Methods and systems using machine learning to provide risk of a chronic disease or condition

Abstract

A device may receive physical activity data for a subject, wherein the physical activity data comprises raw acceleration data. A device may derive a plurality of metrics from the raw acceleration data, wherein the plurality of metrics comprise a plurality of pattern-related metrics. A device may input the plurality of metrics into a deployed machine learning model. A device may be configured to perform receiving, from the deployed machine learning model, a physical activity score. A device may calculate a risk score for at least one disease or condition based on the physical activity score.

Description

Docket Number: 11760-025WO1METHODS AND SYSTEMS USING MACHINE LEARNING TO PROVIDE RISK OF A CHRONIC DISEASE OR CONDITIONCROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U. S. provisional patent application No. 63 / 740,468, filed on December 31, 2024, and titled “METHODS AND SYSTEMS USING MACHINE LEARNING TO PROVIDE RISK OF A CHRONIC DISEASE OR CONDITION,” the disclosure of which is expressly incorporated herein by reference in its entirety.STATEMENT REGARDING FEDERALLY FUNDED RESEARCH

[0002] This invention was made with government support under Grant no.R01AG072445 awarded by the National Institute on Aging. The government has certain rights in the invention.BACKGROUND

[0003] Chronic diseases account for a large economic burden in the United States. Lifestyle modifications are a key element for reducing risk of a wide array of chronic diseases from heart disease to cancer, however there is currently no solution that provides continuous disease risk assessments to individuals that would motivate and reward behavioral change.SUMMARY

[0004] In some aspects, the techniques described herein relate to a method including: receiving physical activity data for a subject, wherein the physical activity data includes raw acceleration data; deriving a plurality of metrics from the raw acceleration data,Docket Number: 11760-025W01 wherein the plurality of metrics include a plurality of pattern-related metrics; inputting the plurality of metrics into a deployed machine learning model: receiving, from the deployed machine learning model, a physical activity score; and calculating a risk score for at least one disease or condition based on the physical activity score.

[0005] In some aspects, the plurality of pattern-related metrics include wavelet energy features. Optionally, the wavelet energy features include a wavelet pattern variability index (WPVI), the WPVI quantifying a variability in daily physical activity patterns.

[0006] In some aspects, the plurality of pattern-related metrics further include fractal complexity features. Optionally, the fractal complexity features are calculated using a detrended fluctuation analysis.

[0007] In some aspects, the plurality of pattern-related metrics are related to periods of activity or inactivity. In some aspects, the plurality of pattern-related metrics include a variability measure or a stability measure.

[0008] In some aspects, the plurality of metrics further include volume or intensity features.

[0009] In some aspects, the deployed machine learning model is trained on a dataset associated with a specific disease or condition.

[0010] In some aspects, the at least one disease or condition is a plurality of diseases or conditions.

[0011] In some aspects, the at least one disease or condition is a plurality of diseases or conditions, and the step of calculating the risk score includes calculating a respective risk score for each of the plurality of diseases or conditions based on the physical activity score. Alternatively, the at least one disease or condition is a plurality of diseases or conditions, the step of receiving, from the deployed machine learning model, the physical activity score includes receiving a plurality of physical activity scores, and the step ofDocket Number: 11760-025W01 calculating the risk score includes calculating a respective risk score for each of the plurality of diseases or conditions based on the plurality of physical activity scores.

[0012] In some aspects, the deployed machine learning model is trained on a dataset associated with the plurality of diseases or conditions.

[0013] In some aspects, the raw acceleration data is measured by a wearable device. Alternatively or additionally, the raw acceleration data is received from the wearable device. Optionally, the wearable device includes a tri-axial accelerometer.

[0014] In some aspects, the method further includes: receiving physiological data for the subject; and deriving one or more physiological metrics from the physiological data, wherein the plurality of metrics and the one or more physiological metrics are input into the deployed machine learning model. Optionally, the physiological data includes one or more of sleep data, heart rate data, electrocardiography data, electromyography data, blood pressure data, blood oxygen saturation data, and blood glucose data.

[0015] In some aspects, the method further includes: receiving clinical data for the subject; and deriving one or more clinical metrics from the clinical data, wherein the plurality of metrics and the one or more clinical metrics are input into the deployed machine learning model.

[0016] In some aspects, the method further includes: receiving test data for the subject; and deriving one or more test metrics from the test data, wherein the plurality of metrics and the one or more test metrics are input into the deployed machine learning model.

[0017] In some aspects, the deployed machine learning model is a supervised machine learning model. Optionally, the supervised machine learning model is a regression model, a ridge regression model, a lasso regression model, an elastic net model, or an artificial neural network.Docket Number: 11760-025W01

[0018] In some aspects, the at least one disease or condition is a cardiovascular disease or condition, a chronic disease or condition, a neurological disease or condition, a metabolic disease or condition, a mental disorder or condition, or cancer.

[0019] In some aspects, the method further includes providing a personalized recommendation for the subject based, at least in part, on the risk score for the at least one disease or condition in the subject. Optionally, the personalized recommendation reduces or improves a risk of the at least one disease or condition.

[0020] In some aspects, the techniques described herein relate to a method including: assessing the risk score for the at least one disease or condition as described herein; and treating the at least one disease or condition based on the risk score.

[0021] In some aspects, the techniques described herein relate to a computing device including: at least one processor and memory operably coupled to the at least one processor, the memory having computer-executable instructions stored thereon that, when executed by the at least one processor, cause the at least one processor to: receive physical activity data for a subject, wherein the physical activity data includes raw acceleration data; derive a plurality of metrics from the raw acceleration data, wherein the plurality of metrics include a plurality of pattern-related metrics; input the plurality of metrics into a deployed machine learning model; receive, from the deployed machine learning model, a physical activity score; and calculate a risk score for at least one disease or condition based on the physical activity score.

[0022] In some aspects, the techniques described herein relate to an electronic device including: a tri-axial accelerometer; and a computing device operably coupled to the tri-axial accelerometer, the computing device including at least one processor and memory operably coupled to the at least one processor, the memory having computer-executable instructions stored thereon that, when executed by the at least one processor, cause the atDocket Number: 11760-025W01 least one processor to: receive physical activity data for a subject, wherein the physical activity data includes raw acceleration data measured by the tri-axial accelerometer; derive a plurality of metrics from the raw acceleration data, wherein the plurality of metrics include a plurality of pattern-related metrics; input the plurality of metrics into a deployed machine learning model; receive, from the deployed machine learning model, a physical activity score; and alert the subject of a risk score for at least one disease or condition, wherein the risk score for at least one disease or condition is calculated based on the physical activity score.

[0023] In some aspects, the electronic device is a wearable device.

[0024] In some aspects, the techniques described herein relate to a method including: receiving physical activity data for a subject, wherein the physical activity data includes raw acceleration data; deriving a plurality of metrics from the raw acceleration data, wherein the plurality of metrics include a plurality of pattern-related metrics; inputting the plurality of metrics into a deployed machine learning model; receiving, from the deployed machine learning model, a physical activity score; and calculating a risk score for mortality for at least one disease or condition within one or more specified ranges of time based on the physical activity score.

[0025] In some aspects, the techniques described herein relate to a method including: receiving physical activity data for a subject, wherein the physical activity data includes raw acceleration data; deriving a plurality of metrics from the raw acceleration data, wherein the plurality of metrics include a plurality of pattern-related metrics; and calculating a risk score for at least one disease or condition based on the plurality of metrics.

[0026] In some aspects, the techniques described herein relate to a method including: receiving physical activity data for a subject, wherein the physical activity data comprises raw acceleration data; deriving a plurality of metrics from the raw acceleration data, wherein the plurality of metrics comprise a plurality of pattern-related metrics; inputtingDocket Number: 11760-025W01 the plurality of metrics into a deployed machine learning model; receiving, from the deployed machine learning model, a physical activity score: calculating a risk score for at least one disease or condition based on the physical activity score; and determining, based on the risk score, a diagnostic indicator.

[0027] In some aspects, the diagnostic indicator comprises a likelihood of a presence or absence of a disease or condition.

[0028] In some aspects, the diagnostic indicator comprises a classification of the risk score as a disease or condition.

[0029] In some aspects, the techniques described herein relate to a method including: receiving physical activity data for a subject, wherein the physical activity data comprises raw acceleration data; deriving a plurality of metrics from the raw acceleration data, wherein the plurality of metrics comprise a plurality of pattern-related metrics; inputting the plurality of metrics into a deployed machine learning model; receiving, from the deployed machine learning model, a physical activity score; and determining, using a large language model, a recommendation output based on the physical activity score.

[0030] In some aspects, the physical activity score is a composite of a plurality of scores, and wherein the recommendation output is based on at least one score of the plurality of scores.

[0031] In some aspects, the recommendation output is output to a user interface for display.

[0032] It should be understood that the above-described subject matter may also be implemented as a computer-controlled apparatus, a computer process, a computing system, or an article of manufacture, such as a computer-readable storage medium.

[0033] Other systems, methods, features and / or advantages will be or may become apparent to one with skill in the art upon examination of the following drawings and detailedDocket Number: 11760-025W01 description. It is intended that all such additional systems, methods, features and / or advantages be included within this description and be protected by the accompanying claims.BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The components in the drawings are not necessarily to scale relative to each other. Like reference numerals designate corresponding parts throughout the several views.

[0035] FIGURE 1 is a flowchart of an example method for providing risk of a chronic disease or condition according to implementations described herein.

[0036] FIGURE 2 is a block diagram of a machine learning model used for predicting physical activity scores according to implementations described herein.

[0037] FIGURE 3 is a block diagram of an example computing device.

[0038] FIG. 4A illustrates an example method of model development, according to implementations of the present disclosure.

[0039] FIG. 4B illustrates an example method of using the model developed in FIG. 4A.

[0040] FIG. 5 A illustrates hazard ratios for risk of incident chronic disease for disease-specific PA composite scores, according to a study of an example implementation of the present disclosure.

[0041] FIG. 5B illustrates hazard ratios for risk of incident chronic disease for mortality-based PA composite scores, according to a study of an example implementation of the present disclosure.

[0042] FIG. 6 A illustrates example C -indices for chronic disease prediction models for disease-specific PA composite scores, according to a study of an example implementation of the present disclosure.Docket Number: 11760-025W01

[0043] FIG. 6B illustrates C-indices for chronic disease prediction models for mortality-based PA composite scores, according to a study of an example implementation of the present disclosure.

[0044] FIG. 7 A illustrates hazard ratios for the composite score, MVPA, and step count, according to a study of an example implementation of the present disclosure. Models are adjusted for age, sex, ethnicity, education, household income, BMI, according to a study of an example implementation of the present disclosure.

[0045] FIG. 7B illustrates c-statistics for composite score, MVPA, and step counts. Models are adjusted for age, sex, ethnicity, education, household income, BMI, according to a study of an example implementation of the present disclosure.DETAILED DESCRIPTION

[0046] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art. Methods and materials similar or equivalent to those described herein can be used in the practice or testing of the present disclosure. As used in the specification, and in the appended claims, the singular forms “a,” “an,” “the” include plural referents unless the context clearly dictates otherwise. The term “comprising” and variations thereof as used herein is used synonymously with the term “including” and variations thereof and are open, non-limiting terms. The terms “optional” or “optionally” used herein mean that the subsequently described feature, event or circumstance may or may not occur, and that the description includes instances where said feature, event or circumstance occurs and instances where it does not. Ranges may be expressed herein as from "about" one particular value, and / or to "about" another particular value. When such a range is expressed, an aspect includes from the oneDocket Number: 11760-025W01 particular value and / or to the other particular value. Similarly, when values are expressed as approximations, by use of the antecedent "about," it will be understood that the particular value forms another aspect. It will be further understood that the endpoints of each of the ranges are significant both in relation to the other endpoint, and independently of the other endpoint.

[0047] As used herein, the terms "about" or "approximately" when referring to a measurable value such as an amount, a percentage, and the like, is meant to encompass variations of ±20%, ±10%, ±5%, or ±1% from the measurable value.

[0048] “Administration” of “administering” to a subject includes any route of introducing or delivering to a subject an agent. Administration can be carried out by any suitable means for delivering the agent. Administration includes self-administration and the administration by another.

[0049] The term “subject” is defined herein to include animals such as mammals, including, but not limited to, primates (e.g., humans), cows, sheep, goats, horses, dogs, cats, rabbits, rats, mice and the like. In some implementations, the subject is a human.

[0050] The term “artificial intelligence” is defined herein to include any technique that enables one or more computing devices or computing systems (i.e., a machine) to mimic human intelligence. Artificial intelligence (Al) includes, but is not limited to, knowledge bases, machine learning, representation learning, and deep learning. The term “machine learning” is defined herein to be a subset of Al that enables a machine to acquire knowledge by extracting patterns from raw data. Machine learning techniques include, but are not limited to, linear regression, logistic regression, support vector machines (SVMs), decision trees, Naive Bayes classifiers, and artificial neural networks. The term “representation learning” is defined herein to be a subset of machine learning that enables a machine to automatically discover representations needed for feature detection, prediction, orDocket Number: 11760-025W01 classification from raw data. Representation learning techniques include, but are not limited to, autoencoders. The term “deep learning” is defined herein to be a subset of machine learning that enables a machine to automatically discover representations needed for feature detection, prediction, classification, etc. using layers of processing. Deep learning techniques include, but are not limited to, artificial neural network or multilayer perceptron (MLP).

[0051] Machine learning models include supervised, semi-supervised, and unsupervised learning models. In a supervised learning model, the model learns a function that maps an input (also known as feature or features) to an output (also known as target or targets) during training with a labeled data set (or dataset).

[0052] The term “composite score” is defined herein as an aggregate measure that combines multiple individual scores or metrics into a single value. Composite scores may be created using a weighted or unweighted sum, average, or more complex combination rules or algorithms (including machine learning). Composite scores are used to summarize multiple dimensions or features.

[0053] The term “wearable device” is defined herein to include small, electronic, and often portable devices designed to be worn on the body, typically as accessories, clothing, or embedded in textiles. Wearable devices collect, process, and optionally transmit data, often in real-time, to provide users with insights, assistance, or connectivity.

[0054] Implementations of the present disclosure include improvements to systems and methods for training and using machine learning models to estimate disease risk based on physical activity. Existing systems and methods using wearable sensor data include significant technical limitations. Conventional metrics like steps and moderate-to-vigorous physical activity (MVP A) fail to include high-frequency signals, temporal variability and circadian rhythm patterns (e.g., sleep data). Attempts to supplement these conventional metrics with additional wearable data sources are limited by the high-dimensional, noisyDocket Number: 11760-025W01 and / or collinear nature of that data. These and other factors make existing wearable data sources inefficient to compute and limit the applicability of many statistical techniques. Additionally, existing machine learning models may not be able to generate accurate, evidence based, and individualized predictions for multiple chronic diseases using long- duration accelerometer data. Moreover, existing systems and methods fail to generalize to multiple diseases or outcomes without retraining. Existing systems and methods typically rely on disease-specific feature representations or outcome-specific model training, such that models trained to predict one disease or condition do not readily transfer to other diseases or outcomes. As a result, extending these approaches to additional diseases generally requires retraining separate models or re-engineering feature sets, increasing computational cost, reducing scalability, and limiting their practical deployment across multiple conditions or large populations. Moreover, simplistic “black box” machine learning models often do not produce interpretable outputs or decomposable features suitable for further processing.

[0055] Implementations of the present disclosure overcome these and other limitations of the prior art. Implementations of the present disclosure include a computer-implemented method and system that can process high-frequency wearable sensor data using novel transformations and / or survival modeling.

[0056] The example system can overcome the limitations of prior art systems and methods for training and using machine learning models by improved feature extraction. The systems and methods herein convert “raw” accelerometer data received from an accelerometer into a structured multivariate feature set that improves the training and use of the machine learning model. The feature set can optionally be constructed using wavelet energy coefficients, the wavelet pattern variability index, fractal complexity measures (e.g., DFA), intra-daily variability and inter-daily stability, intensity distributions, bout duration patterns, circadian metrics, and temporal variability signatures. The effect of the feature set isDocket Number: 11760-025W01 to extract physiologically relevant structures from noisy high-frequency time-series data that cannot be captured by conventional methods like steps or MVPA.

[0057] The example implementations described herein can further include improvements to model architecture and training. The example models trained and used herein can use a Cox proportional hazards framework combined with ridge regression to stabilize coefficients across correlated features. In some implementations a cross-validated lambda term can be included to address multicollinearity and prevent coefficient collapse. The structured training and validation methods described herein enable individualized risk scores that generalize beyond the training datasets used for the models.

