566 Ctrl 5 4 fit_ss_dayt_total_steps_sd GG GA 0.00477680 0.07880345 Ctrl 1 fit ss dayt total steps sd GG AA 0.65850680 0.73280452 Ctrl 2 fit ss dayt total steps sd GA AA 0.62793512 0.72454053 Ctrl 8 3fit_ss_dayt_farily_active_min_sd GG GA 0.00764154 0.07880345 Ctrl 7 fit_ss_dayt_farily_active_min_sd GG AA 0.57594719 0.69045711 Ctrl 7 6 fit_ss_dayt_farily_active_min_sd GA AA 0.16007223 0.29346576 Ctrl 3STD GG GA 0.03753357 0.05663644 adhd 9 6 InterdayCV GG GA 0.01271714 0.03330681 adhd 6InterdaySD GG GA 0.03810088 0.05663644 adhd 2 6 IntradayCV mean GG GA 0.00247677 0.01513583 adhd 2 1 IntradayCV median GG GA 0.00355046 0.01952755 adhd 4 2 IntradaySD_mean GG GA 0.00642850 0.02115482 adhd 4 9 IntradaySD median GG GA 0.00654914 0.02115482 adhd 9 9 fit_ss_day_total_steps_no_el_mean GG GA 0.00197641 0.01506358 adhd 1 9 fit_ss_total_step_mean GG GA 0.00210276 0.01506358 adhd 2 9 fit_ss_total_ave_met_mean GG GA 0.03480371 0.05630013 adhd 7 fit_ss_total_fairly_active_min_mea GG GA 0.01614723 0.03861294 adhd n 2 fit ss fitbit totalsteps mean GG GA 0.00201773 0.01506358 adhd 2 9 fit ss fitbit fairlyactivemin mean GG GA 0.01755617 0.04023289 adhd 1 2 fit_ss_dayt_total_steps_mean GG GA 0.00219106 0.01506358 adhd 8 9 fit_ss_dayt_ave_met_value_mean GG GA 0.02938934 0.05214238 adhd 6 9fit_ss_dayt_farily_active_min_mea GG GA 0.01587768 0.03861294 adhd n 2 2 fit_ss_day_total_steps_no_el_max GG GA 0.00174436 0.01506358 adhd 9 9 fit ss total step max GG GA 0.00132504 0.01506358 adhd 4 9 fit_ss_total_fairly_active_min_max GG GA 0.02717834 0.05154514 adhd 7 fit ss fitbit totalsteps max GG GA 0.00181877 0.01506358 adhd 6 9 fit_ss_fitbit_fairlyactivemin_max GG GA 0.02938934 0.05214238 adhd 6 9 fit_ss_dayt_total_steps_max GG GA 0.00132504 0.01506358 adhd 4 9fit ss dayt farily active min max GG GA 0.02592187 0.05091796 adhd 3 6 fit_ss_day_total_steps_no_el_medi GG GA 0.00630986 0.02115482 adhd an 1 9 fit_ss_total_step_median GG GA 0.00563848 0.02115482 adhd 5 9 fit_ss_total_fairly_active_min_med GG GA 0.03480371 0.05630013 adhd ian 7 fit_ss_fitbit_totalsteps_median GG GA 0.00585475 0.02115482 adhd 4 9 fit ss fitbit fairlyactivemin media GG GA 0.03642051 0.05663644 adhd n 4 6 fit ss dayt total steps median GG GA 0.00563848 0.02115482 adhd 5 9 fit_ss_dayt_farily_active_min_med GG GA 0.03480371 0.05630013 adhd ian 7 fit_ss_day_total_steps_no_el_sd GG GA 0.00909827 0.02502026 adhd 9 7 fit_ss_total_step_sd GG GA 0.00667182 0.02115482 adhd 7 9 fit ss total fairly active min sd GG GA 0.02138517 0.04704739 adhd 8 1fit_ss_fitbit_totalsteps_sd GG GA 0.00831358 0.02406565 adhd 9 3 fit ss fitbit fairlyactivemin sd GG GA 0.02432491 0.04955074 adhd 3 9 fit ss dayt total steps sd GG GA 0.00692339 0.02115482 adhd 8 9 fit_ss_dayt_farily_active_min_sd GG GA 0.02281393 0.04826025 adhd 7 1feature genotype genotype p value FDR value disease group 1 group 2 group fit ss excl day min max AA AC 0.221453753 0.576431093 Ctrl fit ss excl day min max AA CC 0.038728725 0.274199373 Ctrl fit ss excl day min max AC CC 0.332708869 0.690162873 Ctrl fit_ss_excl_day_min_nohr_max AA AC 0.107172407 0.446355484 Ctrl fit_ss_excl_day_min_nohr_max AA CC 0.019472719 0.229778087 Ctrl fit_ss_excl_day_min_nohr_max AC CC 0.31075902 0.690162873 Ctrl fit ss fitbit sedentarymin max AA AC 0.163115629 0.515561899 Ctrl fit_ss_fitbit_sedentarymm_max AA CC 0.084043826 0.391467297 Ctrl fit ss fitbit sedentarymin max AC CC 0.538804068 0.82214069 Ctrl fit ss day min sd AA AC 0.828111623 0.943433167 Ctrl fit ss day min sd AA CC 0.191298127 0.531522966 Ctrl fit_ss_day_min_sd AC CC 0.190582029 0.531522966 Ctrl fit ss excl day min sd AA AC 0.296912632 0.690162873 Ctrl fit ss excl day min sd AA CC 0.081359069 0.38995133 Ctrl fit ss excl day min sd AC CC 0.322160744 0.690162873 Ctrl fit_ss_excl_day_min_nohr_sd AA AC 0.246390912 0.608731151 Ctrl fit_ss_excl_day_min_nohr_sd AA CC 0.062739709 0.382082437 Ctrl fit ss excl day min nohr sd AC CC 0.290983408 0.686720843 Ctrl fit ss total min sd AA AC 0.352039588 0.708079626 Ctrl fit_ss_total_min_sd AA CC 0.064759735 0.382082437 Ctrl fit ss total min sd AC CC 0.227306006 0.58308932 Ctrl fit ss excl day min max AA AC 0.019300777 0.189790975 adhd fit ss excl day min max AA CC 1.10E-06 9.73E-05 adhd fit_ss_excl_day_min_max AC CC 0.000781268 0.019754923 adhd fit_ss_excl_day_min_nohr_max AA AC 0.179938217 0.61248201 adhd fit_ss_excl_day_min_nohr_max AA CC 3.68E-06 0.000217089 adhd fit_ss_excl_day_min_nohr_max AC CC 8.67E-05 0.003836857 adhd fit_ss_fitbit_sedentarymin_max AA AC 0.004675845 0.066497213 adhdfit ss fitbit sedentarymin max AA CC 3.87E-07 6.84E-05 adhd fit ss fitbit sedentarymin max AC CC 0.010582929 0.124878558 adhd fit ss day min sd AA AC 0.812300676 0.910295093 adhd fit ss day min sd AA CC 0.001393072 0.02585576 adhd fit ss day min sd AC CC 0.001414239 0.02585576 adhd fit ss excl day min sd AA AC 0.18917994 0.631789613 adhd fit ss excl day min sd AA CC 0.000666455 0.019660431 adhd fit ss excl day min sd AC CC 0.004883976 0.066497213 adhd fit_ss_excl_day_min_nohr_sd AA AC 0.331653603 0.804146408 adhd fit_ss_excl_day_min_nohr_sd AA CC 0.000666455 0.019660431 adhd fit_ss_excl_day_min_nohr_sd AC CC 0.001713531 0.027572267 adhd fit ss total min sd AA AC 0.138705973 0.580508997 adhd fit ss total min sd AA CC 0.001460777 0.02585576 adhd fit ss total min sd AC CC 0.009374624 0.118522036 adhd

