Deep learning based systems and methods for disease risk phenotyping
Deep learning-based phenotyping models integrate diverse health data to enhance disease risk prediction accuracy, addressing the limitations of traditional models by providing absolute risk estimates and enabling tailored interventions.
Patent Information
- Application Number
- PCT/US2026/027326
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-09
- Filing Date
- 2026-05-08
- Publication Date
- 2026-11-12
AI Technical Summary
Existing disease risk models are inaccurate due to their reliance on simple statistical techniques and failure to leverage comprehensive health data, including genetic and unstructured clinical information, leading to false positives and negatives, which can result in unnecessary tests and late detections.
Deep learning-based phenotyping models that integrate neural networks, including autoencoders and classifiers, to generate unified representations of structured and unstructured health data, using polygenic risk scores and electronic health records, and apply hybrid quantum-classical optimization to improve disease risk estimation.
Enhances disease risk prediction accuracy with improved recall and specificity, enabling tailored screening and interventions based on absolute risk estimates, capturing complex interactions and adapting to genetic discoveries.
Smart Images

Figure US2026027326_12112026_PF_FP_ABST
Abstract
Description
Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTDEEP LEARNING BASED SYSTEMS AND METHODS FOR DISEASE RISK PHENOTYPINGCROSS-REFERENCE TO RELATED APPLICATIONS
[0001] This international application claims priority and benefit of U.S. Provisional Patent Application No. 63 / 803,507, titled “DEEP LEARNING BASED SYSTEMS AND METHODS FOR DISEASE RISK PHENOTYPING” and filed on May 9, 2025, the contents of which are hereby incorporated by reference herein to the maximum extent permitted by applicable law.BACKGROUND
[0002] The present disclosure generally relates to determining disease risk, and more specifically relates to methods for determining disease risk to facilitate early detection of disease. The identification of a disease in its early stages, before it progresses and becomes more severe, can improve health outcomes, expand treatment options, reduce complications, and reduce healthcare costs. For example, early detection of cancer can substantially increase survivability.SUMMARY
[0003] In some aspects, the techniques described herein relate to a method for determining whether a subject is at increased risk for a disease, the method including: generating, by one or more processors, an integrated risk score (IRS) for a subject, wherein the IRS indicates a risk of a particular disease or disease group for the subject, and wherein generating the IRS includes (1) extracting, by one or more first neural network models, values of a plurality of features from health data of the subject, and (2) classifying, by one or more second models and based on the values of the plurality of features, the risk of the particular disease or disease group for the subject; determining one or more recommended actions for the subject based at least in part on the IRS; and providing an indication of the one or more recommended actions for the subject.
[0004] In some aspects, the techniques described herein relate to a method, wherein the one or more first neural network models include an autoencoder, and wherein the one or more second models include a classifier.
[0005] In some aspects, the techniques described herein relate to a method, wherein the autoencoder and the classifier are jointly trained based on training data including, for each individual of a plurality of individuals, (1) health data of the respective individual and (2) label data indicating the risk of the particular disease or disease group for the individual.
[0006] In some aspects, the techniques described herein relate to a method, wherein the IRS for the subject is a first IRS for the subject, wherein the particular disease or disease group is aAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTfirst particular disease or disease group, wherein the plurality of features is a plurality of first features, wherein the health data of the subject are first health data of the subject, wherein the one or more recommended actions are one or more first recommended actions, wherein the autoencoder is a first autoencoder, wherein the classifier is a first classifier, and wherein the method further includes: generating, by one or more processors, a second integrated risk score (IRS) for the subject, wherein the second IRS indicates a risk of a second particular disease or disease group for the subject, and wherein generating the second IRS includes (1) extracting, by a second autoencoder, values of a plurality of second features from second health data of the subject, and (2) classifying, by a second classifier and based on the values of the plurality of second features, the risk of the second particular disease or disease group for the subject; determining one or more second recommended actions for the subject based at least in part on the second IRS; and providing an indication of the one or more second recommended actions for the subject.
[0007] In some aspects, the techniques described herein relate to a method, wherein the IRS for the subject is a first IRS for the subject, wherein the particular disease or disease group is a first particular disease or disease group, wherein the plurality of features is a plurality of first features, wherein the one or more recommended actions are one or more first recommended actions, and wherein the method further includes: generating, by one or more processors, a second integrated risk score (IRS) for the subject, wherein the second IRS indicates a risk of a second particular disease or disease group for the subject, and wherein generating the second IRS includes (1) extracting, by the autoencoder, values of a plurality of second features from the health data of the subject, and (2) classifying, by the classifier and based on the values of the plurality of second features, the risk of the second particular disease or disease group for the subject; determining one or more second recommended actions for the subject based at least in part on the second IRS; and providing an indication of the one or more second recommended actions for the subject.
[0008] In some aspects, the techniques described herein relate to a method, wherein an encoder portion of the autoencoder is an electronic health record (EHR) encoder.
[0009] In some aspects, the techniques described herein relate to a method, wherein the EHR encoder includes a transformer.
[0010] In some aspects, the techniques described herein relate to a method, wherein the risk indicated by the IRS is a first risk, and wherein the health data include a polygenic risk score (PRS) for the subject, wherein the PRS is based on a genotype of the subject and indicates a second risk of the particular disease or disease group for the subject.Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT
[0011] In some aspects, the techniques described herein relate to a method, further including generating the PRS for the subject, wherein generating the PRS for the subject includes applying a polygenic model to the genotype of the subject, wherein the polygenic model is associated with the particular disease or disease group.
[0012] In some aspects, the techniques described herein relate to a method, wherein the health data include values of one or more biomarkers for the subject, wherein the one or more biomarkers are associated with the particular disease or disease group.
[0013] In some aspects, the techniques described herein relate to a method, wherein the health data further include principal components of ethnicity of the subject.
[0014] In some aspects, the techniques described herein relate to a method, wherein the health data further include structured data derived from one or more electronic health records of the subject.
[0015] In some aspects, the techniques described herein relate to a method, wherein the structured data include a plurality of fields, and wherein values of the plurality of fields are one-hot encoded.
[0016] In some aspects, the techniques described herein relate to a method, further including generating, by one or more generative models, a portion of the structured data based on unstructured data in the one or more electronic health records.
[0017] In some aspects, the techniques described herein relate to a method, wherein the one or more generative models include a large language model (LLM).