[0058] The example implementations described herein can further include improvements to model outputs. The systems and methods described herein can enable composite scoring of individualized disease risks. Optionally, the composite score preserves underlying feature weights used, which can enable decomposition into specific risk contributions to the overall composite score.

[0059] Examples 1 and 2, hereto, show the measurable computational and modeling improvements of the example implementation when compared to conventional machine learning and statistics-based approaches. As shown in the Examples 1 and 2, an example implementation of the present disclosure enabled a composite score that outperforms step counts and MVPA in predictive strength and discrimination across multiple disease outcomes, as illustrated by higher C-index and stronger hazard ratios. The examples further show that the ridge-regularized Cox modeling prevents instability from multicollinearity and accommodates the full feature vector for multi-disease adaptability, and that the wavelet- and fractal-based features introduce technically derived representations unavailable in conventional activity metric.Docket Number: 11760-025W01

[0060] In sum, the present disclosure includes advances reflecting improvements in data pipeline design, feature computation, model stability, and deployment efficiency. The implementations described herein can enable a single mortality-trained model to generate predictions for other outcomes without retraining, increasing efficiency and generalizability. The outputs include interpretable and decomposable feature-level contributions, which avoid the limitations of black-box architectures and process high-dimensional wearable data without overfitting, and with suitable computational efficiency for near-real-time and / or on- device inference.

[0061] With reference to Fig. 1, a flowchart is shown of an example method using machine learning for providing risk of a chronic disease or condition. This disclosure contemplates that the operations shown in Fig. 1 can be implemented using one or more computing devices such as the computing devices shown in Fig. 3 (e.g. at least one processor and memory). In some implementations, the operations shown in Fig. 1 can be implemented on an electronic device associated with an individual such as a smartphone or a wearable device. In other implementations, the operations shown in Fig. 1 can be implemented on a computing device that is remote with respect to the individual. In yet other implementations, one or more of the operations shown in Fig. 1 can be implemented on an electronic device associated with an individual while one or more of the operations shown in Fig. 1 can be implemented on a computing device that is remote with respect to the individual. In implementations described below, a trained machine learning model is optionally used to predict a physical activity score based on a plurality of pattern-related metrics. The physical activity score is then used to calculate a risk score that quantifies a risk of chronic disease or condition.

[0062] At step 110, the method includes receiving physical activity data for a subject. The physical activity data includes raw acceleration data. Raw acceleration data canDocket Number: 11760-025W01 be measured by an accelerometer such as a tri-axial accelerometer. An example tri-axial accelerometer is the AXIVITY AX3 accelerometer from Axivity Ltd. of Newcastle upon Tyne, United Kingdom. It should be understood that the AXIVITY AX3 accelerometer is provided only as an example and that other accelerometers can be used with the methods and systems described herein. In some implementations, the raw acceleration data is measured by an electronic device that includes an accelerometer. Alternatively or additionally, the raw acceleration data is received from such electronic device. Optionally, the electronic device is a wearable device.

[0063] At step 120, the method includes deriving a plurality of metrics from the raw acceleration data. The plurality of metrics include a plurality of pattern-related metrics, which are described in more detail below. Alternatively or additionally, the plurality of metrics can optionally further include volume or intensity features. Example metrics are shown in the table below:

[0064] In some implementations, the plurality of pattern-related metrics include wavelet energy features. Optionally, the wavelet energy features include a wavelet pattern variability index (WPVI), the WPVI quantifying a variability in daily physical activity patterns.

[0065] In some implementations, the plurality of pattern -related metrics further include fractal complexity features. Optionally, the fractal complexity features are calculated using a detrended fluctuation analysis.

[0066] In some implementations, the plurality of pattern-related metrics are related to periods of activity or inactivity. In some aspects, the plurality of pattern-related metrics include a variability measure or a stability measure. For example, the variability measure can optionally be intra-daily variability (IV), and the stability measure can optionally be inter-daily stability (IS). It should be understood that IV and IS are provided only as examples.Docket Number: 11760-025W01 This disclosure contemplates that other variability and stability measures can be used with the methods and systems described herein.

[0067] At step 130, the method includes inputting the plurality of metrics into a deployed machine learning model. In some implementations, the deployed machine learning model is a supervised machine learning model. Optionally, the supervised machine learning model is a regression model, a ridge regression model, a lasso regression model, an elastic net model, or an artificial neural network. It should be understood that regression models and artificial neural networks are provided only as example supervised machine learning models. This disclosure contemplates that the deployed machine learning model can be other supervised learning models.

[0068] A regression model is a supervised learning model that uses a function to predict a target. This disclosure contemplates that the regression model can be implemented using a computing device (e.g., a processing unit and memory as described herein).Regression models can be used for classification and regression tasks. Regression models are trained with a data set to maximize or minimize an objective function, for example a measure of the regression model’s performance (e.g., error), during training. This disclosure contemplates that any algorithm that finds the maximum or minimum of the objective function can be used. Example regression models include, but are not limited to, linear and logistic regression models. Additionally, a ridge regression model is a type of linear regression model that includes L2 regularization to reduce overfitting and improve generalization to unseen data. It modifies the standard linear regression objective function by adding a penalty term proportional to the square of the magnitudes of the model's coefficients. This helps to shrink the coefficients towards zero, preventing them from becoming excessively large when the data is noisy or exhibits multicollinearity. Additionally, a lasso regression model is a type of linear regression that uses LI regularization to improveDocket Number: 11760-025W01 model performance and interpretability by penalizing the absolute values of the model coefficients. This regularization technique not only reduces overfitting but can also perform feature selection by driving some coefficients exactly to zero.

[0069] An elastic net model is a supervised learning model that combines LI (lasso) and L2 (ridge) regularization to predict a target. This disclosure contemplates that the elastic net model can be implemented using a computing device (e.g., a processing unit and memory as described herein). Elastic net models are used primarily for regression tasks and are trained with a data set to minimize an objective function, which includes both a loss term and regularization terms that control model complexity. The elastic net penalty is a linear combination of the LI and L2 penalties, making it useful for automatically reducing multicollinearity and processing sparse data. This disclosure contemplates that any algorithm that minimizes the elastic net objective function can be used.

[0070] An artificial neural network (ANN) is a computing system including a plurality of interconnected neurons (e.g., also referred to as “nodes”). This disclosure contemplates that the nodes can be implemented using a computing device (e.g., a processing unit and memory as described herein). The nodes can be arranged in a plurality of layers such as input layer, output layer, and optionally one or more hidden layers. An ANN having hidden layers can be referred to as deep neural network or multilayer perceptron (MLP). Each node is connected to one or more other nodes in the ANN. For example, each layer is made of a plurality of nodes, where each node is connected to all nodes in the previous layer. The nodes in a given layer are not interconnected with one another, i.e., the nodes in a given layer function independently of one another. As used herein, nodes in the input layer receive data from outside of the ANN, nodes in the hidden layer(s) modify the data between the input and output layers, and nodes in the output layer provide the results. Each node is configured to receive an input, implement an activation function (e.g., binary step, linear, sigmoid, tanh, orDocket Number: 11760-025W01 rectified linear unit (ReLU) function), and provide an output in accordance with the activation function. Additionally, each node is associated with a respective weight. ANNs are trained with a dataset to maximize or minimize an objective function. In some implementations, the objective function is a cost function, which is a measure of the ANN’S performance (e.g., error such as LI or L2 loss) during training, and the training algorithm tunes the node weights and / or bias to minimize the cost function. This disclosure contemplates that any algorithm that finds the maximum or minimum of the objective function can be used for training the ANN. Training algorithms for ANNs include, but are not limited to, backpropagation.

[0071] Referring now to Fig. 2, a block diagram of a system including a machine learning model 210 used for predicting physical activity scores is shown. In Fig. 2, the machine learning model 210 is operating in inference mode. The machine learning model 210 has therefore been trained with a data set (or “dataset”) and is configured to make predictions based on new input data. Accordingly, such a model is sometimes referred to herein as a “trained machine learning model” or a “deployed machine learning model.” In some implementations, the machine learning model 210 is a supervised machine learning model. As discussed above, supervised machine learning models include, but are not limited to, a regression model, a ridge regression model, a lasso regression model, an elastic net model, or an artificial neural network.

[0072] As described above, a supervised machine learning model “learns” a function that maps an input 202 (also known as feature or features) to an output (also known as target or targets) during training with a labeled data set. Optionally, in some implementations, the deployed machine learning model is trained on a dataset associated with a specific disease or condition. In some aspects, the deployed machine learning model is trained on a dataset associated with a plurality of diseases or conditions. In someDocket Number: 11760-025W01 implementations, a trained supervised machine learning model is configured to classify the input 202 into one of a plurality of target categories (i.e., the output). In other words, the trained model can be deployed as a classifier. In other implementations, a trained supervised machine learning model is configured to provide a probability of a target (i.e., the output) based on the input 202. In other words, the trained model can be deployed to perform a regression.

[0073] As shown in Fig. 2, the machine learning model 210 is configured to provide output of a physical activity score based on the input 202. In the examples described herein, the input 202 includes a plurality of metrics derived from raw acceleration data, and the output is a physical activity score 204. The machine learning model 210 is therefore trained to map the input 202 to the output. In other words, the input 202 includes one or more “features” that are input into the machine learning model 210, which predicts the output which is therefore the “target” of the machine learning model 210. For example, the physical activity score can integrate the plurality of metrics weighted by respective coefficients obtained from the machine learning model as described in example 1, and example 2 herein. In other words, the output can optionally be a vector that can include scores for each feature, allowing the output to be interpretable by evaluating each feature. In FIG. 2, the machine learning model 210 includes a Cox proportional hazards model 212, but it should be understood that this is only a non-limiting example.

[0074] The input 202 can be a pre-processed input. As described throughout the present disclosure, implementations of the present disclosure include dimensionality reduction techniques to pre-process the input for training to reduce dimensionality and improve computational performance while maintaining necessary information to generate reliable predictions. For example, the input 202 can be pre-processed by epoch summarization, wavelet transforms into wavelet energy coefficients, WPVI (Wavelet PatternDocket Number: 11760-025W01 Variability Index, described in example 1) and IV / IS (intra-daily variability and stability, respectively, described in Example 1).

[0075] Still with reference to Fig. 2, the system can optionally include a three-axis accelerometer 220. The three-axis accelerometer 220 can be the source of accelerometer data used to derive the metrics derived from raw acceleration data input 202. Alternatively or additionally, the system can further include wearable, clinical and questionnaire measures 222. Optionally, the measures 222 can be stored in a database (e.g., on the remote computing device). Alternatively or additionally, the measures 222 can be collected from sensors of the system shown in FIG. 2. FIG. 4A illustrates additional examples of sensors that can be used in implementations of the present disclosure. Optionally, the three-axis accelerometer 220 can be replaced with a single axis accelerometer. The accelerometer can further optionally be part of an inertial measurement unit (IMU) that includes a gyroscope and / or magnetometer to measure angular velocity, orientation, and / or posture-related motion. In some implementations, the system can further include sensors that acquire other physiological, biomechanical, and / or contextual data including but not limited to heart rate sensors, ECG sensors, pressure or force sensors, skin temperature sensors, GPS sensors, ambient light sensors, microphones, and / or barometric sensors.

[0076] The system shown in Fig. 2 can further include a display 230 and / or a remote computing device 240 in some implementations. The display 230 and / or remote computing device 240 can include any or all of the features of the computing device 300 shown and described with reference to Fig. 3. For example, the physical activity score 204 can be output for display to a user by the display 230. Alternatively or additionally, the physical activity score 204 can be transmitted (e.g., by wired or wireless network) to the remote computing device for storage or additional analysis (e.g., as a cloud-based system).Docket Number: 11760-025W01

[0077] It should be understood that the system of Fig. 2 can be implemented using different numbers and arrangements of devices. For example, in some implementations the three-axis accelerometer 220, display 230, and machine learning model 210 can be implemented using the same computing device (e.g., as part of a single wearable device). Alternatively or additionally, it should be understood that in some implementations, the machine learning model 210 can be located on a remote computing device (the same or different from remote computing device 240) and the system can be configured so that data from the three-axis accelerometer 220 is transmitted to the remote computing device for processing by the machine learning model 210, and then the physical activity score 204 is transmitted back to the user device for display by the display 230.

[0078] The present disclosure further contemplates that the display 230 and / or remote computing device 240 can implement dashboards and / or telehealth workflows, for example by storing the physical activity score over time to display as a dashboard or by transmitting the physical activity score to one or more users (patients, physicians, etc.). Optionally, the method can include standardizing the physical activity score 204 for use in a health record system maintained in the remote computing device (e.g., electronic health records). Alternatively or additionally, the physical activity score 204 can be used to set an alert threshold, create a unified index of another health parameter, or configure another visualization of patient health. As yet another example, the physical activity score 204 can be used to prioritize users by relative risk, and / or as an input to another behavior recommendation engine.

[0079] As a non-limiting example, in some implementations of the present disclosure, a recommendation engine 250 can be implemented using one or more large language models 252. The recommendation engine 250 can optionally be implemented by the same computing device as the machine learning model 210 and / or by the remote computingDocket Number: 11760-025W01 device 240. The recommendation engine can optionally be configured to receive a physical activity score and determine, based on a trained large language model 252, a recommendation output. As described throughout the present disclosure, the physical activity score can be a composite score, and the recommendation engine can output the recommendation output by one or more components of the composite score. For example, the large language model 252 can use components of the physical activity score to provide specific and interpretable recommendations when a large language model is used as the recommendation engine. As shown in FIG. 2, the recommendation engine 250 can output the recommendation for display by a display 230.

[0080] Example Recommendation Engines

[0081] In an example implementation, the recommendation engine 250 can include a large language model 252 trained on a dataset of curated content including one or more of the following: Behavioral science (habit formation, intrinsic motivation, identity-based messaging); Psychology (self-efficacy, attention, reinforcement, affect regulation); Physical activity science (dose-response effects, intensity, metabolic pathways); Chronic disease prevention; and / or Neuroscience of stress, fatigue, and reward.

[0082] The use of a curated dataset enables the large language model 252 to produce recommendations that are not only scientifically credible but also emotionally intelligent and engaging. The use of a curated dataset further enables the output recommendations to be safe, consistent, and evidence-aligned by preventing the recommendation engine from outputting medical advice or unverified information, and / or by ensuring that scientific explanations of any recommendations are based on pre-approved text.

[0083] An example of a recommendation that can be output by implementations of the present disclosure includes a personalized micro-intervention that helps users reduce chronic disease risk through small, achievable actions. The use of a curated dataset and largeDocket Number: 11760-025W01 language model 252 overcomes the limitations of conventional systems that use generic pre¬ programmed reminders. The large language models described herein can further be behaviorally intelligent to generate context-aware recommendations that users respond to.

[0084] In an example implementation, the recommendation engine can receive data collected by one or more wearable devices of a user and input it into a machine learning model according to implementations of the present disclosure (e.g., the machine learning model 210 shown in FIG. 2). The physical activity score output by the machine learning model can optionally include any or all of the following: (1) the disease category where the user needs the most support or the category selected by the user as highest priority; (2) whether their risk trend is improving or worsening; (3) what type of short, achievable movement would have the biggest payoff based on accelerometer features driving the score; and / or (4) what the user has responded to in the past (and what they have ignored). In some implementations, the recommendation engine can further take into account contextual information about the user’s local environment, such as weather conditions, time of day, season, or available surroundings (e.g., parks, walking paths, gyms, etc.), to tailor recommendations in a way that is both feasible and relevant to the user at the time and location they are delivered.

[0085] The large language models described herein can be configured to convert the physical activity scores described herein into a personalized message. For example, instead of saying only “move more,” the outputs described herein can weave together behavioral science, motivational framing, and user data. The large language models 252 described herein can be configured to learn from the user’s behavior and improve the output recommendations to generate real, lasting behavioral change.

[0086] Non-limiting examples from the LLMs described herein can include:

[0087] Reinforcing habits with concrete, low-friction actions:Docket Number: 11760-025W01

[0088] E.g., “Consistency beats intensity: take a 5-minute brisk walk right after this notification to keep your momentum going.”

[0089] Building identity through action-linked self-perception:

[0090] “You’ve been showing up for your health. A fast lap around the block this afternoon reinforces that identity.”

[0091] Highlighting immediate psychological benefits:

[0092] “Feeling mentally cluttered? A quick 2-minute movement break can calm your stress response and reset your focus.”

[0093] Connecting diet and movement in a safe, behaviorally meaningful way:

[0094] “Just finished eating? A 10-minute post-meal walk helps your body use that energy more smoothly.”