[0232] Results of the post-hoc univariate test conducted on the two significant loci reported in Table S2. For each locus, we list the wearable-derived static features from the corresponding cluster that show significant differences among genotype groups of individuals, considering the genotype of the individuals at the lead variant for the locus (see Table S2). For a description of each feature see Data S45-46. We performed pairwise comparisons separately for individuals with ADHD and control individuals. For each comparison we report the feature, the two genotype groups used for the comparison, the p value (two-sided Wilcoxon Rank-Sum test), the Benjamini -Hochberg False Discovery Rate, and the disease group (ctrl = control individuals; ADHD = individuals with ADHD). We included features that show an FDR value < 0.1 in at least one of the three ADHD comparisons. See also FIGS. 37-38.Data S54. Co-localization analysis on the loci identified by the continuous univariate GWAS for ADHD, related to Figure 4, Table 1, Figure S4, Table S3, and STAR Methods “GWASs for ADHD” SectionChr Start End Lead Variant p value Trait Publication PP(i-f4> 11 38982793 393848ft) tsl51239852 3.80E-08 Neuroticism measuremorit Turley etal.. Nat Genet 2018 9.17E-04 17 32256997 32283356 rs6505293 1.12E-08 ChroBolype meastiiemeni denes etai. Nat Commun 2019 3.03E-02 17 32256997 32283356 rs6505293 i.izE-08 Chrwiof / pe measurement Jansen el st Nat Genet 2019 9.96E-017 68219282 68338849 13115229175 4.46E-07 ADHD 2023 Dementis etai.. Nat Genet 2019 2.45E-01