[0018] In some aspects, the techniques described herein relate to a method, wherein the health data include one or more diagnostic codes indicating diagnoses of diagnosed health conditions of the subject; treatment records of the subject; one or more procedure codes indicating procedures administered to the subject; demographic data of the subject; admission, discharge and transfer (ADT) records of the subject, results of laboratory tests administered to the subject, vital signs of the subject, diagnostic data related to diagnosed and / or undiagnosed health conditions of the subject; encounter data related to encounters between the subject and one or more healthcare providers; clinical records related to health or treatment of the subject; allergy data of the subject; surgery data of the subject; immunization data of the subject; a family medical history of the subject; a social history of the subject; one or more radiology reports corresponding to radiological scans of the subject; one or more medical images of the subject; medication data of the subject; results of one or more cardiac health tests of the subject; microbiology data obtained from analysis of one or more samples of blood, urine, sputum, or tissue of the subject; pathology data obtained from analysis of one or more samples of blood, urine, sputum, or tissue of theAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTsubject; infusion data related to one or more substances infused into a bloodstream of the subject; and / or treatment data related to one or more treatments administered to the subject.
[0019] In some aspects, the techniques described herein relate to a method, wherein the health data further include physiological monitoring data obtained from one or more wearable sensors, wherein the physiological monitoring data include heart rate data, blood oxygen saturation data, activity data, sleep data, electrodermal activity data, and / or continuous glucose data.
[0020] In some aspects, the techniques described herein relate to a method, wherein the health data further include medical imaging data.
[0021] In some aspects, the techniques described herein relate to a method, wherein the particular disease is breast cancer, lung cancer, prostate cancer, colorectal cancer, pancreatic cancer, ovarian cancer, endometrial cancer, cervical cancer, liver cancer, gastroesophageal cancer, small intestine cancer, bladder cancer, brain cancer, kidney cancer, thyroid cancer, leukemia, multiple myeloma, lymphoma (both Hodgkin and non-Hodgkin), melanoma, head and neck cancer, obesity, coronary artery disease (CAD), type 2 diabetes (T2D), stroke, Alzheimer’s disease (AD), glaucoma, or age-related macular degeneration (AMD).
[0022] In some aspects, the techniques described herein relate to a method, wherein the one or more first neural network models include: an encoder configured to provide first embeddings of first features of the plurality of features based on a first portion of the health data of the subject; one or more subnetworks configured to provide second embeddings of second features of the plurality of features based on respective second portions of the health data of the subject; and a fusion layer configured to provide a unified representation of the plurality of features based on the first embeddings and the second embeddings.
[0023] In some aspects, the techniques described herein relate to a method, wherein classifying of the risk by the one or more second models is based on the unified representation of the plurality of features.
[0024] In some aspects, the techniques described herein relate to a method, wherein the encoder includes a portion of an autoencoder or an EHR encoder.
[0025] In some aspects, the techniques described herein relate to a method, wherein the encoder, the one or more subnetworks, the fusion layer, and the one or more second models are arranged in a fusion neural network, wherein the first portion of the health data include structured data, wherein at least one second portion of the health data includes unstructured text of electronic health records of the subject, wherein the second features include one or more unstructured text features based on the unstructured text of electronic health records of the subject, wherein a first subnetwork of the one or more subnetworks includes one or moreAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTgenerative models, wherein the one or more second models include a classifier, and wherein the method further includes: generating, by the one or more generative models, the one or more unstructured text features; and providing, by the first subnetwork, one or more second embeddings of the one or more unstructured text features.
[0026] In some aspects, the techniques described herein relate to a method, wherein the second features further include a polygenic risk score (PRS) of the subject, a numeric biomarker value of the subject, one or more lab results of the subject, one or more lifestyle factors of the subject, and / or one or more environmental factors of the subject.
[0027] In some aspects, the techniques described herein relate to a method, wherein the first features include a polygenic risk score (PRS) of the subject, a numeric biomarker value of the subject, one or more lab results of the subject, one or more lifestyle factors of the subject, and / or one or more environmental factors of the subject.
[0028] In some aspects, the techniques described herein relate to a method, wherein the one or more generative models include a large language model (LLM) including a generative transformer-based network trained to embed free-text clinical notes.
[0029] In some aspects, the techniques described herein relate to a method, wherein the encoder and the one or more subnetworks are jointly trained in a joint optimization process.
[0030] In some aspects, the techniques described herein relate to a method, further including: applying, by the fusion layer, a gating mechanism or attention mechanism that weights contributions from the one or more second embeddings and contributions from the first embeddings differently based on domain context.
[0031] In some aspects, the techniques described herein relate to a method, wherein the joint optimization process includes a hybrid quantum-classical optimization process in which a quantum processor evaluates at least one quantum circuit associated with a parameter update, a loss function component, a feature interaction, an attention mechanism, or an embedding transformation.
[0032] In some aspects, the techniques described herein relate to a method, wherein measurement results from the quantum processor are used to update one or more parameters of the encoder, the one or more subnetworks, the fusion layer, or the classifier.
[0033] In some aspects, the techniques described herein relate to a method, wherein the encoder includes a transformer configured to model temporal, semantic, or clinical relationships among features extracted from electronic health records, and wherein the hybrid quantum-classical optimization process determines one or more attention weights, query projections, keyAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTprojections, value projections, temporal encodings, or feed-forward layer parameters of the transformer.
[0034] In some aspects, the techniques described herein relate to a method, wherein at least one of the one or more subnetworks is configured to generate a polygenic risk embedding based on genetic-risk data of the subject, and wherein generating the polygenic risk embedding includes applying a quantum-assisted model to genotype data, variant data, single nucleotide polymorphism data, variant effect sizes, ancestry-adjusted variant weights, disease-associated loci, one or more polygenic risk scores, or a combination thereof.
[0035] In some aspects, the techniques described herein relate to a method, wherein the quantum-assisted model is configured to estimate nonlinear interactions among genetic variants, interactions among polygenic risk components, interactions between genetic-risk features and clinical features, or interactions between genetic-risk features and environmental or lifestyle features.
[0036] In some aspects, the techniques described herein relate to a method, wherein the polygenic risk embedding is combined, by the fusion layer, with an embedding generated by a second encoder, wherein the second encoder is transformer-based and is configured to encode one or more electronic health records.
[0037] In some aspects, the techniques described herein relate to a method, wherein the fusion layer learns disease-specific interactions between longitudinal clinical features represented by the embedding generated by the second encoder and inherited genetic-risk features represented by the polygenic risk embedding.
[0038] In some aspects, the techniques described herein relate to a method, wherein the quantum processor includes a gate-based quantum processor, a quantum annealer, a photonic quantum processor, an ion-trap quantum processor, a superconducting quantum processor, or a simulated quantum processor.