[0095] Use creative cues to make activity enjoyable and easier to start:

[0096] “Queue up your favorite podcast and take a short walk during the next break in your schedule.”

[0097] Offering surprising, cross-domain links that increase motivation:

[0098] “Choosing simple whole foods today can keep your energy steadier, perfect for fitting in a quick walk this evening.”

[0099] Reducing sedentary time with micro-break prompts:

[0100] “Been sitting for a while? Stand up and take a 1 -minute lap around the room. It helps your circulation and keeps your energy up for the rest of the day.”

[0101] The outputs can stay “slightly far afield” in ways that surprise users, but can return to reinforce the behavior of movement that meaningfully reduces disease risk.

[0102] Referring again to Fig. 1, at step 140, the method includes receiving, from the deployed machine learning model, a physical activity score. The physical activity score is a composite score (i.e. an aggregate measure) based on a plurality of metrics derivedDocket Number: 11760-025W01 from raw acceleration data as described above. An example machine learning model trained to predict physical activity scores (i.e. a Cox proportional hazards model) is described in Example 1. It should be understood that the Cox proportional hazards model described in Example 1 is provided only as an example. As discussed above, this disclosure contemplates that the deployed machine learning model can be other supervised learning models.

[0103] At step 150, the method includes calculating a risk score for at least one disease or condition based on the physical activity score. The risk score can be used to quantify an individual’s relative risk based on their unique activity patterns. For example, the physical activity score can be derived from a vector (e.g. accelerometer metrics weighted by respective coefficients of the deployed machine learning model) as described in Example 1. It should be understood that the derivation described in Example 1 is provided only as an example. This disclosure contemplates using other methods to calculate the risk score based on the physical activity score.

[0104] The at least one disease or condition can be a cardiovascular disease or condition, a chronic disease or condition, a neurological disease or condition, a metabolic disease or condition, a mental disorder or condition, or cancer. Cardiovascular diseases or conditions include, but are not limited to, heart disease, stroke, and hypertension.Neurological diseases or conditions include, but are not limited to, dementia, Alzheimer’s disease, and Parkinson’s disease. Metabolic diseases or conditions include, but are not limited to, diabetes. It should be understood that the diseases and conditions above are provided only as examples. This disclosure contemplates that risk scores for other diseases and conditions can be calculated with the methods and systems described herein.

[0105] In some implementations, the deployed machine learning model returns a single physical activity score at step 140, and then a single risk score for a specific disease or condition is calculated at step 150.Docket Number: 11760-025W01

[0106] Alternatively, the at least one disease or condition is a plurality of diseases or conditions. In some implementations, the deployed machine learning model returns a single physical activity score at step 140, and then a single risk score for a plurality of diseases or conditions is calculated at step 150. In other implementations, the step of calculating the risk score includes calculating a respective risk score for each of the plurality of diseases or conditions based on the physical activity score. In these implementations, the deployed machine learning model returns a single physical activity score at step 140, and then multiple risk scores for a plurality of diseases or conditions are calculated at step 150. In yet other implementations, the step of receiving, from the deployed machine learning model, the physical activity score includes receiving a plurality of physical activity scores, and the step of calculating the risk score includes calculating a respective risk score for each of the plurality of diseases or conditions based on the plurality of physical activity scores. In these implementations, the deployed machine learning model returns multiple physical activity¬ scores at step 140, and then multiple risk scores for a plurality of diseases or conditions are calculated at step 150.

[0107] In some implementations, the method of FIG. 1 can be used to automatically output a diagnostic indicator based on the physical activity score and / or risk score. The risk score can represent the likelihood or potential for future disease development. The diagnostic indicator can include a classification of the risk score as indicating the presence or absence of a disease or condition in a subject. Alternatively or additionally, the diagnostic indicator can include a diagnosis of a disease or condition.

[0108] Optionally, in some implementations, the method further includes providing a personalized recommendation for the subject based, at least in part, on the risk score for the at least one disease or condition in the subject. Optionally, the personalizedDocket Number: 11760-025W01 recommendation reduces or improves a risk of the at least one disease or condition. This disclosure contemplates that the personalized recommendation can be provided through an electronic communication channel such as a display device, email, SMS, mobile app, or website.

[0109] Optionally, in some implementations, the techniques described herein relate to a method including: assessing the risk score for the at least one disease or condition as described herein; and treating the at least one disease or condition based on the risk score.

[0110] As described above, the physical activity score output by the deployed machine learning model is a composite score based on a plurality of metrics derived from raw acceleration data. Optionally, the physical activity score output by the deployed machine learning model is a composite score based on a plurality of metrics derived from raw acceleration data and one or more metrics derived from physiological, clinical, and / or test data. In other words, the features input into the deployed machine learning model can optionally include features derived from raw acceleration data as well as features derived from other types of data. For example, in some implementations, the method optionally further includes: receiving physiological data for the subject; and deriving one or more physiological metrics from the physiological data, where the plurality of metrics and the one or more physiological metrics are input into the deployed machine learning model. As described above, the deployed machine learning model predicts the physical activity score. In other words, the features input into the deployed machine learning model include pattern- related metrics and physiological metrics in these implementations. Optionally, the physiological data includes one or more of sleep data, heart rate data, electrocardiography¬ data, electromyography data, blood pressure data, blood oxygen saturation data, and blood glucose data. It should be understood that the physiological data above are provided only asDocket Number: 11760-025W01 examples. This disclosure contemplates that other physiological data can be used with the methods and systems described herein.

[0111] Alternatively or additionally, in some implementations, the method optionally further includes: receiving clinical data for the subject; and deriving one or more clinical metrics from the clinical data, where the plurality of metrics and the one or more clinical metrics are input into the deployed machine learning model. As described above, the deployed machine learning model predicts the physical activity score. In other words, the features input into the deployed machine learning model include pattern-related metrics and clinical metrics in these implementations (and optionally physiological metrics).Additionally, this disclosure contemplates that clinical data can be obtained from electronic medical records, subject reports of medical history, or other sources of clinically-related data or personal characteristics (e.g., demographics, family history, etc.).

[0112] Alternatively or additionally, in some implementations, the method optionally further includes: receiving test data for the subject; and deriving one or more test metrics from the test data, wherein the plurality of metrics and the one or more test metrics are input into the deployed machine learning model. This disclosure contemplates that test data can include results of subjective questions (e.g. surveys), objective tests of cognitive and behavioral performance, or physical tests. Optionally, the test data can be acquired over time and linked or synchronized with the raw accelerometry data to provide information on activities performed by the individual during specified periods of accelerometer-measured physical activity. As described above, the deployed machine learning model predicts the physical activity score. In other words, the features input into the deployed machine learning model include pattern-related metrics and test metrics in these implementations (and optionally physiological and / or clinical metrics).Docket Number: 11760-025W01

[0113] In some implementations, the techniques described herein relate to a computing device including: at least one processor and memory operably coupled to the at least one processor, the memory having computer-executable instructions stored thereon that, when executed by the at least one processor, cause the at least one processor to: receive physical activity data for a subject, wherein the physical activity data includes raw acceleration data; derive a plurality of metrics from the raw acceleration data, wherein the plurality of metrics include a plurality of pattern-related metrics; input the plurality of metrics into a deployed machine learning model; receive, from the deployed machine learning model, a physical activity score; and calculate a risk score for at least one disease or condition based on the physical activity score.

[0114] In some implementations, the techniques described herein relate to an electronic device including: a tri-axial accelerometer; and a computing device operably coupled to the tri-axial accelerometer, the computing device including at least one processor and memory operably coupled to the at least one processor, the memory having computer¬ executable instructions stored thereon that, when executed by the at least one processor, cause the at least one processor to: receive physical activity data for a subject, wherein the physical activity data includes raw acceleration data measured by the tri-axial accelerometer; derive a plurality of metrics from the raw acceleration data, wherein the plurality of metrics include a plurality of pattern-related metrics; input the plurality of metrics into a deployed machine learning model; receive, from the deployed machine learning model, a physical activity score; and alert the subject of a risk score for at least one disease or condition, wherein the risk score for at least one disease or condition is calculated based on the physical activity score. In some aspects, the electronic device is a personal computing device such as a smartphone. In some aspects, the electronic device is a wearable device. Example wearable devices include, but are not limited to, a device that is worn on the body or clothing, a patch-Docket Number: 11760-025W01 like device that adheres to the body or clothing, and jewelry (e.g., a ring). This disclosure contemplates that the subject can be alerted through an electronic communication channel such as a display device, email, SMS, mobile app, or website; sound; vibration; or combinations thereof.

[0115] In some implementations, the techniques described herein relate to a method including: receiving physical activity data for a subject, wherein the physical activity data includes raw acceleration data; deriving a plurality of metrics from the raw acceleration data, wherein the plurality of metrics include a plurality of pattern-related metrics; inputting the plurality of metrics into a deployed machine learning model; receiving, from the deployed machine learning model, a physical activity score; and calculating a risk score for mortality for at least one disease or condition within one or more specified ranges of time based on the physical activity score.

[0116] In some implementations, the techniques described herein relate to a method including: receiving physical activity data for a subject, wherein the physical activity data includes raw acceleration data; deriving a plurality of metrics from the raw acceleration data, wherein the plurality of metrics include a plurality of pattern-related metrics; and calculating a risk score for at least one disease or condition based on the plurality of metrics. Optionally, the step of calculating a risk score can include inputting the plurality of metrics into a deployed machine learning model.

[0117] It should be appreciated that the logical operations described herein with respect to the various figures may be implemented (1) as a sequence of computer implemented acts or program modules (i.e., software) running on a computing device (e.g., the computing device described in Fig. 3), (2) as interconnected machine logic circuits or circuit modules (i.e., hardware) within the computing device and / or (3) a combination of software and hardware of the computing device. Thus, the logical operations discussed hereinDocket Number: 11760-025W01 are not limited to any specific combination of hardware and software. The implementation is a matter of choice dependent on the performance and other requirements of the computing device. Accordingly, the logical operations described herein are referred to variously as operations, structural devices, acts, or modules. These operations, structural devices, acts and modules may be implemented in software, in firmware, in special purpose digital logic, and any combination thereof. It should also be appreciated that more or fewer operations may be performed than shown in the figures and described herein. These operations may also be performed in a different order than those described herein.

[0118] Referring to Fig. 3, an example computing device 300 upon which the methods described herein may be implemented is illustrated. It should be understood that the example computing device 300 is only one example of a suitable computing environment upon which the methods described herein may be implemented. Optionally, the computing device 300 can be a well-known computing system including, but not limited to, personal computers, servers, handheld or laptop devices, multiprocessor systems, microprocessorbased systems, network personal computers (PCs), minicomputers, mainframe computers, embedded systems, and / or distributed computing environments including a plurality of any of the above systems or devices. Distributed computing environments enable remote computing devices, which are connected to a communication network or other data transmission medium, to perform various tasks. In the distributed computing environment, the program modules, applications, and other data may be stored on local and / or remote computer storage media.

[0119] In its most basic configuration, computing device 300 typically includes at least one processing unit 306 and system memory 304. Depending on the exact configuration and type of computing device, system memory 304 may be volatile (such as random access memory (RAM)), non-volatile (such as read-only memory (ROM), flashDocket Number: 11760-025W01 memory, etc.), or some combination of the two. This most basic configuration is illustrated in Fig. 3 by box 302. The processing unit 306 may be a standard programmable processor that performs arithmetic and logic operations necessary for operation of the computing device 300. The computing device 300 may also include a bus or other communication mechanism for communicating information among various components of the computing device 300.

[0120] Computing device 300 may have additional features / functionality. For example, computing device 300 may include additional storage such as removable storage 308 and non-removable storage 310 including, but not limited to, magnetic or optical disks or tapes. Computing device 300 may also contain network connection(s) 316 that allow the device to communicate with other devices. Computing device 300 may also have input device(s) 314 such as a keyboard, mouse, touch screen, etc. Output device(s) 312 such as a display, speakers, printer, etc. may also be included. The additional devices may be connected to the bus in order to facilitate communication of data among the components of the computing device 300. All these devices are well known in the art and need not be discussed at length here.

[0121] The processing unit 306 may be configured to execute program code encoded in tangible, computer-readable media. Tangible, computer-readable media refers to any media that is capable of providing data that causes the computing device 300 (i.e., a machine) to operate in a particular fashion. Various computer-readable media may be utilized to provide instructions to the processing unit 306 for execution. Example tangible, computer- readable media may include, but is not limited to, volatile media, non-volatile media, removable media and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. System memory 304, removable storage 308, and non-removable storage 310 are all examples of tangible, computer storage media. Example tangible,Docket Number: 11760-025W01 computer-readable recording media include, but are not limited to, an integrated circuit (e.g., field-programmable gate array or application-specific IC), a hard disk, an optical disk, a magneto-optical disk, a floppy disk, a magnetic tape, a holographic storage medium, a solid- state device, RAM, ROM, electrically erasable program read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices.

[0122] In an example implementation, the processing unit 306 may execute program code stored in the system memory 304. For example, the bus may carry data to the system memory 304, from which the processing unit 306 receives and executes instructions. The data received by the system memory 304 may optionally be stored on the removable storage 308 or the non-removable storage 310 before or after execution by the processing unit 306.

[0123] It should be understood that the various techniques described herein may be implemented in connection with hardware or software or, where appropriate, with a combination thereof. Thus, the methods and apparatuses of the presently disclosed subject matter, or certain aspects or portions thereof, may take the form of program code (i.e., instructions) embodied in tangible media, such as floppy diskettes, CD-ROMs, hard drives, or any other machine-readable storage medium wherein, when the program code is loaded into and executed by a machine, such as a computing device, the machine becomes an apparatus for practicing the presently disclosed subject matter. In the case of program code execution on programmable computers, the computing device generally includes a processor, a storage medium readable by the processor (including volatile and non-volatile memory and / or storage elements), at least one input device, and at least one output device. One or more programs may implement or utilize the processes described in connection with the presentlyDocket Number: 11760-025W01 disclosed subject matter, e.g., through the use of an application programming interface (API), reusable controls, or the like. Such programs may be implemented in a high level procedural or object-oriented programming language to communicate with a computer system. However, the program(s) can be implemented in assembly or machine language, if desired. In any case, the language may be a compiled or interpreted language and it may be combined with hardware implementations.

[0124] Examples

[0125] The following examples are put forth so as to provide those of ordinary skill in the art with a complete disclosure and description of how the compounds, compositions, articles, devices and / or methods claimed herein are made and evaluated, and are intended to be purely exemplary and are not intended to limit the disclosure. Efforts have been made to ensure accuracy with respect to numbers (e.g., amounts, temperature, etc.), but some errors and deviations should be accounted for.

[0126] Example 1:

[0127] A study was performed of an example implementation of the present disclosure. The study includes an example implementation of the present disclosure including a disease risk algorithm that harnesses the underlying richness of wearables data to provide users with key motivation for lifestyle modifications. The study includes data from wearable accelerometers as a case study, however this method is flexible and can include a wide array of data from wearables, as well as clinical data and questionnaire-based data as shown in FIGS. 4A and 4B.

[0128] METHODS

[0129] Study Design and Participants

[0130] The study used data from the UK Biobank (community dwelling participants in England, Scotland, and Wales) with baseline data collected between 2006 andDocket Number: 11760-025W012010. In a sub-study from 2013 to 2015, 103,684 participants agreed to wear an Axivity AX3 tri-axial accelerometer 24 hours per day for 7 days on their dominant wrist. The study restricted the analysis to indi viduals who were free of the di sease of interest prior to the accelerometer sub-study and who had at least 3 valid days (more than 16 hours per day) of wear-time. Participants were followed from accelerometer wear date until their first disease of interest diagnosis, death, loss to follow-up, or to the last date of hospital admission from the respecti ve database

[0131] Physical Activity (PA) and Sedentary Behavior (SB ) Features

[0132] PA and SB features were derived from raw accelerometer data using R statistical computing software. Accelerometer data was first processed using the GGIR package. Using this package, the Euclidean Norm Minus One (ENMO) was calculated with negative values rounded to zero. A group of features was extracted from the GGIR output and used in model selection (see Supplementary Table 1 below for an example list of features). Specific features are described below but we divide the features into either physical activity volume / intensity metrics or metrics that describe paterns of activity and inactivity across the day. Unless noted below, metrics described were derived from GGIR output.