[0233] For each tested locus (listed in Table 1 and Table S3), we report genomic coordinates in assembly GRCh38, the lead variant and corresponding p value as obtained fromour GWAS, the trait for which we conducted the colocalization analysis, the publication that provided the GWAS summary statistics for the tested trait, and the Posterior Probability for accepting the colocalization hypothesis (“PP(H4)”).Data S55. Significant features identified by the post-hoc test for the continuous multivariate GW AS for behavioral traits, related to Figure 5, Table 1, Figure S5, Table S4, and STAR Methods “Continuous Multivariate GWAS for behavioral traits” Section feature genotype genotype p value FDR value group 1 group 2Mean AA AG 1.60E-08 3.37E-07 Mean AA GG 1.05E-06 7.38E-06 Mean AG GG 0.50413190 0.699486134 1 Median AA AG 1.96E-07 2.05E-06 Median AA GG 2.84E-06 1.49E-05 Median AG GG 0.45046366 0.699486137 1 InterdayCV AA AG 0.00126307 0.005894369 7 InterdayCV AA GG 0.02346393 0.070391805 5 InterdayCV AG GG 0.94948754 0.995789845 7 IntradayCV mean AA AG 0.02215692 0.070391805 5 IntradayCV mean AA GG 0.27274886 0.528954772 IntradayCV mean AG GG 0.58608553 0.745927059 IntradayCV median AA AG 0.00708075 0.029739188 3 IntradayCV_median AA GG 0.22556888 0.498625953 1IntradayCV median AG GG 0.48074236 0.699486131 IntradayCV_sd AA AG 0.01806357 0.063222515 2 IntradayCV_sd AA GG 0.04424862 0.116152645 IntradayCV_sd AG GG 0.76207226 0.889084304 8 IntradaSD sd AA AG 0.01806357 0.063222515 2 IntradaSD_sd AA GG 0.04424862 0.116152645IntradaSD sd AG GG 0.76207226 0.889084304 8 IntradayMean mean AA AG 1.12E-08 3.37E-07 IntradayMean mean AA GG 1.90E-06 1.14E-05 IntradayMean mean AG GG 0.61284227 0.757040455 8 IntradayMean median AA AG 2.75E-08 3.85E-07 IntradayMean median AA GG 2.62E-07 2.20E-06 IntradayMean median AG GG 0.30279232 0.528954779 2feature genotype genotype p value FDR value group 1 group 2fit_ss_outbed_minutes_mean AA AG 0.00019892 0.002652281 5 fit_ss_outbed_minutes_mean AA GG 0.12865967 0.291160253 2 fit_ss_outbed_minutes_mean AG GG 0.71119062 0.820604561 3 fit_ss_wa ke u p_mi n utes_mea n AA AG 0.00029716 0.003565975 5 fit_ss_wa ke u p_mi n utes_mea n AA GG 0.13102211 0.291160253 2 fit_ss_wa ke u p_mi n utes_mea n AG GG 0.65088219 0.765743754 7 f it_ss_o ut bed_m in utes_mi n AA AG 2.04E-05 0.000408385 f it_ss_o ut bed_m in utes_mi n AA GG 0.01154540 0.072918379 5f it_ss_o ut bed_m in utes_mi n AG GG 0.31945360 0.518032877 7 fit_ss_wakeup_minutes_min AA AG 2.92E-05 0.000438149 fit_ss_wakeup_minutes_min AA GG 0.02257390 0.117776895 7 fit_ss_wakeup_minutes_min AG GG 0.38311991 0.581954309 7 fit_ss_o ut bed_m in utes_med i a n AA AG 0.00177667 0.019381948 7 f it_ss_o ut bed_m in utes_med i a n AA GG 0.07168643 0.20981394 f it_ss_o ut bed_m in utes_med i a n AG GG 0.47919166 0.665721856 6 f it_ss_wa ke u p_ m i n ut es_med i a n AA AG 0.00267030 0.026703066 3 f it_ss_wa ke u p_ m i n ut es_med i a n AA GG 0.06026995 0.206341936 5 f it_ss_wa ke u p_ m i n ut es_med i a n AG GG 0.39672856 0.595092842 4 fit_ss_total_light_active_min_min AA AG 0.10914391 0.261945408 2 fit_ss_total_light_active_min_min AA GG 0.00671433 0.047395282 1 fit_ss_total_light_active_min_min AG GG 0.04212305 0.173754752 fit_ss_total_very_active_min_min AA AG 2.27E-21 9.57E-20 fit_ss_total_very_active_min_min AA GG 0.00404467 0.033244131 2 fit_ss_total_very_active_min_min AG GG 0.81779459 0.899120291 4 fit_ss_fitbit_lightlyactivemin_min AA AG 3.43E-02 1.53E-01 fit_ss_fitbit_lightlyactivemin_min AA GG 0.00858081 0.057205412 6 fit_ss_fitbit_lightlyactivemin_min AG GG 0.06653062 0.206341938 5 f it_ss_f it b it_ve ry active mi n_m in AA AG 2.39E-21 9.57E-20 f it_ss_f it b it_ve ry active mi n_m in AA GG 0.00415551 0.033244137 2 fit_ss_fitb it_ve ry active mi n_m in AG GG 0.81779459 0.899120291 4 fit_ss_dayt_light_active_min_min AA AG 1.00E-01 0.253887635 fit_ss_dayt_light_active_min_min AA GG 0.00645918 0.047395282 1 fit_ss_dayt_light_active_min_min AG GG 0.03968755 0.170089502 fit_ss_dayt_very_active_min_min AA AG 1.14E-21 9.57E-20 fit_ss_dayt_very_active_min_min AA GG 0.00389303 0.033244133 2 fit_ss_dayt_very_active_min_min AG GG 0.82419360 0.899120293 4feature genotype genotype p value FDR value group 1 group 2fit_ss_outbed_minutes_mean GG GA 0.00227555 0.018204446 9 fit_ss_outbed_minutes_mean GG AA 0.00029144 0.002914444 2 fit_ss_outbed_minutes_mean GA AA 0.00828689 0.041434487 5 fit_ss_wa ke u p_mi n utes_mea n GG GA 0.02084613 0.086259876 3 fit_ss_wa ke u p_mi n utes_mea n GG AA 0.00022721 0.002478726 1 fit_ss_wa ke u p_mi n utes_mea n GA AA 0.00324857 0.024252886 6 f it_ss_o ut bed_m in utes_mi n GG GA 1.00E-07 1.71E-06 f it_ss_o ut bed_m in utes_mi n GG AA 0.00016792 0.002180468 8 f it_ss_o ut bed_m in utes_mi n GA AA 0.04509278 0.117633368 fit_ss_wakeup_minutes_min GG GA 1.33E-06 2.00E-05 fit_ss_wakeup_minutes_min GG AA 0.00018170 0.002180466 8 fit_ss_wakeup_minutes_min GA AA 0.02884443 0.091873589 4 f it_ss_l ig ht_min utes_mi n GG GA 0.43537082 0.580494436 4 f it_ss_l ig ht_min utes_mi n GG AA 0.04169467 0.113712742 2 f it_ss_l ig ht_min utes_mi n GA AA 0.03234252 0.094661044 6 fit_ss_deep_minutes_min GG GA 0.02411412 0.090736816 8 fit_ss_deep_minutes_min GG AA 0.58198008 0.698671072 7 fit_ss_deep_minutes_min GA AA 0.28034464 0.410260452 2 f it_ss_wa ke_count_min GG GA 0.72596720 0.770938627 6 f it_ss_wa ke_count_min GG AA 0.02909330 0.091873582 4 f it_ss_wa ke_count_min GA AA 0.03898038 0.108782451 8 f it_ss_o ut bed_m in utes_med i a n GG GA 0.00432889 0.028859268 fit_ss_o ut bed_m in utes_med i a n GG AA 0.00125835 0.010785939 3 f it_ss_o ut bed_m in utes_med i a n GA AA 0.01366660 0.060740452 1 fit_ss_wakeup_minutes_median GG GA 0.00343582 0.024252886 6 fit_ss_wakeup_minutes_median GG AA 0.00096138 0.008874325 4 fit_ss_wakeup_minutes_median GA AA 0.01202657 0.055507264 5f i t_ss_d ay_tota l_st e ps_n o_e l_m i n GG GA 0.12888581 0.220947115 2 f i t_ss_d ay_tota l_st e ps_n o_e l_m i n GG AA 0.00939984 0.045119255 4 f i t_ss_d ay_tota l_st e ps_n o_e l_m i n GA AA 0.03113812 0.093414389 6 fit_ss_tota l_ste p_m in GG GA 0.08850155 0.163387497f it_ss_t ot a l_ste p_m in GG AA 0.00729459 0.039788671 6 fit_ss_t ota l_ste p_m in GA AA 0.02997242 0.092222859 7 fit_ss_total_light_active_min_min GG GA 0.04904339 0.121597146 8 fit_ss_total_light_active_min_min GG AA 0.00552814 0.033168899 4 fit_ss_total_light_active_min_min GA AA 0.02722104 0.090736815 8 f it_ss_f it b it_tot a lsteps_mi n GG GA 2.24E-01 3.50E-01 f it_ss_f it b it_tot a lsteps_mi n GG AA 0.01508399 0.064645672 8 f it_ss_f it b it_tot a lsteps_mi n GA AA 0.04509278 0.117633368 fit_ss_fitbit_lightlyactivemin_min GG GA 7.70E-02 1.47E-01 fit_ss_fitbit_lightlyactivemin_min GG AA 0.00788639 0.041146393 6 fit_ss_fitbit_lightlyactivemin_min GA AA 0.03552939 0.101512542 9 fit_ss_dayt_light_active_min_min GG GA 0.07224235 0.142116098 fit_ss_dayt_light_active_min_min GG AA 0.00506987 0.032020231 9 fit_ss_dayt_light_active_min_min GA AA 0.02236618 0.089464759 5 feature genotype genotype p value FDR value group 1 group 2fit_ss_excl_n ig ht_mi n_mea n AA AG 2.07E-05 0.000180439 fit_ss_excl_n ig ht_mi n_mea n AA GG 0.02861137 0.066114343 6 fit_ss_excl_n ig ht_mi n_mea n AG GG 0.35653996 0.462538336 4 fit_ss_excl_n ig ht_mi n_h r50_mea n AA AG 9.97E-07 6.98E-05 fit_ss_excl_n ig ht_mi n_h r50_mea n AA GG 0.00120128 0.005241977 9 fit_ss_excl_n ig ht_mi n_h r50_mea n AG GG 0.11147947 0.188760275 fit_ss_excl_n ig ht_min_h r_rept_mea AA AG 0.00228567 0.00914267 n 9 fit_ss_excl_n ig ht_min_h r_rept_mea AA GG 0.07055481 0.13280906n 4 2fit_ss_excl_n ig ht_min_h r_rept_mea AG GG 0.30815751 0.40524823 n 2 fit_ss_excl_sleep_min_mean AA AG 0.00011623 0.000743873 fit_ss_excl_sleep_min_mean AA GG 0.05414895 0.10973194 fit_ss_excl_sleep_min_mean AG GG 0.40295165 0.508991563 2 fit_ss_excl_sleep_min_hr50_mean AA AG 2.92E-06 6.98E-05 fit_ss_excl_sleep_min_hr50_mean AA GG 0.00287489 0.011039595 8 fit_ss_excl_sleep_min_hr50_mean AG GG 0.14255355 0.231951557 1 fit_ss_excl_sleep_min_rept_mean AA AG 0.00684002 0.020987544 7 fit_ss_excl_sleep_min_rept_mean AA GG 0.12024745 0.199030279 7 fit_ss_excl_sleep_min_rept_mean AG GG 0.41674461 0.519577692 9 fit_ss_excl_n ig ht_mi n_mi n AA AG 0.00170803 0.00712928 fit_ss_excl_n ig ht_mi n_mi n AA GG 0.00031626 0.001597943 f it_ss_excl_ n ig ht_mi n_mi n AG GG 0.00729926 0.021234212 6 fit_ss_excl_n ig ht_mi n_h r50_mi n AA AG 0.19587364 0.29381046 fit_ss_excl_n ig ht_mi n_h r50_mi n AA GG 0.02200635 0.057097564 7 fit_ss_excl_n ig ht_mi n_h r50_mi n AG GG 0.05884890 0.115295801 6 fit_ss_excl_sleep_min_min AA AG 0.00342454 0.012644474 2 fit_ss_excl_sleep_min_min AA GG 0.02338332 0.057558941 fit_ss_excl_sleep_min_min AG GG 0.15625237 0.250003793 6 fit_ss_excl_sleep_min_hr50_min AA AG 0.18829222 0.286921487 9 fit_ss_excl_sleep_min_hr50_min AA GG 0.00580019 0.018560639 7 fit_ss_excl_sleep_min_hr50_min AG GG 0.02537658 0.060903805 5 fit_ss_excl_n ig ht_mi n_max AA AG 1.71E-05 0.000164461 fit_ss_excl_n ig ht_mi n_max AA GG 0.06553029 0.125818165 6 f it_ss_excl_ n ig ht_mi n_max AG GG 0.65423187 0.692125704 4 f it_ss_excl_ n ig ht_mi n_h r50_max AA AG 5.09E-06 6.98E-05 fit_ss_excl_n ig ht_mi n_h r50_max AA GG 0.00452883 0.015681014 4 fit_ss_excl_n ig ht_mi n_h r50_max AG GG 0.22745462 0.325905149fit_ss_excl_n ig ht_min_h r_rept_max AA AG 0.00021707 0.001225847 9 fit_ss_excl_n ig ht_min_h r_rept_max AA GG 0.10715291 0.187030546 4 fit_ss_excl_n ig ht_min_h r_rept_max AG GG 0.59456674 0.656073655 fit_ss_excl_sleep_min_max AA AG 5.85E-05 0.000408097 fit_ss_excl_sleep_min_max AA GG 0.05023201 0.107161649 1 fit_ss_excl_sleep_min_max AG GG 0.44896968 0.552578075 4 fit_ss_excl_sleep_min_hr50_max AA AG 4.76E-06 6.98E-05 fit_ss_excl_sleep_min_hr50_max AA GG 0.00521017 0.017247459 fit_ss_excl_sleep_min_hr50_max AG GG 0.23643678 0.328955522 3 fit_ss_excl_sleep_min_rept_max AA AG 0.00089360 0.004085069 8 fit_ss_excl_sleep_min_rept_max AA GG 0.02261983 0.057144834 fit_ss_excl_sleep_min_rept_max AG GG 0.16973251 0.262811631 fit_ss_excl_n ig ht_mi n_med ia n AA AG 0.00457362 0.015681019 4 fit_ss_excl_n ig ht_mi n_med ia n AA GG 0.04121418 0.089921855 9 fit_ss_excl_n ig ht_mi n_med ia n AG GG 0.23593128 0.328955527 3 fit_ss_excl_n ig ht_min_h r50_med ia AA AG 2.44E-06 6.98E-05 nfit_ss_excl_n ig ht_min_h r50_med ia AA GG 5.95E-05 0.00040809 n 7 fit_ss_excl_n ig ht_min_h r50_med ia AG GG 0.00968061 0.02733349 n 3 6 fit_ss_excl_sleep_min_median AA AG 0.01421503 0.038223572 1 fit_ss_excl_sleep_min_median AA GG 0.27506732 0.377235196 fit_ss_excl_sleep_min_median AG GG 0.56483333 0.646306114 fit_ss_excl_sleep_min_hr50_media AA AG 4.50E-06 6.98E-05 nfit_ss_excl_sleep_min_hr50_media AA GG 0.00029425 0.00156933 n 1 fit_ss_excl_sleep_min_hr50_media AG GG 0.02892502 0.06611434 n 7 6 fit_ss_excl_n ig ht_mi n_sd AA AG 9.37E-06 9.99E-05 f it_ss_excl_ n ig ht_mi n_sd AA GG 0.11207641 0.188760273 5 f it_ss_excl_ n ig ht_mi n_sd AG GG 0.76833963 0.786049726 fit_ss_excl_n ig ht_mi n_h r50_sd AA AG 3.39E-06 6.98E-05fit_ss_excl_n ig ht_mi n_h r50_sd AA GG 0.00699584 0.020987549 7 fit_ss_excl_n ig ht_mi n_h r50_sd AG GG 0.29795936 0.401246843 6 fit_ss_excl_n ig ht_mi n_h r_rept_sd AA AG 0.00016216 0.000973018 1 fit_ss_excl_n ig ht_mi n_h r_rept_sd AA GG 0.09647887 0.171517991 4 fit_ss_excl_n ig ht_mi n_h r_rept_sd AG GG 0.57350389 0.647722044 5 fit_ss_excl_sleep_min_sd AA AG 2.90E-05 0.000232349 fit_ss_excl_sleep_min_sd AA GG 0.08188571 0.151173636 fit_ss_excl_sleep_min_sd AG GG 0.55365674 0.646306111 4 fit_ss_excl_sleep_min_hr50_sd AA AG 6.82E-06 8.18E-05 fit_ss_excl_sleep_min_hr50_sd AA GG 0.01433383 0.038223579 1 fit_ss_excl_sleep_min_hr50_sd AG GG 0.39014034 0.499379644 1 fit_ss_excl_sleep_min_rept_sd AA AG 0.00064893 0.003114896 2 fit_ss_excl_sleep_min_rept_sd AA GG 0.04061952 0.089921857 9 fit_ss_excl_sleep_min_rept_sd AG GG 0.30093513 0.401246844 6