[0039] In some aspects, the techniques described herein relate to a method, wherein the one or more recommended actions for the subject include undergoing a screening test, a blood test, lab work, and / or a medical imaging procedure.
[0040] In some aspects, the techniques described herein relate to a method, wherein generating the integrated risk score includes: determining a missingness characteristic of electronic health record data of the subject; determining, based at least in part on the missingness characteristic, one or more weights for genetic-risk data of the subject and electronic health record data of the subject; and generating the integrated risk score by combining a genetic-risk component derivedAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTfrom the genetic-risk data and a clinical-risk component derived from the electronic health record data according to the one or more weights.
[0041] In some aspects, the techniques described herein relate to a method, wherein the missingness characteristic includes an amount of missing electronic health record data, a pattern of missing electronic health record data, a type of missing electronic health record data, cohortspecific missingness, feature-specific missingness, or combinations thereof.
[0042] In some aspects, the techniques described herein relate to a method including: (a) obtaining first embeddings of unstructured health data of a subject, wherein obtaining the first embeddings includes applying a large language model (LLM) to at least a portion of the unstructured health data; (b) obtaining second embeddings of structured health data of the subject, wherein obtaining the second embeddings includes applying an encoder to at least a portion of the structured health data, and wherein the structured health data include at least one polygenic risk score (PRS) of the subject, wherein the encoder includes (1) a portion of an autoencoder or (2) a transformer; (c) combining the first embeddings and the second embeddings in a fusion layer of a deep neural network, thereby producing a unified representation indicative of a risk of a particular disease or disease group; (d) classifying, based on the unified representation, the risk of the particular disease or disease group for the subject; (e) identifying a therapy for the subject based at least in part on the classified risk; and (f) providing a recommendation that the therapy be administered to the subject to reduce the risk of the particular disease or disease group.
[0043] In some aspects, the techniques described herein relate to a computer-implemented method for training a disease risk model, the method including: obtaining training data including health data for a plurality of individuals and label data indicating disease labels for the plurality of individuals; generating, by one or more neural network models, a plurality of embeddings of the health data, wherein the health data include at least two types of health data, wherein the two types of health data include electronic health record data, unstructured clinical text, polygenic risk score data, biomarker data, medical imaging data, lifestyle data, environmental data, and / or principal components of ethnicity; combining the plurality of embeddings using a fusion layer to produce a plurality of unified representations for the plurality of individuals; evaluating, using a hybrid quantum-classical optimization process, a loss function associated with predictions generated based on the plurality of unified representations; updating one or more parameters of the one or more neural network models based at least in part on output of the hybrid quantum-classical optimization process; and storing the updated one or more parameters for subsequent use in generating a disease risk score for a subject.Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT
[0044] In some aspects, the techniques described herein relate to a method, wherein the disease labels indicate disease status, disease risk, and / or disease outcome for one or more diseases for the plurality of individuals.
[0045] In some aspects, the techniques described herein relate to a system for determining whether a subject is at increased risk for a disease, the system including: one or more processors; and a computer-readable storage medium storing instructions that, when executed by the one or more processors, cause the system to perform operations including: (a) obtaining first embeddings of unstructured health data of the subject, wherein obtaining the first embeddings includes applying a large language model (LLM) to at least a portion of the unstructured health data; (b) obtaining second embeddings of structured health data of the subject, wherein obtaining the second embeddings includes applying one or more encoders to one or more portions of the structured health data; (c) obtaining third embeddings of medical image data of the subject, wherein obtaining the third embeddings includes extracting, by a first neural network, one or more features from the medical image data and encoding, by an encoder, the one or more features, thereby producing the third embeddings; (d) combining the first embeddings, the second embeddings, and the third embeddings, in a fusion layer of a deep neural network, thereby producing a unified representation indicative of a risk of a particular disease or disease group; (e) classifying, based on the unified representation, the risk of the particular disease or disease group for the subject; (f) determining one or more recommended actions for the subject based at least in part on the classified risk; and (g) providing an indication of the one or more recommended actions.
[0046] In some aspects, the techniques described herein relate to a system, wherein the structured health data include a polygenic risk score (PRS) of the subject, a numeric biomarker value of the subject, one or more lab results of the subject, one or more lifestyle factors of the subject, and / or one or more environmental factors of the subject.
[0047] In some aspects, the techniques described herein relate to a system, wherein the one or more lab results include microbiology data obtained from analysis of a sample of blood, urine, sputum, or tissue of the subject, and wherein obtaining the microbiology data includes: obtaining the sample of blood, urine, sputum, or tissue of the subject; and analyzing the sample of blood, urine, sputum, or tissue of the subject.
[0048] In some aspects, the techniques described herein relate to a system, wherein providing the indication of the one or more recommended actions includes providing a recommendation that a therapy be administered to the subject to reduce the risk of the particular disease or disease group.Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT
[0049] In some aspects, the techniques described herein relate to a system, wherein the first neural network includes a convolutional neural network (CNN), a vision transformer (ViT), or a pretrained image foundation model.
[0050] In some aspects, the techniques described herein relate to a method including: (a) obtaining first embeddings of unstructured health data of a subject, wherein obtaining the first embeddings includes applying a large language model (LLM) to at least a portion of the unstructured health data; (b) obtaining second embeddings of second health data of the subject, wherein obtaining the second embeddings includes applying an encoder to at least a portion of the second health data, wherein the second health data include structured health data, and wherein the structured health data include at least one polygenic risk score (PRS) of the subject; (c) combining the first embeddings and the second embeddings in a fusion layer of a deep neural network, thereby producing a unified representation indicative of a risk of a particular disease or disease group; (d) classifying, based on the unified representation, the risk of the particular disease or disease group for the subject; (e) determining one or more recommended actions for the subject based at least in part on the classified risk; and (f) providing an indication of the one or more recommended actions, wherein the encoder includes (1) a portion of an autoencoder or (2) a transformer.
[0051] In some aspects, the techniques described herein relate to a method, wherein applying the LLM to at least the portion of the unstructured health data includes prompting the LLM to extract one or more first features from at least the portion of the unstructured health data.
[0052] In some aspects, the techniques described herein relate to a method, wherein the LLM, encoder, and one or more subnetworks of the deep neural network are jointly trained or finetuned in a joint optimization process.
[0053] In some aspects, the techniques described herein relate to a method, wherein the joint optimization process minimizes both reconstruction error and classification loss.