[0133] Supplementary Table 1: Example Variables included in models according to implementations of the present disclosure.Metric Description.. ACC..day„mg..wei Average daily acceleration _dur_day_total_IN_inin_wei average daily time spent in inactivity _dur_day_total_LIG_min_wei average daily time spent in light PA „totalMVPAdur average daily time spent in MVP Adaily inactivity accumulated in bouts of 10-30 _dur_day_IN_bts_ 10_30_min_wei minutesdaily inactivity accumulated in bouts greater _dur_day_IN_bts_30_roin_wei than 30 minutesdaily light PA accumulated in bouts of 10 or _dur_day_LIG_bts_ 10_min_wei more minutesdaily MVP A accumulated in bouts of 10 or_d ur_day_M VP A_bts_ 10_min_ wei more minutesDocket Number: 11760-025W01._ig_day._gradient_ wei intensity gradient slope_ig_day_intercept_wei intensity gradient intercept_dfaResult dfa alpha (measure of fractal complexity)_I VJntradail y variabili ty Intradaily VariabilityJLS jnterdailystability Intradaily Stability_StepsDayAvg Mean steps per day_StepsDayMed Median steps per day_StepsDayMin Minimum step count across wear time.. StepsDayMax Maximum step count across wear time _CadencePeakl.steps.min. highest 1 minute cadence _CadencePeak30.steps.min. highest 30 minute cadencetime stamp for lowest 5hr activity level as _AD_L5hr_ENMO_mg_0.24hr measured by ENMOAverage of ENMO during the least active five hours in the day that is the lowest rolling _AD_. L5_ENMO_mg_0.24hr average value of ENMO.time stamp for highest 5hr activity level as _AD_M5hr_ENMO_mg_0.24hr measured by ENMOAverage of ENMO during the most active five hours in the day that is the highest rolling._AD_M5._ENMO.mg_0.24hr average value of ENMO.Average number of bouts / day of MVP A greater _Nbouts_day_MVPA_bts_l 0_wei than 10 minutesAverage number of bouts / day of M VP A_Nbou ts_day_M VP A_bts_5_ 10_wei between 5 and 10 minutesAverage number of bouts / day of MVPA _Nbouts_day_MVPA_bts_l_5_wei between 1 and 5 minutesAverage number of inactivity bouts greater than _Nbouts_day_IN_bts_30_wei 30 minutesAverage number of inactivity' bouts between 10 _Nbouts_day_IN_bts_10_30_wei and 30 minutesAverage number of inactivity bouts between 1 _Nbouts_day_IN_bts_l_10_wei and 10 minutesAverage number of light PA bouts greater than _Nbouts_day_LIG_bts_ 10_wei 10 minutesAverage number of light PA bouts between 5 _Nbouts_day_LIG_bts_5_10_wei and 10 minutesAverage number of light PA bouts between 1 _Nbouts_day_LIG_bts_ l_5_wei and 5 minutesMean wavelet energy at the first decomposition _mean_wavelet_energy_W 1 level CW1) calculated across all days.Mean wavelet energy at the first decomposition_mean._wavelet._energy.. W2 level (W2) calculated across all days.Docket Number: 11760-025W01Mean wavelet energy at the first decomposition _mean_ wav elet_energy_W 3 level (W3) calculated across all days.Mean wavelet energy at the first decomposition _mean_wavelet_energy_W4 level (W4) calculated across all days.Wavelet Pattern Variability Index, whichmeasures the variability in the wavelet energy patterns across different days. It is the mean Euclidean distance of daily wavelet energy_wpvi distributions from their centroid.

[0134] Supplementary Table 2: Sleep metrics used in alternative models metric definitiontotal sleep duration Sleep within the Steen period time window.onset of sleep expressed in hours since the midnight in the night preceding the night ofsleep onset interest, e.g. 26 is 2am.percentage of sleep within the sleep period timesleep efficiency window

[0135] Physical activity and inactivity volumes and intensities

[0136] The study included overall average waking time acceleration magnitude to capture average PA. To assess intensity, the study divided PA into three categories: Inactivity (IN), Light Physical Activity (LPA), and Moderate-to-Vigorous Physical activity (MVPA). These categories were determined by accelerometer thresholds. One example of thresholds that can be used in implementations of the present disclosure is: (IN < 45mg; LPA >=45 mg and < lOOmg; MVPA >= lOOmg ). It should be understood that these values are only non-limiting examples used for the study, and that different thresholds and / or numbers of thresholds are possible in various implementations of the present disclosure. For each category, the example implementation included total daily duration in that intensity, as well as time spent in bouts, or uninterrupted stretches of time. For IN, the example implementation used bout lengths of 10-30 minutes as well as bouts ofDocket Number: 11760-025W01 over 30 minutes to capture sustained periods of inactivity. For LPA and MVPA, the example implementation used bouts of 10 or more minutes, as 10 minute bouts were defined as health enhancing PA. For each of these bout lengths, the example implementation also used the number of bouts within these time windows per day. The study assessed how intensities are distributed throughout the day using metrics that account for the least active and most active 5 -hour periods. The example implementation included both the average accelerometer magnitudes during these time periods as well as the time stamp for the start of these periods in the example model.

[0137] To capture the overall distribution of activity intensities across the day, the example implementation used an intensity gradient. The intensity gradient method involves calculating the slope and intercept of the log-log regression line between the time spent at various intensity levels and the corresponding intensity values. The intensity gradient, represented by the slope, indicates the rate of decline in time spent as activity intensity increases, while the intercept provides a baseline measure of activity intensity. This approach enables a comprehensive assessment of PA, capturing both the distribution and magnitude of activity intensities, which may be critical for understanding the relationship between physical activity and health outcomes.

[0138] In addition to intensity categories, the example implementation used step counts to characterize PA. The example implementation studied included average daily step counts, median daily step counts, minimum and maximum daily step counts, and cadence (1 - and 30 -minute peak steps / min).

[0139] Paterns of Activity

[0140] The example implementation used a variety of variables that capture different aspects of PA patterns. This study calculated Intra-daily Variability (IV) and Inter-daily Stability (IS) from accelerometer data to assess the regularity and fragmentation ofDocket Number: 11760-025W01 participants’ physical activity patterns. Intra-daily Variability (IV) quantifies the fragmentation of physical activity by measuring the frequency and extent of transitions between activity and inactivity within a 24 -hour period. This was computed using the standard deviation of the differences between consecutive hourly activity counts, normalized by the total activity count. Inter-daily Stability (IS), on the other hand, assesses the consistency of daily activity patterns over the monitoring period. It is calculated as the ratio of the variance between the daily mean activity counts and the total variance of hourly activity counts. Higher IS values indicate more stable and consistent daily activity patterns. These metrics provide insights into the rhythmicity and predictability of physical activity behaviors, which are crucial for understanding their impact on health outcomes.

[0141] The study used Detrended Fluctuation Analysis (DFA) to calculate the fractal complexity of the accelerometer signal using data summarized into 5 second epochs. DFA was conducted for each participant using the nonlinear Tseries package in R version 3.3.1. This process begins with the integration of time-series data after the global mean of the signal has been removed. The integrated time-series is then divided into nonoverlapping segments or windows of varying size, t. In each window, the root mean square deviation from a least-squares fitted line is calculated to detrend the data and to determine the fluctuation amplitude, F(t). This procedure is repeated for different window sizes, ranging from 10 minutes to 7 hours, to assess a wide range of fluctuations. The fractal complexity of the time¬ series is then quantified by the slope, a, of the least-squares regression line that relates F(t) to window size t. A slope ( a ) of 0.5 indicates no long-range correlations in fluctuations (akin to white noise), while slopes between 0.5 and 1.0 suggest the presence of positive long- range correlations and increasing fractal complexity as a approaches 1.0. Slopes greater than 1.0 imply more regular or predictable fluctuations, increasing towards 1.5. In the context ofDocket Number: 11760-025W01 motor activity data, a values close to 1.0 can be indicative of healthy fractal physiological complexity.

[0142] The example implementation further applied wavelet transformation techniques to extract significant features related to physical activity patterns using the wavelets package in R. The wavelet energy coefficients were calculated for each data file using the discrete wavelet transform (DWT) with the "la8" filter at four decomposition levels. This approach allows the example implementation to capture both high-frequency and low-frequency components of the accelerometer signals. For each day of data, wavelet energy coefficients were computed and used to calculate mean wavelet energy. These features provide a comprehensive representation of the variability and complexity of the physical activity patterns.

[0143] The example implementation studied used a technique referred to herein as “Wavelet Pattern Variability Index (WPVI)” which includes determining the Euclidean distance of each day's wavelet energy distribution from the centroid of all daily distributions and taking a mean of these values. This metric quantifies the variability in daily physical activity patterns, with higher values indicating greater variability. This technique improves over previous techniques for quantifying the variability in daily physical patterns. In addition to quantifying variability in daily physical activity patterns, the WPVI, when used as one component of a multivariate feature set in a machine learning model, can provide several technical advantages over existing variability metrics. WPVI is derived from daily multi-scale wavelet-energy profiles and quantifies day-to-day deviations in temporal organization relative to a subject-specific centroid. WPVI contributes to dimensionality reduction and reduces redundancy with conventional volume-based metrics by capturing aspects of temporal structure and variability that are not represented by traditional physical activity measures. Inclusion of WPVI within the feature set improves numerical stability andDocket Number: 11760-025W01 generalizability of downstream penalized regression and survival modeling frameworks, while providing an interpretable, geometrically defined measure of pattern variability suitable for feature -level decomposition and automated downstream processing.

[0144] Sleep

[0145] The study calculated nighttime sleep metrics using GGIR. The study included total sleep duration, sleep efficiency, and time of sleep onset. These variables are included in a separate model with PA and SB metrics described above to determine whether inclusion of sleep data significantly improves the models.

[0146] Disease and Health Condition Outcomes

[0147] Hospital inpatient records and death registries were used to determine health conditions and diseases for the following diagnoses (e.g. heart disease, cancer, diabetes, osteoarthritis, chronic obstructive pulmonary disease (COPD), stroke, hypertension, chronic kidney disease, depression, dementia, and Parkinson's disease). The International Classification of Diseases Tenth Revision (ICD-10) codes were used to classify participants with disease outcomes.

[0148] Feature Selection and Machine Learning Based Model Development

[0149] In the study, the features were first scaled to have a mean of 0 and standard deviation of 1. The dataset was evenly divided into a training set and a testing set. This division was intended to maintain both the robustness and efficacy of the model validation process, especially for diseases / conditions where there are smaller numbers of cases. A Cox proportional hazards model, enhanced with Ridge regression regularization, was employed to predict the incidence of disease using variables related to PA and demographic data. Ridge regression improves model accuracy and interpretability by penalizing the square of the magnitude of the model coefficients. This penalization approachDocket Number: 11760-025W01 reduces the complexity of the model and helps prevent overfitting by shrinking the coefficients, thereby ensuring that only the most significant variables exert a strong influence.

[0150] Model parameters were determined through a 10 -fold stratified cross-validation. This method involves dividing the training data into ten equal parts, ensuring each part is representative of the overall distribution of disease / condition incidence. Participants were randomly assigned to these folds while maintaining stratification (the proportion of disease / condition cases and non -disease / condition cases in each fold was approximately the same as in the entire dataset). Each part was used once as a validation set while the remaining parts served as the training set. This process was repeated ten times, ensuring each subset was used exactly once as validation data.

[0151] For each fold, Ridge regression models (alpha = 0 ) were fitted across a range of regularization parameter values (lambda). Lambda is a regularization parameter that controls the degree of shrinkage applied to the coefficients, helping to manage multicollinearity among predictors. Higher values oflambda increase the amount of shrinkage, thus reducing the impact of multicollinearity by penalizing large coefficients. The lambda that minimized the cross-validated prediction error for each fold was recorded. The average of these optimal lambda values from all folds was selected for the final model configuration. The final Ridge regression model was then trained using the entire training dataset and the selected lambda value.

[0152] Within each fold of the cross-validation, the Cox model was refitted to the corresponding subset of training data using the selected lambda. The model's performance was assessed on the validation subset using Harrell's concordance index (C -index) to measure the model's prediction accuracy and evaluate its ability to correctly rank pairs of individuals by their survival times.Docket Number: 11760-025W01

[0153] After model validation, a composite risk score was calculated for each individual in the held-out testing set. This score integrated the accelerometer metrics weighted by their respective coefficients obtained from the Cox model. The composite score quantifies an individual's relative risk based on their unique activity patterns. To evaluate the models using this held out test dataset, C-statistics and bootstrap confidence intervals were computed for three metrics: steps per day, total moderate -to-vigorous physical activity duration, and the composite score. Hazard ratios, 95% confidence intervals, and pvalues were used to compare effect sizes for these metrics for each disease / condition.

[0154] The study applied these methods to two sets of models to train models according to implementations of the present disclosure. First, the study trained models on each disease / condition separately and tested coefficient weightings that are disease / condition-specific. Second, the study trained one model on all-cause mortality and applied these coefficient weightings to other diseases / conditions to determine whether a single composite PA score could function well across a wide array of chronic diseases and health conditions.

[0155] In addition to single-variable models, the study evaluated the composite PA score in models that included a range of covariates. Models were adjusted for a range of covariates measured at the baseline exam prior to the accelerometer study. Education was coded as having a college or university degree vs. no college or university degree.Socioeconomic status was assessed by the Townsend Deprivation Index, which is calculated using four variables obtained from census data: unemployment rate, non-car ownership, nonhome ownership, and household overcrowding. These variables are combined to create a single composite score, which is then standardized to have a mean of zero and a standard deviation of one. Higher scores on the Townsend Deprivation Index indicate higher levels of socioeconomic deprivation, while lower scores represent lower levels of deprivation.Ethnicity data were obtained from self-report using a set of sequential branching questionsDocket Number: 11760-025W01 with fixed categories. Ethnicity was scored as white or non-white in our models. Smoking status was self-reported as never smoker, former smoker, or current smoker. Alcohol use was determined by using self-reported alcohol consumption. The volume of alcohol consumed was multiplied by the alcohol content (percent) and then divided by 0.6 ounces of alcohol per drink-equivalent. This value was then converted to grams where 1 drink equivalent contained 14 g of alcohol. Moderate consumption was considered > 0 to <= 14 g / day for women and > 0 and <= 28 g / day for men, and we coded alcohol consumption as: never, moderate, or excessive. Body mass index (BMI) was included in fully adjusted models. Adherence to a healthy diet was calculated. A healthy diet was considered one that included at least 4 of the following 7 categories: 1 ) >= 3 servings / day of fruit; 2 ) >= 3 servings / day of vegetables: 3 ) >= 2 servings / week of fish; 4) <= 1.5 servings / week of unprocessed red meats; 5) <= 1 serving / week of processed meats; 6) >= 3 servings / day of whole grains; 7) <= 1.5 servings / day of refined grains. The study tested the proportionality of hazards assumption using Schoenfeld residuals ( p > 0.05 for all models).

[0156] Finally, the study examined whether inclusion of sleep in the metric list altered model performance. The study ran a separate set of analyses including all PA and SB variables and sleep variables. These were run as a second analysis because many users of consumer wearables charge their devices at night during sleep, making a non-sleep metric potentially more widely acceptable.

[0157] RESULTS

[0158] Sample sizes and case counts for each disease and health condition are provided in Table 1.

[0159] FIGS. A-5B illustrate hazard ratios for risk of incident chronic disease for disease-specific PA composite scores as shown in FIG. 5A and mortality-basedDocket Number: 11760-025W01PA Disease-specific composite score as shown in FIG. 5B. For comparison, light-gray points are HRs for MVP A and blue points are HRs for steps / day. Horizontal lines are 95% Cis.

[0160] The three PA metrics were significantly associated with all analyzed diseases / conditions in single variable models as shown in FIGS. 5A-5B. Using PA metrics scaled to have a mean of 0 and SD of 1, HRs are significantly lower per SD of metric for the composite PA score compared with MVPA or step counts for most diseases / conditions as shown in FIGS. 5A-5B. In fact, MVPA had a similar hazard ratio compared with the composite PA score only for Parkinson's disease as shown in FIGS. 5A-5B.

[0161] Across most diseases / conditions analyzed in the study, the composite PA score outperformed two traditional PA metrics, MVPA and step counts in terms of the c-index as shown in FIG. 6A-6B, Table 1). FIGS. 6A-6B illustrate C-indices for chronic disease prediction models. FIG. 6A illustrates example disease-specific PA composite scores. FIG. 6B illustrates mortality-based PA composite scores. For comparison, light-shaded points are c-indices for MVPA and blue points are c-indices for steps / day. Horizontal lines are 95% Cis.