[0234] Results of the post-hoc univariate test conducted on the four significant loci reported in Table S4. For each locus, we list the wearable-derived static features from the corresponding cluster that show significant differences among genotype groups of individuals, considering the genotype of the individuals at the lead variant for the locus (see Table S4). For a description of each feature see Data S45-S46. For each comparison we report the feature, the two genotype groups used for the comparison, the p value (two-sided Wilcoxon Rank-Sum test), and the Benjamini -Hochberg False Discovery Rate. We included features that showed an FDR value < 0.1 in at least one of the three genotype comparisons. See also Data S33-36.Data S56. Genome-wide significant loci identified by the continuous multivariate GWAS for behavioral traits on chromosome X, related to Figure 5, Table S4, and STAR Methods“Continuous GWAS for behavioral traits” & “Statistical Significance and Functional Dissection of GWAS” SectionsLocus Chr Start End Lead Variant Position p value GWAS ran1 X 124,254, 095 124,257,377 rs 12012355 124,254,420 4.80E-08 multivariate GWAS, mates duster 1 2 X 7,037,241 7,670,194 - 7,432,023 5.17E-09 multivariate GWAS, males cluster 7 3 X 9,806,961 9,618,398 rs112248037 9,618.398 4.54E-08 multivariate GWAS, mates cluster 7

[0235] Analogous representation to Table S4 for variants located on chromosome X.Example 2 - Digital phenotyping disentangles genetic and environmental determinants of severity and progression rate in Parkinson’sAbstract