[0054] In some aspects, the techniques described herein relate to a method, wherein the fusion layer applies a gating mechanism or attention mechanism that weights contributions from the first embeddings and the second embeddings differently based on domain context.
[0055] In some aspects, the techniques described herein relate to a method, wherein the LLM is a generative transformer-based network configured to embed free-text clinical notes, and wherein the encoder is configured to embed one or more numeric biomarker values of the subject, one or more lab results of the subject, one or more lifestyle factors of the subject, one or more environmental factors of the subject, and / or the at least one polygenic risk score (PRS) for the subject.Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT
[0056] In some aspects, the techniques described herein relate to a method, wherein the encoder includes a portion of an autoencoder.
[0057] In some aspects, the techniques described herein relate to a method, wherein the encoder is an electronic health record (EHR) encoder.
[0058] In some aspects, the techniques described herein relate to a method, wherein the encoder includes a transformer.
[0059] In some aspects, the techniques described herein relate to a method, wherein the second health data further include medical imaging data.
[0060] In some aspects, the techniques described herein relate to a system including: one or more processors; and a computer-readable storage medium storing instructions that, when executed by the one or more processors, cause the system to perform any of the above-described methods.
[0061] In some aspects, the techniques described herein relate to a computer program product including at least one non-transitory computer-readable storage medium storing instructions that, when executed, cause a system to perform any of the above-described methods.
[0062] The foregoing is a non-limiting summary of some embodiments, and does not limit the scope of any claim.BRIEF DESCRIPTION OF DRAWINGS
[0063] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component that is illustrated in various figures is represented by a like numeral. For purposes of clarity, not every component may be labeled in every drawing. In the drawings:
[0064] FIG. l is a block diagram of an example disease risk phenotyping model, according to some embodiments.
[0065] FIG. 2A is a block diagram of another example disease risk phenotyping model, according to some embodiments.
[0066] FIG. 2B is a block diagram of an example encoder of a disease risk phenotyping model, according to some embodiments.
[0067] FIG. 2C is a block diagram of another example encoder of a disease risk phenotyping model, according to some embodiments.
[0068] FIG. 3 is a flowchart of an example disease risk phenotyping method, according to some embodiments.Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT
[0069] FIG. 4 is a block diagram of an example computer device, according to some embodiments.
[0070] FIG. 5 illustrates example metrics for evaluating a disease risk phenotyping technique.
[0071] FIGS. 6A and 6B illustrate the precision and specificity, respectively, of a set of risk models for coronary artery disease.
[0072] FIGS. 7A and 7B illustrate the precision and specificity, respectively, of a set of risk models for breast cancer.
[0073] FIG. 8A illustrates the receiver operating characteristic (ROC) curves for a set of risk models for Alzheimer’s disease.
[0074] FIG. 8B illustrates the ROC curves for a set of risk models for glaucoma.
[0075] FIG. 8C illustrates the ROC curves for a set of risk models for colorectal cancer.
[0076] FIG. 8D illustrates the ROC curves for a set of risk models for prostate cancer.
[0077] FIG. 9 shows a table of area under receiver operating characteristic (AUROC) curve values and area under precision-recall curve (AUPRC) values for sets of risk models for Alzheimer’s disease, stroke, chronic kidney disease, coronary artery disease, glaucoma, prostate cancer, colorectal cancer, ovarian cancer, and lung cancer.DETAILED DESCRIPTIONIntroduction
[0078] Many traditional models for disease risk are inaccurate. False positive diagnoses of disease (or predictions of disease risk), which are common, can lead to expensive, unnecessary, and potentially harmful tests and treatments. Thus, the accuracy (e.g., recall, specificity, precision, negative predictive value, etc.) of an early disease detection method is an important consideration. False negative assessments of disease risk can lead to late detection of disease, such that the benefits of early detection are unavailable.
[0079] The inaccuracy of many traditional models for disease risk arises from their reliance on relatively simple statistical analysis techniques and their failure to leverage large portions of an individual’s available health data, including genetic data. For example, an individual’s 10-year risk of developing coronary artery disease (CAD) is often estimated using the Pooled Cohort Equations (PCEs), which are based on the application of multivariable regression models to a small set of health parameters including the individual’s age, sex, race, smoking status, systolic blood pressure, diabetes status, hypertension treatment status, total cholesterol level, and HDL cholesterol level. Likewise, an individual’s 10-year risk of developing breast cancer is often estimated using the Tyrer-Cuzick (TC) model, which includes a genetic segregation model thatAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTconsiders family history and a few specific genetic mutations (e.g., BRCA1 and BRCA2) and a regression model that considers a small set of health parameters including the individual’s hormone levels, weight, age, and breast density. These models often miss the complex, dynamic interactions that influence disease development. Risk models for many other diseases are equally or more simplistic.
[0080] The present disclosure describes deep learning based (DL-based) phenotyping models. In some examples, these DL-based phenotyping models can estimate an individual’s disease risk with improved recall, sensitivity and / or specificity relative to traditional models. These DL-based phenotyping models can include neural networks (e.g., deep neural networks, autoencoders, fusion networks, generative models, etc.) and, optionally, other machine learning models (e.g., classifiers) trained to estimate an integrated risk score (IRS) indicating an individual’s risk of developing a specific disease based on the individual’s health data, including electronic health records, polygenic risk score (PRS), and / or principal components of ethnicity (PC-ethnicity).
[0081] In some examples, DL-based phenotyping models triage individuals to appropriate interventions before symptoms arise based on integrated analysis of polygenic risk scores (PRS) and clinical factors. In contrast to traditional risk models that predict relative risk, some examples of DL-based phenotyping models provide absolute risk estimates, which are better suited for clinical care and early screening. In some examples, by using deep neural networks (DNNs) to analyze a rich set of health data (e.g., structured and unstructured data from electronic health records (EHRs), polygenic risk scores, principal components of ethnicity, etc.), DL-based phenotyping models improve phenotyping, increase variant discovery, and provide actionable insights for disease prevention and early intervention. Furthermore, in some examples, the insights gained from analyzing the predictions of DL-based phenotyping models can be used to update PRS models, thereby ensuring that the predictions of phenotyping models evolve with and adapt to emerging genetic discoveries.
[0082] In some examples, the improved accuracy of DL-based phenotyping models can yield more tailored screening and intervention relevant for most of the adult population. In some examples, DL-based phenotyping models can be used to predict an individual’s risk for common and / or severe diseases for which no other accurate early-detection models are currently available. In some examples, a specialized treatment protocol can be initiated (e.g., by a clinician) based on a model’s indication that an individual’s risk for a particular disease exceeds a threshold level.