[0162] Table 1: Results for single variable models for disease-specific and mortality based composite PA score and comparative metrics (steps / day and MVPA).c- C- c- index HR pindex mdex Compo HR HR Compo compos cases / c Model Disease Steps MVPA site PA Steps p- steps MVPA pmvpa site PA ite ontrols exclude 0.55 0.58 0.62 0.82 0.73 0.34Heart_ (0.54.0. (0.57.0. (0.61,0. (0.8, 0.8 2.05E- (0.71,0. 4.76E- (0.37,0. 2.13E- 9000 / 6 Disease 56) 59) 63) 5) 33 76) 72 32) 164 1143 14093 0.52 0.55 0.56 0.92 0.84 0.21(0.51,0. (0.54,0. (0.56,0. (0.89,0, 7.77E- (0.81,0. 5.54E- (0.26,0. 6.19E- 9151 / 5 Cancer 53) 56) 57) 95) 08 87) 27 17) 51 9870 15215 0.64 0.64 0.69 0,57 0.54 0.29Diabete (0.62,0. (0.63,0. (0.67,0. (0.53,0. 6.8 IE- (0.49,0. 3.79E- (0.33,0. 1.24E- 1604 / 766) 66) 71) 62) 42 59) 44 26) 81 9095 3537 0.54 0,54 0.61 0.86 0.88 0.22Osieoar (0.53.0. (0.53,0. (0.6, 0.6 (0.82,0. 1.99E- (0.84,0. 8.62E- (0.26,0. 7.27E- 5015 / 6 thritis 55) 55) 2) 89) 13 91) 10 19) 76 7524 11697 0.67 0.67 0.72 0.51 0.49 0.25(0.65,0. (0.64,0. (0.69,0. (0.46,0. 2.33E- (0.43.0. 2.50E- (0.29,0. 7.07E- 1042 / 8 COPD 69) 69) 74) 57) 37 54) 36 22) 77 1805 1389 0.58 0.64 0.65 0.73 0.57 0.1(0.56,0. (0.62,0. (0.63,0. (0.67,0. 2.03E- (0.52,0. 1.56E- (0.13,0, 1.48E- 1390 / 8Stroke 6) 66) 67) 79) 14 62) 33 07) 51 1338 1508Docket Number: 11760-025WO1Mortality-based 0.56 0.58 0.62 0.79 0.74 0.34Hypeit (0.55,0. (0.57,0. (0.6, 0.6 (0.76,0, 9.83E- (0.7,0.7 1.52E- (0.38,0. 2.35E- 4925 / 5 ension 57) 59) 3) 83) 27 7) 39 31) 88 7395 21916 Chronic.. Kidn 0.63 0.66 0.69 0.61 0.52 0.31ey_Dis (0.61.0. (0.64.0. (0.67,0. (0.56,0. 1.04E- (0.48,0. 1.07E- (0.35,0. 3.16E- 1660 / 8 ease 65) 68) 71) 66) 34 57) 47 28) 90 1012 1564 0.59 0.57 0,6 0.71 0.78 0.16Depres (0.57,0. (0.55,0. (0,58,0, (0.65,0. 2.27E- (0.72.0. 7.74E- (0.23,0. 4.97E- 1366 / 7 sion 61) 59) 62) 77) 16 85) 09 11) 7961 4909 0.62 0.77 0.79 0.65 0.25 0.25Parkins (0.57.0. (0.73.0. (0.75,0. (0.55,0. 1.71E- (0.2, 0,3 1.15E- (0.29,0. 1.59E- 395 / 83 OHS 66) 82) 83) 76) 07 1) 34 21) 66 688 153 0.59 0.67 0.72 0.69 0.47 0.26Demen (0.55.0. 10.63,0. (0.69,0. (0.6.0.7 1.75E- (0.4,0.5 2.58E- (0.32,0. 4.23E- 468 / 83 tia 62) 7) 75) 9) 07 5) 19 21) 41 714 54 0.62 0.67 0.69 0.63 0.49 0.33(0.6, 0.6 (0.65,0. (0.67,0. (0.59,0. 3.50E- <0.46,0. 1.74E- (0.35,0. 6.02E- 2825 / 8 Death 68) 7) 66) 54 52) 95 3) 177 1406 5 0.55 0.59 0.62 0.82 0.7 0.49Heart_ (0.54.0. (0.58.0. (0.61,0. (0.8, 0.8 2.66E- (0.68,0. 9.40E- (0.52,0. 7.26E- 9000 / 6 Disease 56) 6) 63) 5) 34 73) 90 47) 175 1143 14093 0.52 0.55 0.56 0.91 0.84 0.72(0.52,0. (0.54,0. (0.55,0. (0.88,0, 1.02E- (0.82,0. 5.51E- (0.75,0. 6.43E- 9151 / 5 Cancer 53) 55) 57) 94) 09 87) 26 68) 41 9870 15215 0.61 0.63 0.66 0,64 0.58 0.4Diabete (0.6, 0.6 (0.61,0. (0.64,0. (0.59,0. 2.48E- <0.53,0. 6.11E- (0.45,0. 7.04E- 1604 / 74) 65) 67) 69) 29 63) 36 36) 62 9095 3537 0.54 0.55 0.59 0.85 0.85 0.61Osieoar (0.53.0. (0.54,0. (0.57,0. (0.82,0. 1.00E- (0.81,0. 4.95E- (0.66,0. 5.80E- 5015 / 6 thritis 56) 56) 6) 89) 13 88) 14 58) 48 7524 11697 0.66 0.67 0.71 0.54 0.49 0.31(0.64,0. (0.64,0. (0.69,0. (0.49,0. 2.51E- (0.44.0. 1.26E- (0.35,0. 3.25E- 1042 / 8 COPD 68) 69) 73) 6) 54) 36 27) 74 1805 1389 0.57 0.63 0.65 0.78 0.59 0.41(0.54,0. (0.61,0. (0.63,0. (0.72,0. 1.24E- (0.54,0. 1.27E- (0.46,0. 2.32E- 1390 / 8 Stroke 59) 65) 67) 84) 09 65) 28 36) 50 1338 1508 0.56 0.59 0.62 0.8 0.7 0.49Hypert (0.55,0. <0.58,0. (0.61,0. (0.77,0. 1.14E- (0.67,0. 3.12E- (0.52,0. 4.59E- 4925 / 5 ension 57) 6) 63) 84) 23 73) 52 46) 100 7395 21916 Chronic_Kidn 0.63 0.65 0.69 0.6 0.53 0.34ey_Dis (0.61.0. (0.63.0. (0.67,0. (0.56,0. 2.72E- (0.49,0. 2.72E- (0.38,0. 6.61E- 1660 / 8 ease 65) 67) 71) 65) 37 58) 46 31) 95 1012 1564 0.59 0.58 0.59 0.69 0.74 0.54Depres (0.57,0. (0.56,0. (0.57,0. (0.63,0, 3.64E- (0.68,0. 2.04E- (0.61,0. 1.40E- 1366 / 7 sion 62) 61) 62) 75) 19 8) 12 48) 23 7961 4909 0.64 0.79 0.79 0,59 0.2 0.21Parkins (0.59,0. (0.76,0. (0.76,0. (0.51,0. 7.23E- <0.16,0. 1.10E- (0.25,0. 1.54E- 395 / 83 ons 68) 83) 82) 69) 11 25) 44 17) 62 688 153 0.59 0.66 0.71 0.7 0.5 0.29Demen (0.54.0. (0.62.0. (0.67,0. (0.61,0. 1.15E- (0.42,0. 3.37E- (0.35,0. 1.52E- 468 / 83tia 63) 7) 74) 81) 06 59) 16 24) 37 714 54 00163] In general, daily step counts provide the lowest discriminant power across diseases / conditions. For Parkinson's disease and stroke, the composite score and MVP A provide similar discriminant power. For depression, all three metrics provide similar discriminant power. Other diseases / conditions show the clear benefits of using the composite score. For example, dementia, diabetes, heart disease, COPD, and osteoarthritis all show a clear advantage for the composite PA score in terms of the cindex values.Docket Number: 11760-025W01Disease-specific Mortality-based _

[0164] In models according to implementations of the present disclosure, adjusted for a wide range of covariates, for all diseases / conditions, the composite PA score and MVPA remain significant predictors of incident disease / condition, however step counts are no longer significant predictors of osteoarthritis. In all cases except for osteoarthritis, adding the composite PA score to a model with co variates leads to an improvement in the c- index compared with a model with covariates alone (Table 2). This increase in the c-index is larger than the c-index change associated with adding either step counts or MVPA to the covariate only model (see Table 2).

[0165] Table 2: Change in C-index when adding PA metrics to model that includes only covariatesmod C -difference C -difference C -difference el disease Steps MVPA Composite PA Heart.. Disease 0.001 4.62E-04 0.002 Cancer 0.000 0.001 0.001 Diabetes 0.004 0.005 0.009 Osteoarthritis -1.36E-05 0.001 0.001 COPD 0.007 0.002 0.009 Stroke 0.002 0.004 0.006 Hypertension 0.002 0.001 0.003 Chroni c_Kidney__Di sease 0.004 0.003 0.006 Depression 0.005 0.003 0.007 Parkinson’s 0.027 0.085 0.093 Dementia 0.003 0.006 0.016 Heart-Disease 0.000 0.001 0.002 Cancer 0.001 0.001 0.001 Diabetes 0.002 0.003 0.003 Osteoarthritis -1.71E-06 0.001 8.54E-05 COPD 0.009 0.007 0.013 Stroke 4.68E-04 0.005 0.004 Hypertension 0.001 0.002 0.004 Chronic. Kidney „Disease 0.003 0.002 0.005 Depression 0.006 0.005 0.010 Parkinson’s 0.022 0.082 0.082Dementia 0.003 0.006 0.015Docket Number: 11760-025W01 Disease-specific

[0166] In addition to disease / condition specific models, the study examined whether a model trained on all-cause mortality according to the methods described herein would produce scores that predict other chronic disease / condition outcomes. HRs are generally smaller for the mortality-based score compared with disease / condition-specific scores, however the composite PA score still shows the largest effect sizes across diseases / conditions (FIGS. 5A-5B). While c-statistics are also slightly lower for the mortality based composite score compared with disease / condition-specific scores, the mortality-based composite still outperforms MVP A and step counts in most cases (FIGS. 6A-6B). When covariates are included in models, all three metrics remain significant predictors of disease / condition risk, and the composite PA score has the largest effect size (Table 3). Similar to disease / condition-specific scores, adding the mortality-based composite PA score to models that include covariates increases the c-index. This increase is generally larger than the increases linked with adding step counts or MVP A (see Table 2).

[0167] Table 3: Results for models for disease-specific and mortality based composite PA score and comparative metrics (steps / day and MVPA) including covariates.C- PM inde P HR co case ex 0 inde X index P- HR m Comp nip s / co cl d X MVP Compo HR ste MVP vp osite osit ntrol ud el Disease Steps A site PA Steps ps A a PA e s e 0.74 0.74 0.95 1.5 0.94 1.2(0.73 (0.73 0.74 (0.91 3E (0.91 8E 0.76 9.8 900 14 Heart.. D,0.74,0.74 (0.73,0.,0.98,0.98 (0.84, 7E- 0 / 61 09 isease ) ) 75) ) 03 ) 03 0.69) 09 143 33.4 0.94 1.80.61 0.61 0.61 0.97 3E (0.91 3E 0.58 1.7 915 15 (0.6, (0.6, (0.61,0. (0.93.0.97 (0.75, 8E- 1 / 59 21 Cancer 0.62) 0.62) 62), D 02 ) 04 0.46) 05 870 50.77 7.2 2.40.78 0.78 0.79 (0.71 IE 0.77 7E 0.56 1.0 160 Diabete (0.77 (0.77 (0.78,0.,0.84 (0.7, (0.65, 1E- 4 / 79 35s,0.8),0.8) 81) ) 09 0.84) 08 0.48) 14 095 37Docket Number: 11760-025WO1 Mortality-based 0.67 0.67 1 8.6 1.09 5.3(0.66 (0.66 0.67 (0.96 4E (1.05 5E 0.73 1.5 501 11 Osteoart,0.68,0.68 (0.66,0.,1.05,1.14 (0.89, SE5 / 67 69 hritis ) ) 68) ) 01 ) 05 0.6) 05 524 70.83 0.83 0.64 2.2 0.73 2.1(0.81 (0.81 0.83 (0.58 IE (0.65 3E 0.44 5.5 104,0.85,0.84 (0.82,0.,0.72,0.82 (0.52, 2E- 2 / 81 13 COPD ) ) 85) ) 14 ) 07 0.36) 19 805 890.75 0.75 2.6 0.75 3.6(0.73 (0.73 0.75 0.83 8E (0.68 4E 0.25 3.4 139,0.77,0.77 (0.74,0. (0.76,0.82 (0.36, 5E- 0 / 81 15 Stroke ) ) 77),0.9) 05 ) 09 0.17) 14 338 080.7 0.7 0.88 1.5 0.91 2.4(0.69 (0.69 0.7 (0.84 IE (0.87 3E 0.67 5.3 492 21 Hyperte,0.71,0.71 (0.69,0.,0.92.0.96 (0.76, 2E- 5 / 57 91 nsion ) ) 71) ) 07 ) 04 0.59) 10 395 6 Chronic 0.78 0.78 0.78 2.9 9.7_Kidney (0.76 (0.76 0.78 (0.71 0E 0.77 IE 0.58 2.0 166 „Diseas,0.79,0.79 (0.76,0.,0.85 (0.7, (0.67, SE0 / 81 15 e ) ) 8) ) 08 0.85) 08 0.5) IS 012 640.77 0.76 0.8 1.4 0.85 8.4(0.75 (0.75 0.77 (0.73 5E (0.78 9E 0.28 7.9 136 Depress,0.78,0.78 (0.75,0.,0.87.0.94 (0.43, 2E- 6 / 77 49 ion ) ) 79) ) 06 ) 04 0.18) 09 961 090.78 0.84 0.59 5.3 0.24 4.2(0.75 (0.81 0.84 (0.48 0E (0.19 7E 0.24 9.4 395 / Parkins,0.82,0.87 (0.81,0.,0.71,0.32 (0.29, 1E- 836 15 ons ) ) 88) ) 08 ) 26 0.2) 45 88 30.82 0.81 7.5 0.73 3.9(0.79 0.82 0.83 (0.69 3E (0.61 8E 0.45 1.6 468 / Dementi,0.84 (0.8, (0.81,0.,0.94,0.87 (0.58, 5E- 837 a ) 0.84) 85) ) 03 ) 04 0.36) 10 14 540.74 0.74 0.94 6.5 0.92 1.5(0.73 (0.73 0.74 (0.91 5E (0.88 4E 0.81 1.3 900 14 Heart-D,0.75,0.75 (0.73,0.,0.98,0.95 (0.86, 0E- 0 / 61 09 isease ) ) 75) ) 04 ) 06 0.76) 12 143 30.94 2.5 7.30.61 0.61 0.61 (0.91 8E 0.93 9E 0.88 1.1 915 15 (0.6, (0.6, (0.6, 0.6,0.97 (0.9, (0.93, SE1 / 59 21 Cancer 0.62) 0.62) 2) ) 04 0.97) 05 0.83) OS 870 50.78 0.78 1.1 1.4(0.76 (0.76 0.78 0.87 9E 0.82 9E 0.72 2.1 160 Diabete,0.79,0.79 (0.76,0. (0.8, (0.75 (0.82, 9E- 4 / 79 35 s ) ) 79) 0.95) 03,0.9) 05 0.63) 06 095 370.68 0.68 1 8.8 1.08 6.1(0.67 (0.67 0.68 (0.96 4E (1.03 5E 1.05 2 3 501 11 Osteoart,0.69,0.69 (0.67,0..1.05.1.13 (1.14, 5E- 5 / 67 69hritis ) ) 69) ) 01 ) 04 0.97) 01 524 7Docket Number: 11760-025W010.82 0.68 6.4 2.0(0.81 0.82 0.83 (0.61 6E 0.71 6E 0.51 1.8 104,0.84 (0.8, (0.81,0.,0.76 (0.63 (0.6,0 1E- 2 / 81 13 COPD ) 0.84) 84) ) 12,0.8) 08.43) 15 805 890.74 0.74 0.88 6.5 2.2(0.72 (0.72 0.74 (0.81 0E 0.77 2E 0.62 4.0 139,0.76,0.76 (0.72,0.,0.97 (0.7, (0.72, 6E- 0 / 81 15 Stroke ) ) 76) ) 03 0.85) 07 0.53) 10 338 080.91 8.4 0.89 2.60.7 0.71 0.71 (0.87 4E (0.85 3E 0.76 6.4 492 21 Hyperte (0.7, (0.7, (0.7, 0.7,0.95,0.94 (0.83, 0E- 5 / 57 91 nsion 0.72) 0.72) 2) ) 05 ) 06 0.71) 11 395 6 Chronic 0.78 0.79 8.9 0.83 5.2_Kidney 0.78 (0.76 0.78 (0.73 9E (0.76 0E 0.66 5.0 166 „Diseas (0.76,0.79 (0.76,0.,0.86.0.91 (0.76, 9E- 0 / 81 15 e,0.8) ) 8) ) 08 ) 05 0.58) 09 012 640.76 0.76 5.7 0.8 3.2(0.75 (0.74 0.76 0.77 5E (0.72 4E 0.61 5.3 136 Depress,0.78,0.78 (0.75,0. (0.7,,0.88 (0.71, 9E- 6 / 77 49 ion ) ) 79) 0.85) 08 ) 06 0.53) 11 961 090.77 0.83 0.61 6.2 0.24 7.3(0.75 (0.81 0.83 (0.51 9E (0.19 5E 0.2 5.5 395 / Parkins,0.81,0.87 (0.81,0.,0.73.0.31 (0.26, 3E- 836 15 ons ) ) 87) ) 08 ) 29 0.16) 38 88 30.81 0.81 5.2 0.73 7 2(0.78 (0.79 0.82 0.85 4E (0.61 IE 0.45 2.0 468 / Dementi,0.84,0.84 (0.8, 0.8 (0.73,0.88 (0.59, 9E- 837a ) ) 5) 4) 02 ) 04 0.35) 09 14 54

[0168] Finally, the study ran these same models and analyses with data for sleep included. Using accelerometer derived sleep and sleep quality metrics (C-statistics and HRs for the composite score were similar to those calculated without the sleep data and compared similarly with models that do not include sleep data, as shown in Tables 4 and 5 below).