[0236] Digital health technologies are transforming disease detection and management by providing scalable, non -invasive measures of human physiology and behavior.Neurodegenerative disorders (NDDs) such as Parkinson’s disease (PD) are particularly well-suited to digital phenotyping, as symptoms evolve gradually and heterogeneously, complicating diagnosis and staging. Wearable devices can capture temporal activity patterns reflective of disease severity, but existing approaches often reduce signals to summary features or binary case-control labels, limiting their clinical and mechanistic utility. Here, we present an interpretable Al framework that models temporal activity patterns from wearable-derived accelerometry to generate three complementary, disease-anchored digital phenotype constructs: (i) a severity score that captures case-control PD-likeness, (ii) a progression (stagelikeness) score that distinguishes prodromal from diagnosed PD, and (iii) a progression rate / velocity. To ensure interpretability, severity scores were decomposed into feature-level components, producing multi-dimensional digital phenotypes that align wearable signals with PD-relevant behavioral domains. Using accelerometry and genotype data from the UK Biobank, we applied these phenotypes in genome-wide association studies (GWAS). Univariate continuous GWAS (UCGWAS) with severity scores increased power compared to traditional binary phenotypes, while multivariate continuous GWAS (MVCGWAS) of decomposed features uncovered additional associations masked in composite analyses (Gene X and Y). Beyond increased discovery, this approach enables interpretability: genetic variants were mapped not only to overall disease risk but to specific activity- and time-resolved features, establishing a tri-axial link between genotype, digital phenotype, and disease. Identified loci implicated pathways related to brain function, metabolism, and mitochondria, consistent with PD biology. Together, these findings demonstrate the potential of interpretable digital phenotype GWAS to uncover disease mechanisms and advance precision neurology.Introduction

[0237] Digital health technologies are increasingly adopted for disease detection, management, and care quality. Wearable devices, in particular, have emerged as a key tool within the broader digital health toolbox, offering continuous, non-invasive, and scalable means to capture physiological and behavioral changes in real-world settings. This is especially relevant for neurodegenerative disorders (NDDs) such as Parkinson’s disease (PD), where symptoms evolve gradually and heterogeneously over time. The ability of wearables to generate continuous multichannel time-series data creates new opportunities for characterizing disease stratification and progression; however, two major challenges remain. First, more advanced methods are required to develop and validate Al frameworks that can fully leverage wearable-derived digital phenotypes to capture disease dynamics. Second, there remains a gap in integrating these rich digital phenotypes with genomic data: most genome-wide association studies (GWAS) still rely on univariate, binary case-control definitions, which fail to capture the complexity of such diseases. Moreover, identifying a significant GWAS locus using case-control phenotypes does not directly explain its functional role or mechanistic link to disease. Therefore, addressing both challenges - advancing interpretable Al approaches for digital phenotyping and integrating them with genetic discovery - will be essential to better dissect mechanisms of NDDs such as PD and to advance precision medicine.

[0238] PD, the second most common NDD and the leading age-related movement disorder, illustrates the challenges and opportunities of digital phenotyping. The prodromal stage is often characterized by non-motor symptoms such as REM sleep behavior disorder, while the diagnosed stage is dominated by motor impairments including bradykinesia, tremor, and reduced mobility. Overlap between motor and non-motor features complicates staging and subtyping, resulting in considerable clinical heterogeneity. Recent studies, including large-scale analyses of accelerometer-derived features, have demonstrated that wearable data can improve discrimination of PD cases versus controls. However, most existing approaches reduce complex sensor signals to summary statistics or a condensed set of features and rely on traditional linearmodels. As a result, they are limited in their ability to capture temporal dynamics or disease severity, both of which are critical for characterizing heterogeneous, evolving conditions such as PD

[0239] Another major challenge lies in leveraging digital phenotypes for genetic discovery. PD is estimated to have a genetic heritability of 15-36%, with numerous risk loci identified through GWAS. While GWAS has been instrumental in mapping common variants, the effect sizes of individual SNPs are small, necessitating very large cohorts (and often proxy cases) for adequate power. More fundamentally, most GWAS continue to rely on univariate, binary case-control phenotypes, which fail to capture the nuanced patterns of disease expression and progression revealed through digital phenotyping. Some progress has been made by extending GWAS to univariate, continuous wearable-derived traits, such as sleep scores, but these measures remain largely opaque in their biological interpretation. Therefore, there is a need to develop frameworks that not only bridge quantitative digital phenotypes with genetic discovery but also ensure interpretability by decomposing composite scores into functional components that can connect genetic associations to their mechanistic relevance.

[0240] To address these challenges, we developed an Al prediction framework that captures temporal activity patterns of PD, using them not only for disease prediction but also as a foundation to connect digital phenotypes with genetics and enable a more interpretable dissection of functional roles in disease. We applied our approach to the UK Biobank, a large prospective population cohort with multimodal data including wearable-derived accelerometry and genotype information collected between 2013 and 2016. To comprehensively characterize PD heterogeneity, we processed hourly resolved activity features derived from accelerometers, capturing both temporal and behavioral dynamics. These features, combined with demographic covariates, were used in machine learning models to predict clinical diagnosis. From these models, we extracted individual severity scores — univariate continuous values that distinguish PD cases from healthy controls and reflect disease progression. To ensure interpretability, we decomposed each severity score into 648 dimensions using SHapley Additive exPlanations (SHAP), quantifying the contribution of each wearable feature to the overall score. For each individual, the resulting decomposed vector constitutes a multi-dimensional digital phenotype, suitable for downstream genetic discovery. This decomposition is a key step, as it ensures that each component reflects a PD-relevant contribution rather than an arbitrary raw sensor measure,thereby grounding genetic analyses in biologically and clinically interpretable traits. To link these digital phenotypes with genetic discovery, we performed a series of genome-wide association studies (GWAS). We first used the severity score in a univariate continuous (UC) GWAS. To further leverage the decomposed feature space, we conducted multivariate continuous (MVC) GWAS, revealing how specific activities and behaviors relate to genetic variation and clinical diagnosis. Beyond demonstrating that both approaches increase power over traditional casecontrol GWAS, the MVC framework uncovered additional associations that remained masked in the UC analysis. Overall, this strategy bridges digital phenotyping with genetic discovery, establishing a tri -axial link between genotype, digital phenotype, and disease.ResultDefining the PD cohort and feature engineering

[0241] To explore the temporal and behavioral dynamics of PD, we identified 648 wearable-derived features from accelerometer data across PD cases and healthy controls. These features covered nine activity types, averaged over seven days: raw acceleration (milligravity) and activities inferred from acceleration, including metabolic equivalent of task (MET; kcal / kg / hour), bicycling, mixed activities, sitting / standing, sleep, vehicle time, walking, and device wear time (proportion of time per hour). Each activity type was recorded with hourly resolution (24 time points) across three time contexts - weekday, weekend, and combined -resulting in a total of 648 features. In addition, we also employed demographic features of each participant in our analysis, including sex, BMI, age, socioeconomic status (SES), and smoking status. Our cohort consisted of 320 PD cases (114 prodromal and 316 diagnosed) and 2,758 controls (see methods for more details). The mean age of cases was 75.80 years (SD: 5.62), which was significantly higher (p<0.001) than that of controls (66.19 years, SD: 7.26). To address this imbalance, we performed 1:1 age matching for downstream modeling analyses. Intercorrelation and temporal trend of wearable digital hourly features

[0242] To assess the similarity among wearable-derived features, we performed hierarchical clustering based on the Pearson correlation to identify groups of highly related features across activity type, time context, and hourly time points. Our analysis revealed that most features clustered primarily by activity type rather than by time context or hourly time points. This finding suggests that activity type is the primary driver of patterns in the data.However, a subset of features did cluster according to morning hours, indicating that certain temporal patterns also contribute to feature similarity.