[0083] In some examples, a DL-based phenotyping model for disease risk has a unique architecture that enables the model (1) to learn, from a large amount of health data, aAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTrepresentation of the health data that is useful for determining an individual’s risk with respect to one or more diseases, and (2) to jointly analyze that representation of health data, additional structured health data (e.g., polygenic risk scores), and unstructured text relating to a person’s health (e.g., unstructured text obtained from an individual’s electronic health records) to determine the individual’s disease risk. In some examples, the phenotyping model includes an autoencoder arranged to learn a useful representation of a large amount of health data (e.g., structured health data extracted from an individual’s electronic health records), a subnetwork arranged to analyze additional health data, and a generative model (e.g., LLM) arranged to extract information from the unstructured text of electronic health records (EHRs). However, jointly analyzing such disparate data types in a unified deep neural network presents technical challenges, including dimensional mismatch of features (e.g., embeddings) and data type heterogenicity. In some examples, the phenotyping model includes a unique merge layer that surmounts these technical challenges, thereby enabling the phenotyping model to analyze a richer set of data than prior models, yielding improved accuracy of disease risk scoring.
[0084] Many conventional computational models for disease-risk estimation provide inaccurate or poorly calibrated risk outputs. False positive risk assessments can cause computing systems used in clinical workflows to route subjects to unnecessary follow-up procedures, testing, or review queues, while false negative risk assessments can prevent the computing systems from identifying subjects for whom earlier follow-up is appropriate. Accuracy metrics such as recall, specificity, precision, negative predictive value, calibration, and discrimination are accordingly important technical performance characteristics of computerized disease-risk phenotyping systems.
[0085] Conventional disease-risk models often rely on relatively small sets of manually selected variables and relatively simple statistical techniques. For example, an individual’s ten-year risk of developing coronary artery disease may be estimated using pooled cohort equations based on a limited set of variables, and an individual’s risk of developing breast cancer may be estimated using models that consider family history, selected genetic mutations, and a limited set of clinical parameters. Such models generally do not provide a computational architecture for jointly processing heterogeneous, high-dimensional health data including structured electronic health record data, unstructured clinical text, polygenic risk score data, principal-component ethnicity data, image-derived data, laboratory-derived data, temporally ordered clinical events, and / or other health data. As a result, conventional models may fail to capture nonlinear, temporal, and cross-modal interactions among health-data features that are relevant to diseaserisk phenotyping.Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT
[0086] Heterogeneous health data present technical challenges when used in a computerized risk-phenotyping model. Structured electronic health record fields, free-text clinical notes, polygenic risk scores, genetic or ancestry features, laboratory values, medical images, and longitudinal event sequences have different formats, dimensions, sparsity patterns, noise characteristics, time scales, and levels of reliability. Some modalities may be missing, stale, incomplete, or associated with different confidence levels. A model that merely applies a generic classifier to such data can lose clinically relevant information, overemphasize unreliable inputs, or fail to combine features generated by different encoders in a computationally coherent manner.
[0087] The disclosed systems and methods address these technical challenges using a particular deep learning architecture that converts different health-data modalities into modalityspecific embeddings and combines the modality-specific embeddings in a merge layer to generate a unified representation for disease-risk classification. In some examples, the architecture includes an encoder configured to generate embeddings from unstructured clinical notes or to extract structured health data from unstructured clinical notes, and one or more subnetworks configured to generate embeddings from additional health data including polygenic risk score data, biomarker data, temporal data, image data, structured health data, or sample-derived laboratory data, a merge layer configured to combine the embeddings, and a classifier configured to generate disease-risk data based on the unified representation. The merge layer may use projection, attention, gating, concatenation, or combinations thereof to combine embeddings having different dimensions, sources, and reliability characteristics.
[0088] In some examples, the merge layer improves the operation of the computerized phenotyping model by resolving dimensional mismatch and data-type heterogeneity among embeddings generated by different branches of the model. For example, the merge layer may compute modality weights or gate values based on domain context, including target disease, disease category, prediction horizon, subject attributes, availability of data modalities, missingdata patterns, recency of clinical information, confidence values associated with extracted features, or reliability measures associated with data sources. By selectively weighting, gating, projecting, or attending to modality-specific embeddings, the model can generate a unified representation that preserves cross-modal interactions while reducing the influence of missing, stale, low-confidence, or less reliable data.
[0089] In some examples, the architecture is trained end-to-end using training records for multiple subjects and labels indicating observed disease status, disease onset, time-to-event outcomes, risk categories, or other clinical outcomes. Gradients may be backpropagated from theAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTclassifier through the merge layer and into one or more upstream branches, including a structured-data encoder, an unstructured data encoder, a polygenic-risk-score subnetwork, a temporal subnetwork, or an image-processing subnetwork. The training objective may include classification loss, reconstruction loss, survival loss, regularization terms, missing-data consistency terms, modality-dropout terms, calibration terms, contrastive terms, or combinations thereof. Joint training enables upstream embeddings and the merge layer to be optimized for the downstream disease-risk classification task rather than being generated as unrelated intermediate values.
[0090] The disease-risk data generated by the disclosed systems and methods are not merely an observation of a natural correlation or a mental assessment of disease risk. Rather, the disease-risk data are produced by a trained model that transforms heterogeneous health data into embeddings, combines the embeddings into a unified representation using a merge layer configured to address cross-modal technical constraints, and classifies disease risk from the unified representation. In some examples, the resulting disease-risk data may be used to provide a risk-based recommendation for further screening, diagnostic imaging, laboratory testing, clinical evaluation, lifestyle modification, dietary modification, or other follow-up action.Terminology and ConceptsTerminology and Concepts Relating to Health Data
[0091] Electronic health records (EHRs) include a digital representation of an individual’s health data. EHRs may include a wide variety of health data, including an individual’s demographics, medical history, health conditions and diagnoses, medications, treatment plans, immunization dates, allergies, medical images, laboratory test results, medical procedures, etc. In some examples, an individual’ s EHRs include notes regarding a health care providers’ interactions with or treatment of the individual (e.g., progress notes and / or visit summaries). Such notes may be formatted as unstructured text.