[0169] Table 4: Results for single variable models for disease-specific and mortality based composite PA score that includes sleep metrics and comparative metrics (steps / day and MVPA).M P HR case ex 0 inde index P- HR m Comp P s / co cl d X inde Compo HR ste MVP vp osite co ntrol udel Disease Steps X site PA Steps ps A a PA mp sDocket Number: 11760-025WO1 Disease-specific MVP ositA e0.55 0.58 2.0 0.73 4.7(0.54 (0.57 0.62 0.82 5E (0.71 6E 0.34 2.8 900 14 IIeart_D,0.56,0.59 (0.61,0. (0.8,,0.76 (0.37, 8E- 0 / 61 09 isease ) ) 63) 0.85) 33 ) 72 0.32) 165 143 30.52 0.55 0.92 7.7 0.84 5.5(0.51 (0.54 0.56 (0.89 7E (0.81 4E 0.21 3.5 915 15,0.53,0.56 (0.56,0.,0.95,0.87 (0.25, 41 - 1 / 59 21 Cancer ) ) 57) ) 08 ) 27 0.17) 51 870 50.64 0.64 0.57 6.8 0.54 3.7(0.62 (0.63 0.69 (0.53 IE (0.49 9E 0.29 1.6 160 Diabete,0.66,0.66 (0.68,0.,0.62,0.59 (0.33, 1E- 4 / 79 35 s ) ) 71) ) 42 ) 44 0.26) 85 095 370.54 0.54 0.86 1.9 0.88 8.6(0.53 (0.53 0.61 (0.82 9E (0.84 2E 0.23 1.8 501 11 Osteoarl.0.55,0.55 (0.6, 0.6,0.89,0.91 (0.27. 41.- 5 / 67 69 hritis ) ) 2) ) 13 ) 10 0.2) 76 524 70.67 0.67 0.51 2.3 0.49 2.5(0.65 (0.64 0.72 (0.46 3E (0.43 0E 0.26 7.8 104,0.69,0.69 (0.7, 0.7,0.57,0.54 (0.3,0 3E- 2 / 81 13 COPD ) ) 5) ) 37 ) 36.22) 82 805 890.64 0.73 2.0 0.57 1.50.58 (0.62 0.65 (0.67 3E (0.52 6E 0.09 3.3 139 (0.56,0.66 (0.63,0.,0.79,0.62 (0.12, 41.- 0 / 81 15 Stroke,0.6) ) 67) ) 14 ) 33 0.06) 52 338 080.56 0.58 0.79 9.8 1.5(0.55 (0.57 0.62 (0.76 3E 0.74 2E 0.34 3.6 492 21 Hyperte,0.57,0.59 (0.61,0.,0.83 (0.7, (0.38, 5E- 5 / 57 91 nsion ) ) 63) ) 27 0.77) 39 0.3) 90 395 6 Chronic 0.63 0.66 0.61 1.0 0.52 1.0JKidney (0.61 (0.64 0.69 (0.56 4E (0.48 7E 0.3 7.6 166 „Diseas.0.65,0.68 (0.67.0.,0.66,0.57 (0.34, 9E- 0 / 81 15 e ) ) 71) ) 34 ) 47 0.27) 90 012 640.59 0.57 0.71 2.2 0.78 7.7(0.57 (0.55 0.6 (0.65 7E (0.72 4E 0.12 7.7 136 Depress,0.61,0.59 (0.58,0.,0.77,0.85 (0.19, 7E- 6 / 77 49 ion ) ) 62) ) 16 ) 09 0.08) 23 961 090.62 0.77 0.65 1.7 1.1(0.57 (0.73 0.79 (0.55 IE 0.25 5E 0.23 1.0 395 / Parkins.0.66,0.82 (0.75,0.,0.76 (0.2, (0.27, 8E- 836 15 ons ) ) 83) ) 07 0.31) 34 0.19) 64 88 30.59 1.7 2.5(0.55 0.67 0.72 0.69 5E 0.47 8E 0.26 1.9 468 / Dementi,0.62 (0.63 (0.69,0. (0.6, (0.4, (0.31, 3E- 837 a ),0.7) 75) 0.79) 07 0.55) 19 0.21) 41 14 540.67 0.63 3.5 0.49 1.70.62 (0.65 0.69 (0.59 0E (0.46 4E 0.32 2.3 282 (0.6,,0.68 (0.67.0.,0.66,0.52 (0.35, 7E- 5 / 81Death 0.63) ) 7) ) 54 ) 95 0.3) 174 406 5Docket Number: 11760-025WO1 blid Mttoraasey- 0.55 3.7 0.71 1.5(0.54 0.59 0.62 0.83 IE (0.69 7E 0.49 1.7 900 14 Heart_D,0.56 (0.58 (0.61,0. (0.8,,0.74 (0.51, 6E- 0 / 61 09 isease ),0.6) 63) 0.86) 30 ) 82 0.46) 164 143 30.53 0.55 0.91 3.9 0.84 3.9(0.52 (0.54 0.56 (0.88 5E (0.81 7E 0.71 1.1 915 15,0.53,0.56 (0.55,0.,0.93,0.87 (0.75, 7E- 1 / 59 21 Cancer ) ) 57) ) 10 ) 26 0.68) 39 870 50.62 0.63 0.64 1.7 0.57 8.1(0.59 (0.61 0.66 (0.59 0E (0.52 2E 0.38 4.5 160 Diabete,0.64,0.65 (0.64,0.,0.69,0.62 (0.42, 0E- 4 / 79 35 s ) ) 68) ) 27 ) 36 0.34) 62 095 370.54 0.55 2.3 0.86 5.0(0.53 (0.54 0.58 0.86 3E (0.82 5E 0.61 3.8 501 11 Osteoart,0.55,0.56 (0.57,0. (0.82.0.89 (0.66, 9E- 5 / 67 69 hritis ) ) 59),0.9) 12 ) 12 0.57) 44 524 70.66 0.67 1.3 0.49 1.6(0.64 (0.64 0.71 0.54 3E (0.43 9E 0.3 8.9 104,0.68,0.69 (0.69,0. (0.49,0.55 (0.34, 6E- 2 / 81 13 COPD ) ) 74),0.6) 30 ) 34 0.26) 71 805 890.56 0.63 0.78 1.1 0.59 9.5(0.54 (0.61 0.65 (0.72 1E (0.53 9E 0.4 5.3 139,0.59,0.66 (0.63,0.,0.85.0.65 (0.46, 6E- 0 / 81 15 Stroke ) ) 67) ) 08 ) 28 0.36) 47 338 080.56 0.81 6.3 0.71 2.6(0.55 0.59 0.62 (0.78 5E (0.67 7E 0.48 1.2 492 21 Hyperte,0.57 (0.58 (0.61,0.,0.85,0.74 (0.52, 1E- 5 / 57 91 nsion ),0.6) 63) ) 21 ) 47 0.45) 92 395 6 Chronic 0.65 0.62 4.5 1.9_Kidney 0.62 (0.63 0.68 (0.57 IE 0.55 4E 0.34 1.6 166 _Diseas (0.6,,0.66 (0.66,0.,0.67 (0.5, (0.38, 4E- 0 / 81 15 e 0.64) ) 7) ) 31 0.6) 40 0.31) 81 012 640.59 0.69 2.7 0.75 5.2(0.57 0.58 0.59 (0.64 7E (0.68 8E 0.55 2.3 136 Depress,0.61 (0.56 (0.57,0.,0.75,0.81 (0.62, 2E- 6 / 77 49 ion ),0.6) 62) ) 17 ) 11 0.48) 20 961 090.79 2.0 0.2 8.30.64 (0.75 0.8 0.59 1E (0.16 0E 0.19 1.7 395 / Parkins (0.6,,0.82 (0.76,0. (0.5,.0.25 (0.23, 4E- 836 15 ons 0.68) ) 83) 0.69) 10 ) 43 0.16) 60 88 30.58 0.72 8.3 0.5 1.6(0.54 0.66 0.71 (0.62 3E (0.42 9E 0.28 1.7 468 / Dementi,0.62 (0.62 (0.67,0.,0.83,0.59 (0.34, 9E- 837a ),0.7) 74) ) 06 ) 15 0.23) 35 14 54Docket Number: 11760-025W01 Disease-specific

[0170] Table 5: Results for models for disease-specific and mortality based composite PA score that includes sleep metrics and comparative metrics (steps / day and MVPA) including covariates.C- PM inde P HR co case ex 0 inde X index P- HR m Comp mp s / co cl d X MVP Compo HR ste MVP vp osite osit ntrol ud el Disease Steps A site PA Steps ps A a PA e s e 1.2 0.91 1.20.69 0.69 0.7 0.93 2E (0.88 7E 0.68 2.2 900 14 Heart_D (0.69 (0.69 (0.69,0. (0.9,,0.94 (0.75, 8E- 0 / 61 09 isease •0.7),0.7) 7) 0.96) 05 ) 07 0.62) 16 143 30.96 6.9 2.20.61 0.61 0.61 (0.93 5E 0.93 4E 0.56 2.2 915 15 (0.6, (0.6, (0.6, 0.6,0.99 (0.9, (0.71, 4E- 1 / 59 21 Cancer 0.62) 0.62) 2) ) 03 0.96) 05 0.44) 06 870 50.75 5.6 0.74 6.00.78 0.78 0.79 (0.69 3E (0.67 8E 0.53 6.7 160 Diabete (0.76 (0.77 (0.77,0.,0.82,0.81 (0.61, 7E- 4 / 79 35 s,0.8),0.8) 8) ) 11 ) 11 0.46) 19 095 370.66 0.66 0.98 4.2 1.07 9.5(0.65 (0.65 0.67 (0.94 8E (1.03 4E 0.65 8.1 501 11 Osteoart,0.67,0.68 (0.65,0.,1.03,1.12 (0.79, 1E- 5 / 67 69 hritis ) ) 68) ) 01 ) 04 0.54) 06 524 70.83 0.65 4.8 0.72 2.9(0.81 0.82 0.83 (0.58 6E (0.64 3E 0.43 3.5 104,0.84 (0.8, (0.81,0.,0.72.0.81 (0.51, 3E- 2 / 81 13 COPD ) 0.84) 85) ) 15 ) 08 0.36) 22 805 890.74 0.74 0.8 7.6 1.1(0.72 (0.72 0.74 (0.74 6E 0.73 4E 0.21 5.5 139,0.76,0.76 (0.73,0.,0.88 (0.66 (0.3,0 2E- 0 / 81 15 Stroke ) ) 76) ) 07,0.8) 10.15) 17 338 080.7 0.7 0.88 2.3 0.91 5.0(0.69 (0.69 0.7 (0.84 7E (0.87 0E 0.65 1.7 492 21 Hyperte,0.71,0.71 (0.69,0.,0.92,0.95 (0.74, 7E- 5 / 57 91 nsion ) ) 71) ) 08 ) 05 0.58) 11 395 6 Chronic 0.76 0.76 5.2 0.75 6.3_Kidney (0.74 (0.74 0.76 0.76 2E (0.68 2E 0.54 7.9 166 _Diseas,0.78,0.78 (0.75,0. (0.7,,0.82 (0.63, 6E- 0 / 81 15 e ) ) 78) 0.83) 10 ) 10 0.47) 17 012 640.64 0.63 0.78 5.1 0.83 8.3(0.62 (0.61 0.64 (0.71 7E (0.76 6E 0.19 1.4 136 Depress,0.66,0.66 (0.63,0.,0.85,0.91 (0.31, 7E- 6 / 77 49 ion ) ) 67) ) 08 ) 05 0.12) 11 961 090.83 0.84 0.25 0.23 2.7 395 / Parkins 0.78 (0.8, (0.81,0. 0.59 2.5 (0.2, 9.1 (0.28, 9E- 836 15ons (0.75 0.87) 87) (0.49 3E 0.33) 1E 0.19) 45 88 3Docket Number: 11760-025WO1 blid Mttoraasey-,0.82,0.71) ) 08 270.82 1.1 0.69 1.5(0.79 0.82 0.83 0.78 OE (0.58 8E 0.41 2.7 468 / Dementi,0.84 (0.8, (0.81,0. (0.67,0.82 (0.52, 2E- 837 a ) 0.84) 86),0.9) 03 ) 05 0.33) 13 14 540.7 0.7 8.8 0.89 4.7(0.69 (0.69 0.7 0.93 5E (0.86 8E 0.75 in 900 14 Heart_D,0.71,0.71 (0.69,0. (0.9,,0.92 (0.8,0 3E- 0 / 61 09 isease ) ) 71) 0.96) 06 ) 11.71) 21 143 34.6 0.92 6.00.61 0.61 0.61 0.93 2E (0.89 4E 0.87 8.1 915 15 (0.6, (0.6, (0.6, 0.6 (0.9,,0.96 (0.93, 1E- 1 / 59 21 Cancer 0.62) 0.62) 2) 0.97) 05 ) 06 0.82) 06 870 50.77 0.77 0.86 2.7 0.8 1.2(0.76 (0.76 0.78 (0.79 9E (0.73 8E 0.68 1.7 160 Diabete,0.79,0.79 (0.76,0.,0.93,0.88 (0.78, BE4 / 79 35 s ) ) 79) ) 04 ) 06 0.59) 05 095 370.68 0.68 0.99 6.2 1.07 3.4(0.67 (0.67 0.68 (0.95 7E (1.02 8E 1.01 8.9 501 11 Osteoart,0.69,0.69 (0.67,0.,1.03,1.11 (1.09, 6E- 5 / 67 69 hritis ) ) 69) ) 01 ) 03 0.93) 01 524 72.8 0.7 1.10.82 0.81 0.82 0.67 9E (0.62 IE 0.48 2.2 104 (0.8, (0.8, (0.81,0. (0.6,,0.78 (0.57, 3E- 2 / 81 13 COPD 0.84) 0.83) 84) 0.75) 13 ) 09 0.41) 18 805 890.73 0.74 1.7 0.76 3.6(0.71 (0.72 0.74 0.87 0E (0.69 3E 0.6 2.3 139,0.75,0.76 (0.72,0. (0.8,,0.84 (0.7,0 6E- 0 / 81 15 Stroke ) ) 76) 0.95) 03 ) 08.52) 11 338 080.7 0.7 0.91 2.2 0.88 1.2(0.69 (0.69 0.7 (0.87 6E (0.84 9E 0.74 4.4 492 21 Hyperte,0.71,0.71 (0.69,0.,0.95,0.92 (0.81, 8E- 5 / 57 91 nsion ) ) 71) ) 05 ) 07 0.69) 13 395 6 Chronic 0.76 0.76 0.77 2.0 0.8 1.6_Kidney (0.75 (0.75 0.77 (0.71 6E (0.73 0E 0.62 6.4 166 _Diseas,0.78,0.78 (0.75,0.,0.84,0.88 (0.71, 6E- 0 / 81 15) ) 78) ) 09 ) 06 0.54) 12 012 64 0.63 0.63 0.75 2.0 0.78 2.1(0.62 (0.61 0.64 (0.68 2E (0.71 IE 0.58 1.5 136 Depress,0.66,0.66 (0.62,0.,0.82,0.86 (0.67, 7E- 6 / 77 49 ion ) ) 67) ) 10 ) 07 0.5) 13 961 090.78 0.84 6.6 2.1(0.75 (0.81 0.84 0.59 5E 0.23 5E 0.19 2.1 395 / Parkins,0.81,0.87 (0.81,0. (0.5, (0.18 (0.24, 2E- 836 15 ons ) ) 87) 0.71) 09,0.3) 31 0.15) 41 88 30.81 0.81 0.84 2.7 0.74 6.0(0.78 (0.79 0.82 (0.72 2E (0.62 0E 0.45 1.3 468 / Dementi,0.83,0.84 (0.8, 0.8,0.98,0.88 (0.59, 8E- 837a ) ) 5) ) 02 ) 04 0.35) 09 14 54Docket Number: 11760-025W01

[0171] DISCUSSION

[0172] For 11 key age-related chronic disease and health condition outcomes, a composite PA variable generated from the machine learning methods described herein outperforms standard PA metrics in survival analyses. Inclusion of sleep metrics provides similarly strong composite activity variables for this set of chronic disease / condition outcomes. The study shows that the comprehensive approach to accelerometer data according to implementations of the present disclosure can provide enhanced predictive metrics for chronic diseases and health conditions. In addition, these metrics can be useful for individuals to monitor from their own wearable devices.