[0243] To investigate the heterogeneity of Parkinson’s disease (PD) across different activity types, we compared temporal trends in wearable-derived features among healthy control, diagnosed, and prodromal groups. The 24-hour temporal trend averaged across all seven days of the week was assessed for nine activity types, across three time contexts (weekend, weekday, and combined). Acceleration was measured in milligravity (mg), MET in kilocalories per kilogram per hour (kcal / kg / hour), and the remaining activity types were quantified as the fraction of a given hour spent in each activity. By first examining the temporal trends in acceleration across the three groups, we observed a consistent pattern: acceleration increased starting around 4:00 am, rose gradually to a peak between 10:00-11:59 am, and then declined over the remainder of the day. However, the magnitude of this increase differed across groups. Controls showed the most pronounced rise, with a weekend peak at 10:00-10:59 am (59.59 mg), compared to lower peaks in the prodromal (42.47 mg) and diagnosed (38.65 mg) groups.Additionally, the value of another movement-related feature, MET, was higher in control compared to prodromal and diagnosed group across all three time contexts, particularly peaked between 12:00-12:59 pm during the weekend (2.37, 2.20, and 2.20 kcal / kg / hour, respectively).

[0244] To further compare behavioral patterns across groups, we found that prodromal group slept more between 12 am and 7 am compared to the other two groups, which showed similar sleep patterns. Prodromal group slept the least during the morning period with lowest proportion observed at 3-3:59am compared to controls and diagnosed groups for weekend (0.85, 0.95, 0.92, respectively). In contrast, the prodromal group had a highest proportion sit or stand consecutively during early morning hours (12am - around 9am) across three time contexts. For example, prodromal groups sit or stand the most compared to controls and diagnosed groups at 3-3:59am during weekend (0.13, 0.03, 0.07). Similarly, prodromal groups have highest vehicle time across 8am - 11:59pm across three time contexts compared to diagnosed and controls, where it peaked at ll-ll:50am weekday (0.17, 0.13, 0.11, respectively). In addition, we observed controls had consecutively higher walking throughout the day (eg., ~6am to 11:59pm) across three time contexts compared to diagnosed and prodromal, where controls walk the most at 12-12:59pm during weekend (0.22, 0.12, 0.10, respectively).Identifying wearable and demographic features contributing to prediction of diagnosed, prodromal, and healthy controls

[0245] Given the substantial differences across prodromal, diagnosed, and healthy control groups in wearable-derived activity features, we developed a framework to systematically evaluate which features are most informative for predicting disease status.Therefore, we constructed an Al-based prediction pipeline using four age-matched comparison cohorts (1: 1 matching): case versus control (n = 774), diagnosed versus control (n = 574), prodromal versus control (n = 226), and prodromal versus diagnosed (n = 220). For each comparison, we developed three distinct feature sets to evaluate and compare model performance: (1) a baseline model incorporating demographic variables alone, (2) an intermediate model including both acceleration data and demographics, and (3) a full model combining acceleration, demographics, and 648 wearable-derived temporal activity features. To improve interpretability of temporal and behavioral contributions to disease severity and progression, we applied SHapley Additive exPlanations (SHAP) analysis to decompose model predictions at the feature level.Case vs control

[0246] To compare model prediction performance, we evaluated the mean AUROC for models using baseline, intermediate, and full models which achieved AUROCs of 0.638 ± 0.031, 0.789 ± 0.021, and 0.838 ± 0.032, respectively. Relative to the baseline model using demographic features alone, the inclusion of average acceleration significantly improved predictive performance (p = 0.008). The full model further improves upon both the baseline and intermediate model (p=0.008, p=0.008, respectively). This suggests that incorporating the full set of wearable-derived features increased discriminative power of time-series activity phenotypes in distinguishing between clinical groups.

[0247] To assess the most important features contributing to model predictions, we ranked features by their mean absolute SHAP values. Notably, acceleration at 9:00-9:59 am, along with sleep and sit / stand at 3:00— 3: 59 am across all seven days of the week, are the top three activity types at specific temporal hours contributing most to model predictions. To further interpret how these features relate to disease progression, we analyzed their raw values to determine the directionality of their association with disease status. Individuals with lower average acceleration between 9:00-9:59 am across the 7 days of a week were more likely to beclassified as cases. During 3:00-3:50am, individuals who slept the least while spent more time sitting or standing were more likely to be classified as cases. This result demonstrates that activities beyond acceleration also contribute to distinguishing PD cases from controls.Diagnosed vs control (Supplemental FIG. 1A)

[0248] To compare model prediction performance, the mean AUROC for baseline, intermediate, and full models were 0.653 ± 0.037, 0.802 ± 0.035, and 0.818 ± 0.027, respectively. Compared to baseline model, the inclusion of average acceleration alongside demographic features significantly enhanced model performance (p = 0.008), while the addition of the full wearable-derived feature set did not result in a statistically significant further improvement (p = 0.690). This indicated that average acceleration is the primary driver of model accuracy. Consistently, overall acceleration and hourly-specific acceleration features ranked among the top 10 contributors to model predictions. Reduced acceleration during the morning, afternoon, and late afternoon, such as at 9:00-9:59am and 12:00-12:59pm, was associated with a higher likelihood of being classified as diagnosed.Prodromal v.s Control

[0249] To compare model prediction performance, the mean AUROC for baseline, intermediate, and full models were 0.646 ± 0.028, 0.754 ± 0.022, and 0.888 ± 0,066, respectively. Compared to the baseline model, the addition of average acceleration alone or along with wearable-derived features statistically increased the model performance (p=0.012, p=0.008, respectively). Moreover, the addition of wearable-derived features upon the intermediate model further enhanced discriminative power (p=0.012). To further understand feature importance in driving the disease status, hourly- specific activity types beyond acceleration, such as sleep, sit / stand, and MET at 3am in weekdays or seven days of a week, emerged as predominant contributors to model prediction. This highlights that, in comparisons between prodromal and control groups, which are closer along the disease spectrum than diagnosed vs. control, more subtle behavioral features are necessary to capture meaningful differences.Prodromal vs Diagnosed

[0250] To compare model prediction performance, the mean AUROC for baseline, intermediate, and full models were 0.448 ± 0.077, 0.483 ± 0.045, and 0.762 ± 0048,respectively. The addition of average acceleration to demographics did not significantly improve model performance (p=0.548), while incorporating the full set of wearable-derived features significantly increased discriminative ability (p=0.008). These results indicate that wearable-derived features provide higher-resolution behavioral signals that offer valuable insights into the transition from the prodromal to diagnosed stages.