[0092] Polygenic risk scores (PRS) have been shown to effectively stratify disease risk in some diseases, including some cancers and coronary artery disease (CAD). Polygenic risk scores may include statistically-derived scores that estimate an individual’s genetic predisposition to a specific health condition by combining the effects of multiple genetic variations across the genome. Thus, a PRS can indicate the combined influence of many genetic variations on a specific disease risk for an individual, even when each of the variations individually has a small impact on the individual’s overall risk.Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT
[0093] Principal components of ethnicity (PC-ethnicity) may refer to summaries of the genetic variations associated with different ethnic groups captured via principal component analysis (PCA). In the context of ethnicity, PCA can reduce the complexity of genetic data by identifying the main axes (principal components) along which the genetic data vary. These components can represent the genetic ancestry and population structure, which can help DL-based phenotyping models understand how genetic differences among ethnic groups influence disease risk, account for population stratification, and avoid spurious associations that may arise due to genetic differences between ethnic groups, thereby improving the accuracy of disease risk predictions.Terminology and Concepts Relating to Machine Learning and Neural Networks
[0094] “Machine learning” may refer to the application of certain techniques (e.g., pattern recognition and / or statistical inference techniques) by computer systems to perform specific tasks. Machine learning techniques may be used to train models based on sample data (e.g., “training data”) and to validate the models using validation data (e.g., “testing data”). The sample and validation data may be organized as sets of records (e.g., “observations” or “data samples”), with each record indicating values of specified data fields (e.g., “independent variables,” “inputs,” “features,” or “predictors”) and corresponding values of other data fields (e.g., “dependent variables,” “outputs,” or “targets”). Machine learning techniques may be used to train models to infer the values of the outputs based on the values of the inputs. When presented with other data (e.g., “inference data”) similar to or related to the sample data, such models may accurately infer the unknown values of the targets of the inference dataset.
[0095] Features can have data types. For instance, a feature can have a numerical data type, a categorical data type, a time-series data type, a text data type (e.g., a structured text data type or an unstructured (“free”) text data type), an image data type, or any other suitable data type. In general, a feature’s data type is categorical if the set of values that can be assigned to the feature is finite.
[0096] Deep learning (DL) is a subset of machine learning (ML) that focuses on artificial neural networks (ANN) and their ability to learn and make decisions. Deep learning involves the use of complex algorithms to train ANNs to recognize patterns and make predictions based on large amounts of data. An important difference between DL and traditional ML algorithms is that DL algorithms can learn multiple layers of representations, allowing them to model highly nonlinear relationships in the data.
[0097] Most DL methods use ANN architectures, which is why DL models are often referred to as deep neural networks (DNNs). The term “deep” may refer to the presence or abundance ofAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCThidden layers in the neural network. For example, DNNs can have as many as 150 hidden layers (or more). DL uses these hidden layers to progressively extract higher-level features from the raw input. DL models can be trained using large sets of labeled data and models (e.g., autoencoders) that learn features directly from the data without the need for manual feature extraction.
[0098] An autoencoder is a neural network that includes an encoder and a decoder. The encoder transforms input data into a lower-dimensional latent representation (e.g., a vector embedding) which may capture features that are useful for downstream tasks while also reducing noise. The decoder reconstructs the original input from this latent representation.
[0099] An autoencoder-based classifier is a neural network that includes an autoencoder and a classifier. The classifier utilizes the autoencoder’s latent representation to perform classification tasks, leveraging the compact and informative feature set produced by the encoder for improved accuracy and efficiency. Regularized autoencoders (e.g., sparse, denoising and / or contractive autoencoders) can be particularly effective in learning representations for subsequent classification tasks. By combining dimensionality reduction and classification within a unified neural network model, the autoencoder-based classifier can predict disease risk based on a rich set of health data with high accuracy.
[0100] A fusion network is a type of neural network that integrates multiple data sources or modalities into a unified model for enhanced learning and prediction. A fusion network includes two or more sub-networks, each dedicated to processing a specific set or type of input data. These sub-networks can extract relevant features from their respective inputs and subsequently combine these features in a fusion layer. The fusion layer can synthesize the diverse information, enabling the network to leverage complementary insights from different data sets or data types.
[0101] Generative Al technology generally uses generative models such as Generative Adversarial Networks (GANs), transformer-based models, diffusion models (e.g., stable diffusion models), and / or Variational Autoencoders (VAEs), etc., which are based on artificial neural networks and deep learning. In many generative Al systems, the generative model that generates content is a large language model (LLM). A large language model (LLM) is a type of DL model that can perform a variety of natural language processing (NLP) tasks such as generating and classifying text, answering questions in a conversational manner, and translating text from one language to another. The term Targe’ refers to the number of values (parameters) the language model can change autonomously as it learns. Some LLMs have hundreds of billions of parameters. In general, LLMs are NN models that have been trained using deep learningAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTtechniques to recognize, summarize, translate, predict, and generate content using very large datasets.
[0102] Many state-of-the-art LLMs use a class of deep learning architectures called transformer neural networks (“transformer networks” or “transformers”). A transformer is a neural network that learns context and meaning by tracking relationships between data units, such as the words in a sentence. A transformer can include multiple transformer blocks, also known as layers. For example, a transformer may have self-attention layers, feed-forward layers, and normalization layers, all working together to decipher input to predict (or generate) streams of relevant output. The layers can be stacked to make deeper transformers and powerful language models.
[0103] Two innovations that make transformers particularly adept for large language models are positional encodings and self-attention. Positional encoding embeds the order in which the input occurs within a given sequence. Rather than feeding words within a sentence sequentially into the neural network, with positional encoding, the words can be fed in non-sequentially. Selfattention assigns a weight to each part of the input data while processing it. This weight signifies the importance of that portion of the input in the context of the rest of the input. The use of the attention mechanism enables models to focus on the parts of the input that matter the most. This representation of the relative importance of different inputs to the neural network is learned over time as the model sifts and analyzes data. These two techniques in conjunction allow for analyzing the subtle ways and contexts in which distinct elements influence and relate to each other over long distances, non-sequentially. The ability to process data non-sequentially enables the decomposition of the complex problem into multiple, smaller, simultaneous computations.
[0104] “Completion” may refer to the process of a generative model generating additional content (e.g., text) based on a provided prompt (e.g., text), e.g., providing the next word in a sentence. In some contexts, the additional content (e.g., text) provided by the generative model may also be referred to herein as a “completion.” Completions generated by generative models may include text, audio data (e.g., speech, music, etc.), image data (e.g., images), video data (e.g., videos), time-series data, or any other suitable type of data.
[0105] “Prompting” may refer to a technique in which a generative model (e.g., an LLM) is matched to a desired downstream task by formulating the task as natural language text explaining the desired behavior, such that a generative model can carry out the task by performing text completion. Often these instructions are split into a “system message” containing general task instructions providing general guidance about the desired behavior and a “prompt template” containing the portion of the prompt that contains indicator values that are substitutedAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTin each use. “Fine-tuning” may refer to the process whereby a generative model is adapted to a particular task by changing its parameters.