[0173] The implementations described herein provide improvements over conventional methods of machine learning applied to physical activity datasets. Conventional methods focus on single-condition models of physical activity that fail to capture the complexity of physical activity data.

[0174] Example Novel aspects of PA Composite used in the example implementation.

[0175] The study shows that example implementations of the present disclosure advance applications of wearable device data to chronic disease / condition prediction. First, the integration of fractal complexity and wavelet energy features represents a new methodology and technical improvement over existing methods. As described herein, this can include the incorporation of wavelet transformation techniques to extract significant features from accelerometer data. By capturing both high-frequency and low-frequency components of physical activity patterns, this method allows for a detailed and nuancedDocket Number: 11760-025W01 analysis of the variability and complexity of PA behaviors. Additionally, the use in some implementation of the new Wavelet Pattern Variability Index (WPVI) further quantifies the daily variability in physical activity patterns, providing a novel method of determining metrics that can be then included into the PA composite score to improve the PA composite score.

[0176] Additionally, the inclusion of Intra-daily Variability (IV) and Inter-daily Stability (IS) metrics offers insights into the regularity and fragmentation of physical activity patterns. These metrics are crucial for understanding the rhythmicity and predictability of PA behaviors, which have significant implications for health outcomes. By incorporating these advanced analytical techniques, our model captures patterns of physical activity and inactivity, providing a more detailed and comprehensive understanding of their relationship with chronic disease risk.

[0177] Moreover, this study is the first to demonstrate the predictive power of a composite PA score across a broad spectrum of chronic diseases and health conditions, including dementia, diabetes, heart disease, COPD, and Parkinson’s disease. Unlike existing research that typically focuses on a single disease, implementations of the present disclosure can generate PA scores capable of predicting multiple health outcomes. This broad applicability enhances the utility of the composite score in numerous research, clinical, and public health contexts. The PA score disclosed herein is a new single composite score that performs well across a wide range of diseases and health conditions that encompass a variety of physiological systems. Thus, the single composite score disclosed herein can function well across diseases / conditions and can be broadly applied in clinical and research settings.

[0178] The methodologies employed in this study also incorporate several innovative techniques that extend beyond the use of specific metrics, setting this research apart from existing work. Notably, the application of the Cox proportional hazards modelDocket Number: 11760-025W01 with Ridge regression represents a strength of this study. This approach effectively handles multicollinearity and enhances model stability, ensuring that the most significant variables exert a strong influence on the predictive outcomes. In addition, a ridge regression model allows us to keep all of the variables in predictions for multiple diseases, with coefficients changing based on the disease of interest. With the same feature vector, the example implementation can effectively track trade-offs in the pattern of predicted outcomes when providing personalized recommendations. Other methods that eliminate variables make it harder to determine risk trade-offs when feature variable values change for a given participant. Furthermore, the use of 10- fold stratified cross-validation ensures the robustness and reliability of the model. This rigorous validation process minimizes overfitting and provides a comprehensive assessment of the model's predictive performance across different subsets of the data.

[0179] Another significant improvement of the example implementation is the practical application of the composite PA score as a digital metric that can be integrated into consumer-grade wearable devices. This offers a unique intermediate goal for individuals, bridging the gap between PA engagement and long-term disease prevention. Based on an individual's unique PA profile leading to their composite score, we can provide precision feedback to improve disease / condition risk prediction for each user. By providing a readily accessible metric that individuals can monitor in real-time, our approach addresses a critical barrier to sustained engagement in health enhancing physical activity-motivation through immediate and personalized feedback. This consumer-focused application has the potential to revolutionize how individuals manage their health, offering a practical tool for early intervention and continuous health monitoring.

[0180] In summary, the unique combination of advanced analytical methods, comprehensive disease / condition prediction, and practical consumer application differentiatesDocket Number: 11760-025W01 this work from existing literature. These contributions not only enhance the predictive accuracy and utility of PA metrics but also pave the way for innovative applications in both research and commercial settings.

[0181] Novel addition of other metrics

[0182] As discussed above, the study results represent a case study of our methods based on accelerometer data from wearable devices. However, it should be understood that implementations of the present disclosure can incorporate other metrics into our model and wearable devices currently extend beyond accelerometry. Such measures include, among other sensors, heart rate monitors, blood oxygen sensors, temperature sensors, fluid collection devices for sweat, glucose, with the potential for measurements of other biological systems coming soon (e.g., blood pressure). Additional capabilities that may be incorporated into the model of the example implementation include voice data from microphones, text message or phone data to assess social interactions, electrocardiograph data, sleep stages and polysomnography, IMUs, GPS, electromyography (to measure muscle activity and neuromuscular control), environmental exposures, and proximity to other devices or people.

[0183] Incorporating these other physiological data streams into the model according to an example implementation of the present disclosure based on movement data can strengthen the links between digital markers and health risks. In addition, incorporating medical records information as well as biomarkers from body fluids (e.g., blood, urine, and saliva) into the model can provide even more detailed models for disease and health condition risk assessment. The present disclosure further contemplates incorporating mobile or clinical tests of physical function, strength, and cognition (e.g., balance, grip strength, walking / gait speed, daily cognitive performance). Finally, data from questionnaires that assess demographics, environment, and health status characteristics can all enhance the predictiveDocket Number: 11760-025W01 power of the model. In the end, the model described above represents a case study in developing wearable-device based disease / condition risk in a way that may be useful to clinicians, researchers, and consumers. Incorporation of future data streams from a wide array of physiological systems will generate a highly detailed risk assessment that will grow more precise as data collection methods evolve.

[0184] Personalized recommendations

[0185] Based on these wearables-based metrics, the example model and algorithm can provide unique and personalized recommendations for users to improve health in relation to an array of disease risks. By using the ridge regression methods described above, the example implementation can include an array of metrics that play a role in the composite risk score. Each user can have a unique set of metrics that combine to generate their composite score, and by analyzing each individual separately, the example implementation can provide a personalized set of recommendations that will reduce their disease / condition risk. For example, one user may be spending a large portion of time inactive, driving up disease risk, while another user may have circadian rhythm disruptions driving their disease risk. These users would benefit from different sets of recommendations. The models described herein enable generation of these kinds of personalized recommendations that can change over time as users tackle different aspects of their movement and other behaviors.

[0186] Further, the personalized recommendations of implementations of the present disclosure can be more engaging for the user and also provide a more precise way to use wearable data to improve health compared with more standard advice (e.g., increase step counts). In addition, because users wear their devices continuously over long periods of time, the example implementation can provide the advantage of long-term risk tracking. Yet another advantage of the methods of the present disclosure is that the ridge model keeps theDocket Number: 11760-025W01 entire feature vector intact for each disease, rather than selecting specific variables that may differ between diseases. Using a common shared set of features, this method allows us to determine how changes in one variable may impact risk across all diseases when making personalized recommendations. Trade-offs can be important and improvements in one disease risk prediction may impact disease risk for other diseases. This is another methodological advancement that makes this method generalizable and valuable across a range of chronic diseases and conditions.

[0187] In some implementations, the composite risk score and associated feature-level contributions are used as control inputs to a recommendation engine configured to automatically select, prioritize, and generate personalized recommendations. The recommendation engine can map changes in the composite score and individual feature contributions to predefined recommendation rules or generation parameters, such that increases or decreases in specific components of the composite score dynamically modify the type, timing, frequency, or content of recommendations provided to the user. In this manner, the composite score can actively govern downstream recommendation logic without manual intervention. As the composite score and feature contributions are updated over time based on newly received wearable data, the recommendation engine can automatically adapt the personalized recommendations to reflect evolving disease-risk drivers. Accordingly, the composite score enables coordinated recommendation generation across multiple diseases by allowing the system to computationally evaluate trade-offs among feature-level risk contributors when selecting or adjusting personalized recommendations.

[0188] Other methods and data sources

[0189] While the methods and results described above provide a case study in using wearable or other data to predict disease risk, they are only a non-limiting example of the present disclosure. The present disclosure contemplates that there are alternativeDocket Number: 11760-025W01 implementations. The example implementation of the study represents a non-limiting example approach to predicting disease risk in a way that allows for personalized recommendations to help people reduce their risks of developing a disease or group of diseases. But other implementations of the present disclosure include systems that can be generalized by having a processor (e.g., the computing device of FIG. 3) execute commands that choose variables from a device, set of devices, or other sources that describe a person's behavior, physiological status, and clinical / personal characteristics.

[0190] For example, in some implementations, these methods can be implemented using a combination of wearable hardware and computing systems. For example, data can be collected from one or more wearable devices equipped with motion sensors such as tri-axial accelerometers or inertial measurement units, as well as physiological sensors that measure signals like heart rate or body temperature. These devices can include onboard processors and memory to store and preprocess data, and they communicate wirelessly with a user device or remote server. The collected data can be transmitted to a computing system, such as a smartphone, tablet, or cloud-based server, where the machine-learning models described herein are used to generate composite risk scores. In some cases, parts of this processing can occur directly on the wearable or user device to allow faster feedback, for example using one or more remote computing devices as described herein. The resulting risk scores and related outputs can then be used to automatically drive downstream functions, including the generation and delivery of personalized recommendations to the user. The choice of variables can come from any kind of variable selection method that can include, but is not limited to, regression models, ridge regression models, lasso regression models, elastic net models, neural networks, convolutional neural networks, XGboost, or other methods that select either variables from a set or that apply coefficients to the full or partial set of variables to predict disease. The coefficient valuesDocket Number: 11760-025W01 provided here are one possible set of values which may be the exact numbers or may be these coefficients plus or minus some difference (e.g., plus or minus 10%).

[0191] The models according to implementations of the present disclosure can be built from motion or other behavioral data that can come from accelerometers, IMUs, Wi¬ Fi sensing, GPS, video, or other ways to capture an individual's activity. Thus, while wearable devices were used here, motion data can, for example, come from mobile phones or from external monitoring devices that capture an individual's movement. Motion data can also come from accelerometers or other motion analysis devices (e.g. IMUs) that can be clipped to a watchband, wristband, waistband, hatband, shoe, or any other item attached to the body. The devices can also stick to the body with adhesive. The model can be trained from large datasets, including the UK Biobank or other similar datasets (including but not limited to NHANES, All of Us, Whitehall, ProPass or others), or can be trained from data collected from users of this application. The training data can also be refined by combining data from large epidemiological datasets like the UK Biobank with data collected by users of this application. In addition, data collected for model training in either epidemiological datasets or from users of this or other commercial applications may come from short durations of just a single day to longer durations, including but not limited to one year (e.g., All of Us data).

[0192] Conventional studies have focused on capturing only a snapshot of objective behavior with accelerometers typically worn over a period from days to weeks at a time; however, with people wearing these devices for long extended periods in the community, there is potential to greatly expand the assessment, providing a unique opportunity to give ongoing integrated feedback and recommendations to the user and potentially their health professionals to enhance quality of life and reduce disease risk. In this way, some disease or health condition outcomes may be best predicted by shorter timeDocket Number: 11760-025W01 durations and some outcomes may be better predicted by longer time durations of data collection.

[0193] Based on trained models, this method allows implementations of the present disclosure to predict multiple disease outcomes using a single method, or predict multiple disease outcomes using a model trained on just one disease, condition, or health outcome including but not limited to mortality. Different diseases may have different variable coefficients or may have different combinations of variables.

[0194] Finally, as described in more detail above, the training data set can include other forms of data in addition to or in lieu of motion data. These types of data include, but are not limited to, heart rate data, blood pressure, polysomnography, environmental exposures, electroencephalography / event-related potentials (EEG / ERP), blood biomarkers, urinary biomarkers, audio data (e.g., speech), electronic medical records, questionnaires (e.g., demographics, behavioral rating scales), cognitive tests, physical function tests, and other data streams, as illustrated in FIGS. 4A-4B. In this way, this application is not limited to a single data type or stream.

[0195] Discussion

[0196] Chronic diseases account for a large economic burden in the US. Lifestyle modifications are a key element for reducing risk of a wide array of chronic diseases from heart disease to cancer, however existing systems and methods lack a method to provide continuous disease risk assessments to individuals that would motivate and reward behavioral change. The growing ecosystem of wearable technologies that track and monitor an array of physiological systems presents a unique opportunity to develop metrics can be presented to users that will allow them to monitor disease risk and determine which lifestyle modifications provide the best risk reduction strategies. This enabling of personalized disease risk tracking represents an important advancement in the use of wearables-based movementDocket Number: 11760-025W01 and physiological data and has the promise of changing the way users approach preventive health and precision medicine as shown in FIGS. 4A and 4B.

[0197] The example implementation uses physical activity and movement data here because they are accessible from wearable devices and are linked with electronic health records in large epidemiological datasets. Physical activity is a key element of a healthy lifestyle and higher levels of PA engagement are linked with protection against a suite of chronic diseases. Sedentary Behaviors, in contrast, are the most common waking behavior in adult humans, and are linked with increased risk of chronic disease across the lifespan. Early work in physical activity and inactivity epidemiology focused on self-reported PA and SB from questionnaires, however more recent studies have used wearable accelerometers that are able to more objectively measure lifestyle behaviors in relation to disease risk.

[0198] While individual metrics relating to accelerometer-derived volume and intensity of PA and volume of SB are often used in epidemiological research, fewer studies have examined the richness of the underlying accelerometer signal which may contain more information regarding how patterns of activity relate to disease risk. Recent work has demonstrated that combinations of a variety of PA metrics can show strong associations with health and disease outcomes including for all-cause mortality, cardiometabolic health, Parkinson's disease, depression, and cognition. Researchers have used many approaches for selection of accelerometer metrics to use in models, and combinations of these measures may provide stronger associations with health outcomes compared to more traditional PA metrics like relying solely on step counts or time spent in moderate to vigorous PA intensities (MVP A). However, to date, no study has examined whether the same methodological strategy can be effective across a range of disease outcomes. In addition, previous work has mainly focused on combinations of standard PA metrics including time spent in various intensity levels. One key gap in the literature is the inclusion of more complex PA pattern-Docket Number: 11760-025W01 related metrics that better capture overall movement activities. Composite PA scores that include a wide range of metrics covering intensity, duration, and patterns of PA may provide valuable early markers of disease risk in both clinical and research settings. The goal of this study is to apply a machine learning approach to wearable accelerometer data to examine the utility of combining a wide array of accelerometer metrics to create more precise PA-related risk scores specific to a range of health conditions and chronic diseases.

[0199] In addition, composite PA risk scores for chronic disease can provide key information to consumers. For example, common PA-related metrics like daily step counts can serve as useful goals for individuals motivated to increase their PA levels.However, for individuals interested in improving their PA-related health outcomes, there are currently no metrics available at the consumer level. A significant barrier to behavior change, when linked with long-term goals such as disease prevention, is the lack of easily tracked metrics that can inform users if they are meeting their goals. Self-monitoring is a crucial component of behavior change interventions, especially for increasing or maintaining PA or reducing SB. However, there is currently no widely available technique to self-monitor disease risk using data from wearables on an individual basis. Thus, a PA metric that can be easily derived from commonly used wearable accelerometers may have value to consumers interested in tracking their personal risk for a range of health conditions and chronic diseases.