[0251] Unlike the case vs. control and diagnosed vs. control comparisons, where acceleration at 3:00 am and 9:00 am across all seven days emerged as the dominant temporal activity features, sleep and sit / stand during the early morning hours (2:00-5:00 am) across weekdays, weekends, and the combined week emerged as key contributors in this comparison. Notably, between 5:00-5:59am across all seven days of the week, individuals who spent more time sitting or standing, and less time sleeping or with lower MET values, were more likely to be classified as prodromal cases. Hence, the combination of activity type and specific time windows provides important information for discerning early signs of prodromal PD.Summary across four model comparison sets

[0252] Overall, while average acceleration remains a strong predictor in case vs. control and diagnosed vs. control comparisons, other activity patterns, such as early morning sleep and sit / stand, play a more prominent role in distinguishing disease status closer along the disease spectrum, specifically prodromal vs. diagnosed and prodromal vs. control. Given the symptom and movement overlap between prodromal and diagnosed individuals, increased sitting / standing and disruptions in early morning sleep may serve as early markers of PD risk and indicators of PD heterogeneity.MethodsFrom accelerometers to digital phenotype featuresData Collection Protocols and Preprocessing Pipeline

[0253] Participants wore wrist accelerometers continuously for ~7 days. Raw tri-axial acceleration aft) - (ax(f), ay(t), az( )) data was saved and calibrated to gravitational milligravity (mg) using standard autocalibration on static periods. We resampled to a uniform frequency and removed obvious artifacts (clipping, spikes).

[0254] Non-wear detection employed a standard inactivity and variance methodology, wherein sustained periods of near-zero vector magnitude accompanied by minimal variability across extended temporal windows were identified as indicative of device removal. Non-wear periods were flagged and excluded from activity inference. Wear-time per hour was tracked and subsequently retained as a covariate while also being summarized as an hourly feature for downstream analysis.Epoch-Level Feature Extraction and Activity Characterization

[0255] Signals were segmented into short epochs. For each epoch, we computed vector magnitude VM(t) — y / ax(t~)2-I- ay(t)2+ az(t)2and derived summary features capturing overall activity intensity and variability, along with estimated wear time and time spent in different activity states (sleep, sedentary, light, moderate-to-vigorous, and vigorous). A validated activity-inference model from accelerometer package assigned each epoch to an activity class:A = {sleep, sit / stand, walking, vehicle, bicycling, mixed / other}.In parallel, Metabolic Equivalent of Task MET(t) is derived from VM(t) based on the package’s calibration curve.Hourly aggregation and weekday / weekend contexts

[0256] We aligned epochs to clock time and aggregated to hourly features for each day d and hour h G {0,...,23}:1 - • Acceleration ( vmg): acc,rf>, = - nvdft| ^ 7^tewdihVM(tY7• Activity fractions: for each a E < A,ri,d,h,a = I w I y VepochW - 41 Wd'h 1 tEWd,h1MET: met;dih= MET (t).• i..e|MZd ftnwear|Wear-time fraction: wi d h= —:.Here indexes times within hour h on day d We then computed three time-contexts by averaging across days: weekday, weekend, and combined (all 7 days). This yields for each participant i a concatenated wearable feature vector.Xj [ 3CC, met, ^bicycling! hnixedi Tit standi Tleep' ^vehicle' ^walking' tV ]■24X3 24X3 24x3 24x3 24x3 24x3 24x3 24x324X 3Demographics features Ci(age, sex, BMI, SES, smoking and ancestry PCs) were collected in parallel.

[0257] Inclusion and quality control procedures mandated adequate temporal coverage, specifically a minimum of 16 hours of wear time across at least 5 days. Missing hourly data cells were imputed using a conservative approach that incorporated participant-specific robust central tendency estimates within the appropriate temporal context, and all imputed values were systematically flagged for transparency in subsequent analyses.Supervised models and out-of-fold generation of disease constructs

[0258] We define two supervised tasks and generate out-of-fold predictions to avoid leakage whenever model outputs are used as genetic traits. Each individual i was randomly assigned to one of the fold, where each fold was —1:1 age matched in sample size, with frequency matching in a 5-year old interval.• Case-control task (any PD vs control). Labels ysev) E {0,1}, where 0 is control, 1 is any PD• Progression task (diagnosed vs prodromal PD). Labels yprog) e {0,1}, where 0 is prodromal PD, 1 is diagnosed PDFor each task we train a random forest on (x,, c within A" -fold cross-validation, producing held-out probabilities pt. We are working with probability scale.Construct 1: Severity (case-control)> (sev)Si = PtConstruct 2: Progression (prodromal-diagnosed)Construct 3: Progression rate / velocity

[0259] With diagnosis date tdx, wearable date t;wear, define TTDj -I t,wear-t(dx|(years). ForPD only:-1TTDj

[0260] Interpretation: severity-per-year of moving toward diagnosed-like wearable patterns.

[0261] All three constructs are standardized within ancestry before genetics.SHAP: anchoring digital phenotypes to the disease construct

[0262] We compute SHAP values from shap. Explainer() for out-of-fold predictions. We work on the probability scale for additivity since the default shap. links (identity) used. We conduct SHAP for each X-fold cross-validation. For a model / and participant ipcik) = < Po + <t>i? + 0^7=1 qseverity or progression• 0Orepresents the base value for average prediction across individual i for each k fold • 0k ns the feature-specific contribution, X, telling us how much wearable feature j moved the prediction toward the disease construct (positive) or away (negative) for that specific individual I(c• We drop covariate, C, SHAP 0^ when constructing genetic traits to avoid re-encoding non-genetic factors

[0263] This yields SHAP-anchored digital phenotypes:• Univariate construct (severity or progression):Si* £ =, ’+"■< * Z ’+• Multivariate construct (feature-resolved):*r = <. C)T.. <’)T• Per-feature velocity (optional):V' = / CTTD, vector vfK).

[0264] Why SH P (explicitly). We do not GWAS raw hourly features; we GWAS SHAP contributions. Thus, every downstream association is by construction linked to disease (because SHAP measures movement toward the model’s PD construct) and localized to a specific behavior / time feature (e.g., “sleep@03:00 weekend”). This is the key conceptual advance.

[0265] Dimensionality for multivariate traits. We analyze (i) top-m features by mean I (p |(e.., m = 50), (ii) PCA of SHAP matrices keeping >80-90% variance, and (iii) correlationcluster PCs (hour x activity modules) to preserve interpretability.Genetic data processing, ancestry, and covariates

[0266] We filtered the variants and individuals in our genetic data per the following parameters: Variant QC (call rate >98%, MAF >1%, HWE P > 10-6, INFO >0.3) and sample QC (call rate >99%, sex check, relatedness). The covariates we included in all GWAS models included age, sex, BMI, smoking status, socioeconomic status (SES-England), and 10 ancestry PC (these PCs were provided by the UKB and calculated on their entire genetic cohort, which is not limited to our cohort).GWAS modelsUnivariate GWAS (binary & continuous; PLINK / LMM)

[0267] For SNP dosage G^and covariates zy.• Binary (UVB). logit Pr yt= 1) = a + gGt + yTZ( (case-control or PD stage comparisons).o y_i — Binary outcome for individual i. Set y_i = 1 for the “positive” class (e.g., case, diagnosed, or stage of interest) and y_i = 0 otherwise (e.g., control).o logit P(y_i=l) — The log-odds of the positive outcome: log(p_i / (1 - p_i)), where p i = P(y_i = 1). Inverse link (predicted probability): p i = 1 / (1 + exp( (a + P g G i + y T z_i))).o a (alpha) — Intercept; baseline log-odds of y = 1 when G_i = 0 and all covariates in z_i are 0.o G_i — Genotype for variant g for individual i. Coded as allele dosage: 0, 1, or 2 copies of the effect allele.0P_g (beta g) — Per-allele effect of the variant on the log-odds of y = 1. To obtain an odds ratio per effect allele: OR = exp(P_g). Interpretation: p_g > 0 increases odds of y = 1; P_g < 0 decreases odds of y = 1.o z_i — Vector of covariates for individual i (e.g., age, sex, ancestry PCs, batch, site, assessment center).o y (gamma) — Coefficient vector for the covariates z_i, adjusting for their effects.• Continuous (UVC). T] - a + PgGf + yTZj -I- £j,where the trait TtE {Sj, Ttitv; We report 2GC.Multivariate GWAS on SHAP vectors (MANTA / MANOVA)