[0106] Generative models can analyze existing content, identify patterns in the content, and combine or modify the identified patterns to generate new content. The new content can include text, images, video, music, or any other suitable type of content. Some non-limiting examples of generative models include generative adversarial networks (GANs), variational autoencoders (VAEs), autoregressive models (e.g., large language models (LLMs)), recurrent neural networks (RNNs), transformer-based models, reinforcement learning models for generative tasks, etc. Transformer-based models generally have an encoder-decoder architecture, use an attention mechanism (e.g., scaled dot-product attention, multi-head attention, masked attention, etc.) to model the relationships between different elements in a sequence of content, and perform well when processing long sequences of content. Some non-limiting examples of transformer-based models include Generalized Pre-trained Transformer 4 (GPT-4), DALL-E3, etc. Other examples of generative models with text-processing capability include Jurassic- 1, Command, and Paradigm.
[0107] The term “generative model” as used herein may generally refer to a type of machine learning model that is trained on existing data to enable the generative model to generate, based on an input or prompt, new data that shares characteristics similar to that of the training data. In some examples, a generative model may handle text. In these examples, the generative model may accept text prompts and produce text outputs. Any suitable type of Al model can be used, including predictive models, generative Al (“Gen Al”) models, etc. Predictive models can analyze historical data, identify patterns in that data, and make inferences (e.g., produce predictions or forecast outcomes) based on the identified patterns. Some non-limiting examples of predictive models include neural networks (e.g., deep neural networks (DNNs), convolutional neural networks (CNNs), recurrent neural networks (RNNs), learning vector quantization (LVQ) models, etc.), regression models (e.g., linear regression models, logistic regression models, linear discriminant analysis (LDA) models, etc.), decision trees, random forests, support vector machines (SVMs), naive Bayes models, classifiers, etc.Some Examples of DL-Based Techniques for Disease Risk Phenotyping
[0108] FIG. 1 shows an example architecture of a deep learning based model 100 for disease risk phenotyping. Based on an individual’s health data, the disease risk phenotyping model 100 may provide (e.g., predict or generate) disease risk data 122 (e.g., one or more risk scores) indicating an individual’s risk of having one or more diseases.Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT
[0109] In some examples, a risk score indicates an individual’s risk of having a particular disease (e.g., a specific disease or a disease within a specific group of diseases). In some examples, a risk score indicates an individual’s risk of currently having a particular disease, risk of having the particular disease at any future time in the individual’s life, or risk of having the particular disease within a finite time period (e.g., a finite time period beginning at the present time or at a future time). Any suitable finite time period may be used (e.g., 1-12 months, 1-5 years, 1-10 years, 1-20 years, 5-10 years, 10-20 years, etc.).
[0110] In some examples, an individual’s risk score for a particular disease is compared to one or more thresholds to categorize the individual’s risk level for that disease (e.g., low, normal, elevated, etc.). For example, if an individual’s risk score for a disease exceeds a first threshold, the individual’s risk level for that disease may be elevated. Assuming the risk score is normalized to have a mean of 0 and a standard deviation (SD) of 1, and the odds ratio (OR) per standard deviation is r, the probability of a positive for disease at risk score x is p(x) = Crxwhere C is a constant. Integrating over all possible values of x and associated probabilities reproduces the population incidence, p:pr= -^= f°° Crxe~xdx = Ce^ln^ , hence C = pe~^ln^ .V27rJ-°°
[0111] If individuals with risk scores above a threshold value t are categorized as having elevated risk, the probability P(t) that an individual’s risk score for a disease is above the threshold value t and the individual is positive for the disease can be found as follows:P(t) = -^= J'(00p(x)e~xdx =perxe~xdx = p(l — normcdf(t — a)),where normed f is the normal cumulative distribution. Similarly, the probability N(t) that an individual’s risk score for a disease is above the threshold value t and the individual is negative for the disease is as follows:A(t) = Jt°°(l — p(x))e-2xdx = 1 — normcdf(t) — p(l — normcdf(t — a)).
[0112] Thus, the sensitivity at threshold t is given by:sensitivity(t) = = 1 — normedfft — ln(r)'>).
[0113] Given population incidence p and a risk score with odds ratio r per standard deviation, the threshold t to achieve a specified sensitivity is given by:t =ln In (r) + norminv l — sensitivity) .Likewise, the Positive Predictive Value (PPV) is given by:, t —P(t)— p(1 normcd-f(1-ZnW))P(t~)+N(t) l-normcdf(t~) ’Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTand the specificity is given by:l-p-Nft} normcdf(t)-p(normcdf(t-ln(rf)}specif icity(t) = —± p= - - ■
[0114] In some embodiments, a system for disease risk phenotyping may recommend one or more actions for an individual whose risk score for a disease indicates that the individual is at elevated risk for the disease. Any suitable action(s) may be recommended. For example, if a subject’s risk score for a disease is high, the system may recommend that the subject undergo further diagnostic imaging (e.g., MRI, CT-scan, etc.) or one or more therapies suitable for treating the disease (e.g., therapeutic procedures (e.g., coronary catheterization, biopsy, etc.)). In some examples, if a subject’s risk score for a disease is high, the system may recommend lifestyle changes, dietary changes, additional lab work, etc. In some examples, the system may identify subjects most at risk for a disease and recommend them for additional or continued screening.
[0115] Dietary recommendations can be similar to those made to other at-risk subjects, but providing the recommendations in combination with a more accurate risk score may lead to a greater probability of the subject following the recommendations. Recommended lab work can include more precise (but often more expensive) testing usually reserved for more at-risk patients. Recommended screenings can include screening tests that are recommended in the general population, but they can start at a younger age and be conducted at more frequent intervals. A recommended screening frequency can be determined by comparing the risk score of a patient to the risk score associated with an average person of a specific age. Examples of screening tests that can be recommended on a targeted basis include mammograms for breast cancer and colonoscopies for colon cancer.