[0200] The example implementation includes an improved method of tracking PA-related health risk using a machine learning technique that can optionally be implemented using consumer wearable devices. The example implementation includes models for a range of common, often debilitating, and economically costly chronic health conditions and diseases affecting the United States including heart disease, diabetes, dementia, Parkinson's disease, and cancer. After demonstrating the potential of this method based on accelerometer data, the study further enables incorporating a wider array of behavioral and physiologicalDocket Number: 11760-025W01 data from wearables and from clinical sources, including those illustrated in FIG. 4A and 4B. It should be understood that the present disclosure is not limited to the particular implementation described in the study, but rather includes implementations that utilize information from a range of sources (devices, medical records, questionnaires) and uses a machine learning approach to determine the best combination of this information to predict disease risk. Using these predictions, implementation of the present disclosure enable users to make behavioral, lifestyle, or interventional changes (e.g., medication or other clinical intervention) to delay or prevent a wide range of chronic conditions or diseases.

[0201] The results of the study described herein show that the composite PA measure determined using implementations of the present disclosure based on accelerometer derived metrics can outperform traditional individual PA measures in predicting incident chronic disease and health conditions. Models that include waking behavior can perform as well as models that include waking and sleep behaviors for this set of chronic diseases and conditions. These metrics can be calculated from consumer-grade devices allowing individuals with smart watches to gain access to a digital marker of disease risk. Providing individuals with information relating to chronic disease / condition risk from noninvasive easily collected wearable data may be a powerful incentive for behavior change and a motivation to maintain a healthier PA profile. Incorporation of additional physiological, demographic, behavioral, and environmental data will further strengthen these risk assessments.

[0202] Example 2:

[0203] An example implementation of the present disclosure was studied.

[0204] To evaluate the generalizability of the example implementation of the present disclosure, including the composite physical activity (PA) score described herein, a second validation analysis was performed using data from the National Health and NutritionDocket Number: 11760-025W01 Examination Survey (NHANES), which is designed to be nationally representative of the U. S. population. Table IB, below, illustrates sample demographics. The analytic sample included 9,439 adults with valid mortality follow-up (n= 830 deaths; mean follow-up time of 6.7 years). The study computed composite PA scores in NHANES using the previously derived mortality prediction coefficients from the UK Biobank models and compared their performance to minutes of moderate -to-vigorous physical activity (MVP A) and daily step counts. Cox proportional hazards models were adjusted for age, sex, ethnicity, education, household income, and body mass index.

[0205] FIG. 7A illustrates hazard ratios for the composite score, MVPA, and step counts. FIG. 7B illustrates statistics for composite score, MVPA, and step counts Models are adjusted for age, sex, ethnicity, education, household income, BMI.

[0206] As shown in FIGS. 7A-7B, all three predictors were significantly associated with mortality risk, but the composite score demonstrated the strongest association (HR = 0.025; 95% CI: 0.020-0.032; p = 2.8 x 10-199), compared to MVPA (HR = 0.153; 95% Cl: 0.121-0.193; p = 2.4 x 10-56) and steps (HR = 0.333; 95% CI: 0.281-0.396; p = 1.2 x 10-35)- Discriminative performance was assessed using Harrell’s C with 1,000 bootstrap replications. The composite score again showed the highest concordance (C = 0.857; 95% CI: 0.844-0.870), followed by steps (C = 0.851; 95% CI: O.838-O.865) and MVPA (C = 0.850; 95% CI: 0.836-0.863). The composite Score (C ~ 0.857; HR = 0.025) outperforms MVPA and steps in discrimination and effect size.

[0207] To formally compare predictive performance, we used paired nonparametric bootstrap tests of the difference in Harrell’s C between models. The composite score significantly outperformed MVPA (AC = 0.0129; 95% CI: 0.0059-0.0207) and steps (AC = 0.0662; 95%’ CI: 0.0523-0.0796), and MVPA significantly outperformed steps (AC = 0.0537; 95% CI: 0.0395-0.0675). Tables 2B-4B provide detailed results. Together, theseDocket Number: 11760-025WO1findings show that the composite metric not only retains predictive strength in an independent, nationally representative dataset, but also, similar to the original analysis, it outperforms commonly used physical activity indicators in both strength of association and risk discrimination.

[0208] Table IB: Demographics for the NHANES sampleVariable Category ValueAge Mean (SD) 47.9 (18.4)Sex Male 4534 (48.0%)Sex Female 4905 (52.0%)Ethnicity Mexican American 1136 (12.0%)Ethnicity Other Hispanic 908 (9.6%)Ethnicity Non-Hispanic White 3759 (39.8%)Ethnicity Non-Hispanic Black 2244 (23.8%)Ethnicity Other Race — Multi 1392 (14.7%)Education Less than 9th grade 747 (8.3%)Education 9-11th grade 1226 (13.7%)Education High school grad / GED 1954 (21.8%)Education Some college or AA degree 2746 (30.7%)Education College graduate or above 2277 (25.4%)

[0209] Table 2B: Hazard Ratios for NHANES analysispredictor HR Cljower CI_upper p_ valuePA..comp...scaled 0.0254 0.02 0.0322 2.84E-199 MVPA__scaled 0.153 0.121 0.193 2.37E-56steps_scaled 0.333 0.281 0.396 1.19E-35

[0210] Table 3B: C-stats for N HANES analysispredictor C...boot C...boot... CI..lower C..boot.. CI...upper boot... R PA_comp_scaled 0.857 0.844 0.87 1000 MVPA_scaled 0.85 0.836 0.863 1000steps...scaled 0.851 0.838 0.865 1000

[0211] Table 4B: Comparison of C-stats for NHANES analysis predictor_l predictor_2 mean_diff Cljower CI_upper boot_R PA_comp_scaled MVPA_scaled 0.0129 0.00587 0.0207 1000 PA...comp.. scaled steps...scaled 0.0662 0.0523 0.0796 1000MVPA_scaIed steps_scaled 0.0537 0.0395 0.0675 1000Docket Number: 11760-025W01

[0212] Note: Paired nonparametric bootstrap test of the difference in Harrell’s C for two Cox models. PA composite is significantly better than other predictors though differences are small.

[0213] Although the subject matter has been described in language specific to structural features and / or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing the claims.

Claims

1. Docket Number: 11760-025W01WHAT IS CLAIMED:

1. A method comprising:receiving physical activity data for a subject, wherein the physical activity data comprises raw acceleration data;deriving a plurality of metrics from the raw acceleration data, wherein the plurality of metrics comprise a plurality of pattern-related metrics;inputting the plurality of metrics into a deployed machine learning model; receiving, from the deployed machine learning model, a physical activity score; and calculating a risk score for at least one disease or condition based on the physical activity score.

2. The method of claim 1, wherein the plurality of pattern-related metrics comprise wavelet energy features.

3. The method of claim 2, wherein the wavelet energy features comprise a wavelet pattern variability index (WPVI), the WPVI quantifying a variability in daily physical activity patterns.

4. The method of any one of claims 1-3, wherein the plurality of pattern-related metrics further comprise fractal complexity features.

5. The method of claim 4, wherein the fractal complexity features are calculated using a detrended fluctuation analysis.

6. The method of any one of claims 1-5, wherein the plurality of pattern-related metrics are related to periods of activity or inactivity.

7. The method of claim 6, wherein the plurality of pattern-related metrics comprise a variability measure or a stability measure.

8. The method of any one of claims 1-7, wherein the plurality of metrics further comprise volume or intensity features.Docket Number: 11760-025W019. The method of any one of claims 1-8, wherein the deployed machine learning model is trained on a dataset associated with a specific disease or condition.

10. The method of claim 1, wherein the at least one disease or condition is a plurality of diseases or conditions.

11. The method of claim 1, wherein the at least one disease or condition is a plurality of diseases or conditions, and wherein the step of calculating the risk score comprises calculating a respective risk score for each of the plurality of diseases or conditions based on the physical activity score.

12. The method of claim 1, wherein the at least one disease or condition is a plurality of diseases or conditions, wherein the step of receiving, from the deployed machine learning model, the physical activity score comprises receiving a plurality of physical activity scores, and wherein the step of calculating the risk score comprises calculating a respective risk score for each of the plurality of diseases or conditions based on the plurality of physical activity scores.

13. The method of any one of claims 10-12, wherein the deployed machine learning model is trained on a dataset associated with the plurality of diseases or conditions.

14. The method of any one of claims 1-13, wherein the raw acceleration data is measured by a wearable device.

15. The method of claim 14, wherein the raw acceleration data is received from the wearable device.

16. The method of claim 14, wherein the wearable device comprises a tri-axial accelerometer.

17. The method of any one of claims 1-16, further comprising:receiving physiological data for the subject; andDocket Number: 11760-025W01deriving one or more physiological metrics from the physiological data, wherein the plurality of metrics and the one or more physiological metrics are input into the deployed machine learning model.

18. The method of claim 17, wherein the physiological data comprises one or more of sleep data, heart rate data, electrocardiography data, electromyography data, blood pressure data, blood oxygen saturation data, and blood glucose data.

19. The method of any one of claims 1-18, further comprising:receiving clinical data for the subject; andderiving one or more clinical metrics from the clinical data, wherein the plurality of metrics and the one or more clinical metrics are input into the deployed machine learning model.

20. The method of any one of claims 1-19, further comprising:receiving test data for the subject; andderiving one or more test metrics from the test data, wherein the plurality of metrics and the one or more test metrics are input into the deployed machine learning model.

21. The method of any one of claims 1-20, wherein the deployed machine learning model is a supervised machine learning model.

22. The method of claim 21, wherein the supervised machine learning model is a regression model, a ridge regression model, a lasso regression model, an elastic net model, or an artificial neural network.

23. The method of any one of claims 1-22, wherein the at least one disease or condition is a cardiovascular disease or condition, a chronic disease or condition, a neurological disease or condition, a metabolic disease or condition, a mental disorder or condition, or cancer.

24. The method of any one of claims 1-23, further comprising providing a personalized recommendation for the subject based, at least in part, on the risk score for the at least one disease or condition in the subject.Docket Number: 11760-025W0125. The method of claim 24, wherein the personalized recommendation reduces or improves a risk of the at least one disease or condition.

26. A method comprising:assessing the risk score for the at least one disease or condition according to any one of claims 1-25; andtreating the at least one disease or condition based on the risk score.

27. A computing device comprising:at least one processor and memory operably coupled to the at least one processor, the memory having computer-executable instructions stored thereon that, when executed by the at least one processor, cause the at least one processor to:receive physical activity data for a subject, wherein the physical activity data comprises raw acceleration data;derive a plurality of metrics from the raw acceleration data, wherein the plurality of metrics comprise a plurality of pattern-related metrics;input the plurality of metrics into a deployed machine learning model;receive, from the deployed machine learning model, a physical activity score; and calculate a risk score for at least one disease or condition based on the physical activity score.

28. The computing device of claim 27, wherein the plurality of pattern-related metrics comprise wavelet energy features.

29. The computing device of claim 28, wherein the wavelet energy features comprise a wavelet pattern variability index (WPVI), the WPVI quantifying a variability in daily physical activity patterns.

30. The computing device of any one of claims 27-29, wherein the plurality of pattern-related metrics further comprise fractal complexity features.

31. The computing device of claim 30, wherein the fractal complexity features are calculated using a detrended fluctuation analysis.Docket Number: 11760-025W0132. The computing device of any one of claims 27-31, wherein the plurality of pattern -related metrics are related to periods of activity or inactivity.

33. The computing device of claim 32, wherein the plurality of pattern-related metrics comprise a variability measure or a stability measure.

34. The computing device of any one of claims 27-33, wherein the plurality of metrics further comprise volume or intensity features.

35. The computing device of any one of claims 27-34, wherein the deployed machine learning model is a supervised machine learning model.

36. The computing device of claim 35, wherein the supervised machine learning model is a regression model, a ridge regression model, a lasso regression model, an elastic net model, or an artificial neural network.

37. The computing device of any one of claims 27-36, wherein the memory has further computer-executable instructions stored thereon that, when executed by the at least one processor, cause the at least one processor to provide a personalized recommendation for the subject based, at least in part, on the risk score for the at least one disease or condition in the subject.

38. An electronic device comprising:a tri-axial accelerometer; anda computing device operably coupled to the tri-axial accelerometer, the computing device comprising at least one processor and memory operably coupled to the at least one processor, the memory having computer-executable instructions stored thereon that, when executed by the at least one processor, cause the at least one processor to:receive physical activity data for a subject, wherein the physical activity data comprises raw acceleration data measured by the tri-axial accelerometer;derive a plurality of metrics from the raw acceleration data, wherein the plurality of metrics comprise a plurality of pattern-related metrics;input the plurality of metrics into a deployed machine learning model;Docket Number: 11760-025W01receive, from the deployed machine learning model, a physical activity score; and alert the subject of a risk score for at least one disease or condition, wherein the risk score for at least one disease or condition is calculated based on the physical activity score.

39. The electronic device of claim 38, wherein the electronic device is a wearable device.

40. The electronic device of claim 38 or 39, wherein the plurality of pattern- related metrics comprise wavelet energy features.

41. The electronic device of claim 40, wherein the wavelet energy features comprise a wavelet pattern variability index (WPVI), the WPVI quantifying a variability in daily physical activity patterns.

42. The electronic device of any one of claims 38-41, wherein the plurality of pattern -related metrics further comprise fractal complexity features.

43. The electronic device of claim 42, wherein the fractal complexity features are calculated using a detrended fluctuation analysis.

44. The electronic device of any one of claims 38-43, wherein the plurality of pattern -related metrics are related to periods of activity or inactivity.

45. The electronic device of claim 44, wherein the plurality of pattern-related metrics comprise a variability measure or a stability measure.

46. The electronic device of any one of claims 38-45, wherein the plurality of metrics further comprise volume or intensity features.

47. The electronic device of any one of claims 38-46, wherein the deployed machine learning model is a supervised machine learning model.Docket Number: 11760-025W0148. The electronic device of claim 47, wherein the supervised machine learning model is a regression model, a ridge regression model, a lasso regression model, an elastic net model, or an artificial neural network.

49. The electronic device of any one of claims 38-48, wherein the memory has further computer-executable instructions stored thereon that, when executed by the at least one processor, cause the at least one processor to provide a personalized recommendation for the subject based, at least in part, on the risk score for the at least one disease or condition in the subject.

50. A method comprising:receiving physical activity data for a subject, wherein the physical activity data comprises raw acceleration data:deriving a plurality of metrics from the raw acceleration data, wherein the plurality of metrics comprise a plurality of pattern-related metrics;inputting the plurality of metrics into a deployed machine learning model; receiving, from the deployed machine learning model, a physical activity score; and calculating a risk score for mortality for at least one disease or condition within one or more specified ranges of time based on the physical activity score.

51. A method comprising:receiving physical activity data for a subject, wherein the physical activity data comprises raw acceleration data;deriving a plurality of metrics from the raw acceleration data, wherein the plurality of metrics comprise a plurality of pattern-related metrics; andcalculating a risk score for at least one disease or condition based on the plurality of metrics.

52. A method comprising:receiving physical activity data for a subject, wherein the physical activity data comprises raw acceleration data;deriving a plurality of metrics from the raw acceleration data, wherein the plurality of metrics comprise a plurality of pattern-related metrics;Docket Number: 11760-025W01inputting the plurality of metrics into a deployed machine learning model; receiving, from the deployed machine learning model, a physical activity score; calculating a risk score for at least one disease or condition based on the physical activity score; anddetermining, based on the risk score, a diagnostic indicator.

53. The method of claim 52, wherein the diagnostic indicator comprises a likelihood of a presence or absence of a disease or condition.

54. The method of claim 52, wherein the diagnostic indicator comprises a classification of the risk score as a disease or condition.

55. A method comprising:receiving physical activity data for a subject, wherein the physical activity data comprises raw acceleration data;deriving a plurality of metrics from the raw acceleration data, wherein the plurality of metrics comprise a plurality of pattern-related metrics;inputting the plurality of metrics into a deployed machine learning model; receiving, from the deployed machine learning model, a physical activity score; determining, using a large language model, a recommendation output based on the physical activity score.

56. The method of claim 55, wherein the physical activity score is a composite of a plurality of scores, and wherein the recommendation output is based on at least one score of the plurality of scores.

57. The method of claim 55 or 56, wherein the recommendation output is output to a user interface for display.

Images