[0268] Let Y ETlxmbe a SHAP -based matrix (e.g., ^top-50 features or PCs). After resi dualizing Yand Gon z:12 = n sjrSr* srG~ / 2pScy= -GTY.nWe use shrinkage / PC-space inversion for stability. Genome-wide control uses 5 x 10-8; FDR is applied for exploratory per-feature scans. For PC-space tests we back-project significant loadings to original features to pinpoint which hourxactivity components drive the signal.

[0269] Using the pipeline derived from genomebiology.biomedcentral.eom / articles / 10.1186 / sl3059-023-03076-8 (github.com / dgarrimar / mvgwas-nf), We conducted a Multivariate GWAS on various representations of the set of per-feature components of Sj, including the set of the top 50 principle components (PCs) of CV, the top 50 most model-important features ranked based on mean(abs(SHAPi)) for a given feature iGWAS WGS hit interpretation:• Hit on S[: allele dosage is associated with overall PD-likeness / severity as determined by our wearable-based models.• Hit on: allele dosage is associated with a multidimensional representation of clusters of time-resolved behaviors in the direction that influences PD severity.• Hits on TTjOr Vt: allele relates to stage-likeness / progression status and / or progression rate toward diagnosed patterns.External PRS construction and genetic validation

[0270] Using independent PD meta-GWAS weights w^, c...

Claims

CLAIMSWhat is claimed is:

1. A method of training an artificial intelligence (Al) model to generate digital phenotypes, the method comprising:providing a training set including digital descriptors; andtraining the Al model to generate the digital phenotype using the training set; wherein the training is configured to relate the digital descriptors to a health condition to generate the digital phenotype.

2. The method of claim 1, wherein the digital descriptors are generated from digital information recorded from one or more sensors.

3. The method of claim 2, wherein the one or more sensors include at least one of a biosensor, a wearable sensor, or a combination thereof.

4. The method of claim 3, wherein the wearable sensor comprises at least one of a smartwatch, a fitbit, an accelerometer, or combinations thereof.

5. The method of claim 2, wherein the digital information comprises data that reflects at least one of physical processes of the subject, physiological processes of the subject, inferred higher-order behavioral events of the subject, temporal dynamics, or combinations thereof.

6. The method of claim 5, wherein the physical or physiological processes include at least one of movement, pulse, metabolic intake, or combinations thereof.

7. The method of claim 5, wherein the higher-order behavioral events include at least one of sleep, exercise, or a combination thereof.

8. The method of claim 2, wherein the digital information further includes at least one of genotype data and demographica data.

9. The method of claim 2, wherein the digital information is recorded from control individuals and case individuals.

10. The method of claim 9, wherein:the control individuals are individuals without the health condition; andthe case individuals are individuals with the health condition.

11. The method of claim 2, further comprising generating the digital descriptors from the digital information.

12. The method of claim 11, wherein the generating of the digital descriptors comprises feature engineering the digital information.

13. The method of claim 12, wherein the feature engineering comprises at least one of:generating time-invariant static features from the digital information; and generating time-varying dynamic features from the digital information.

14. The method of claim 13, wherein generating the time-invariant static features includes applying a statistical measure over an arbitrary time window to summarize the corresponding digital information.

15. The method of claim 13, wherein generating the time-varying dynamic features includes preserving a time-varying nature of the corresponding digital information by processing the corresponding digital information using signal imputation and alignment, and then concatenating to form a multichannel time series.

16. The method of claim 12, wherein, prior to the feature engineering, the method includes at least one of preprocessing the digital information, applying a quality control process to the digital information, imputing missing values in the digital information, or combinations thereof.

17. The method of claim 16, wherein:the digital information includes at least two modalities; andthe preprocessing includes combining the at least two modalities into a single matrix.

18. The method of claim 16, wherein the QC process comprises:filtering the length of a time series to be included in downstream analyses; and identifying a time window that provides best alignment across individuals.

19. The method of claim 16, wherein the QC process comprises filtering, from a sample pool, individuals that did not provide a minimum portion of valid measurement within an identified measurement window.

20. The method of claim 16, wherein the imputing step comprises a two-fold imputation strategy.

21. The method of claim 20, wherein the two-fold imputation strategy is employed for both categorical and quantitative data.

22. The method of claim 1, wherein the training comprises a dual -architecture approach employing machine learning for static features of the digital descriptors and deep learning for dynamic features of the digital descriptors.

23. The method of claim 22, wherein the machine learning comprises XGBoost.

24. The method of claim 22, wherein the deep learning comprises Xception.

25. The method of claim 1, wherein the health condition comprises a mental health condition, a neurocognitive disease, or a neurodegenerative disease.

26. The method of claim 25, wherein the mental health condition comprises at least one of a psychiatric disorder, a mental disease, anxiety, attention deficit hyperactivity disorder (ADHD), or combinations thereof.

27. The method of claim 25, wherein the neurodegenerative disease comprises Parkinson’s disease.

28. The method of any one of claims 1-27, further comprising predictive training of the Al model to form a predictive Al model, the predictive Al model configured to predict a health status of the subject based upon the digital phenotype.

29. The method of claim 28, wherein the predictive training comprises:providing a predictive Al training set including digital phenotypes; andtraining the predictive Al model using the predictive training set.

30. The method of claim 29, wherein providing the predictive Al training set includes creating the predictive Al training set.

31. The method of claim 30, wherein creating the predictive Al training set includes generating the digital phenotype for each individual represented by the digital descriptors in the training set.

32. The method of claim 29, wherein the predictive Al model is trained for disease prediction.

33. The method of claim 29, wherein the predictive Al model is trained to predict disease progression.

34. The method of claim 29, wherein the predictive Al model is trained to predict disease progression rate.

35. A method of determining a digital phenotype for a subject, the method comprising:providing digital descriptors for the subject;inputting the digital descriptors to the Al model trained according to any one of claims 1- 27; andreceiving the digital phenotype of the subject as an output of the Al model based upon the digital descriptors.

36. The method of claim 35, further comprising generating the digital descriptors for the subject.

37. The method of claim 36, wherein the generating of the digital descriptors for the subject comprises:providing digital information for the subject; andfeature engineering the digital information to generate the digital descriptors.

38. A method of identifying genetic associations, the method comprising:applying the digital phenotype of the subject according to claim 35 as a univariate or multivariate response variable in a genome-wide association study.

9. A method of predicting a health status of a subject, the method comprising: providing digital phenotype for the subject;inputting the digital phenotype to the predictive Al model trained according to claim 28; andreceiving the predicted health status of the subject as an output of the predictive Al model based upon the digital phenotype.

Images