[0116] The disease risk phenotyping model 100 may produce disease risk data 122 (e.g., a risk score) indicating an individual’s risk of having one or more diseases based on any suitable health data. In some examples, the health data include an individual’s electronic health records (EHR) or data extracted or derived therefrom. Such health data may include, without limitation, one or more diagnostic codes (e.g., ICD-10 diagnostic codes) indicating diagnosed health conditions (e.g., diseases) of the individual; one or more procedure codes indicating procedures administered to the individual; demographic data of the individual (e.g., age, gender, ethnicity, etc.); admission, discharge and transfer (ADT) records of the individual; results of laboratory tests administered to the individual; vital signs of the individual; diagnostic data related to diagnosed and / or undiagnosed health conditions (e.g., diseases) of the individual; physiological monitoring data obtained from one or more wearable sensors (e.g., heart rate, blood oxygen saturation, activity level, sleep patterns, electrodermal activity, or glucose measurements),Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTincluding but not limited to continuous physiological monitoring data; encounter data related to encounters between the individual and one or more healthcare providers; clinical records related to health or treatment of the individual; allergy data indicating one or more allergies of the individual; procedure data indicating one or more procedures (e.g., surgeries) performed on or administered to the individual; immunization data indicating one or more immunizations administered to the individual; a medical history of the individual; a family medical history of the individual; a social history of the individual; one or more radiology reports corresponding to radiological scans of the individual; one or more medical images of the individual; medication data indicating one or more medications prescribed for or used by the individual; results of one or more cardiac health tests (e.g., electrocardiogram (ECG), echocardiogram, etc.) of the individual; microbiology data, pathology data, or any other suitable data obtained from analysis of one or more samples of blood, urine, sputum, or tissue of the individual; infusion data related to one or more substances infused into a bloodstream of the individual; treatment data (e.g., treatment records) related to one or more treatments administered to the individual; and / or any other suitable data pertaining to the individual.
[0117] In some examples, at least a portion of the health data may include temporal data indicating when those portions of health data were obtained. Thus, the health data may provide a longitudinal view of the individual’s health and medical history.
[0118] In some examples, at least a portion of the health data (e.g., a portion of the data in an individual’s EHRs) is structured. In some examples, at least a portion of the health data is unstructured. In some examples, structured health data may be extracted or derived from unstructured health data. Any suitable technique may be used to extract or derive structured health data from unstructured health data including, without limitation, providing the unstructured health data as context to a generative model (e.g., LLM) and prompting the generative model to answer questions about the individual’s health based on the context.
[0119] In some examples, the health data include one or more polygenic risk scores (PRS) of the individual for one or more health conditions (e.g., diseases, traits, etc.). An individual’s PRS for a health condition may be obtained using any suitable techniques. For example, an individual’s PRS for a health condition may be retrieved from a database of PRS or determined based on the individual’s genotype. In some embodiments, the individual’s genotype is determined using whole genome sequences (WGS) (e.g., low coverage WGS) or microarray genotypes at the time of the individual’s screening or as early as the individual’s birth. In some embodiments, an individual’s PRS for a health condition is obtained by (1) analyzing the individual’s genotype (the combination of alleles carried by the individual) for any geneticAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTvariants (e.g., single nucleotide polymorphisms or “SNPs”) associated with the health condition; (2) for each such genetic variant, multiplying the number of risk alleles the individual carries by the corresponding effect size to obtain weighted values; and (3) summing the weighted values across all such genetic variants to produce the PRS for the health condition. In some examples, genetic variants associated with a health condition may be identified based on genome-wide association studies (GWAS). In some embodiments, an individual’s PRS for a health condition is obtained (e.g., generated) by applying a polygenic model associated with the health condition to the genotype of the individual.
[0120] In some examples, the health data include biomarker data associated with one or more health conditions. For example, the biomarker data may indicate whether the individual has one or more biomarkers for one or more health conditions, the individual’s measured levels of biomarkers for one or more health conditions, etc.
[0121] In some examples, the health data include the individual’s principal components of ethnicity. An individual’s principal components of ethnicity may be obtained using any suitable techniques. In some examples, an individual’s principal components of ethnicity are determined based on statistical analysis of genetic variation within one or more ethnic groups to which the individual belongs. In some examples, an individual’s principal components of ethnicity are provided by a model or database based on data identifying one or more ethnic groups to which the individual belongs.
[0122] In some examples, the health data indicate whether an individual exhibits one or more lifestyle factors that can affect health (e.g., lack of physical activity, poor nutrition, inadequate sleep, tobacco use, narcotics use, excessive alcohol consumption, etc.).
[0123] In some examples, the health data include characterizations of aspects of the individual’s environment (e.g., home environment, work environment, etc.) that can affect health (e.g., air quality, water quality, exposure to pollutants, exposure to unsafe conditions, noise level, etc.).
[0124] In some examples, in addition to health-related information (some non-limiting examples of which are described above), the health data may include temporal information. The temporal information may indicate dates and / or times associated with the health-related information. For example, the temporal information may indicate the dates when an individual was diagnosed with various health conditions (e.g., diseases), the dates when various procedures were performed on the individual, the dates when medical images or lab results of the individual were obtained, the dates when subsequently analyzed samples were obtained from the individual, etc. Such temporal information may help the model 100 learn and understand how theAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTprogression of various diseases affect and are reflected in an individual’s health-related information over time.
[0125] In scenarios in which the health data analyzed by the disease risk phenotyping model 100 include two or more distinct types of health data (e.g., two or more types of health data selected from EHR, PRS, biomarker data, and PC-ethnicity), the risk score provided by the disease risk phenotyping model 100 may be referred to as an integrated risk score (IRS).
[0126] The disease risk phenotyping model 100 may provide an individual’s risk scores for any suitable set of one or more diseases. Some non-limiting examples of diseases for which the model can provide risk scores include breast cancer, lung cancer, prostate cancer, colorectal cancer, pancreatic cancer, ovarian cancer, endometrial cancer, cervical cancer, liver cancer, gastroesophageal cancer, small intestine cancer, bladder cancer, brain cancer, kidney cancer, thyroid cancer, leukemia, multiple myeloma, lymphoma (both Hodgkin and non-Hodgkin), melanoma, head and neck cancer, obesity, coronary artery disease (CAD), type 2 diabetes (T2D), stroke, Alzheimer’s disease (AD), glaucoma, and age-related macular degeneration (AMD).
[0127] Still referring to FIG. 1, the disease risk phenotyping model 100 may have any suitable architecture. In some embodiments, the model 100 includes an encoder 110 and a classifier 120. Optionally, the model 100 may also include one or more subnetworks 130 and a merge layer 140. Some embodiments of components of the model 100 are described in further detail below.
[0128] In some embodiments, the encoder 110 provides encodings 112 of one or more features based on first health data 102 of an individual (e.g., extracts feature encodings 112 from the first health data). In some examples, the feature encodings 112 are embeddings (e.g., vector embeddings, latent embeddings, latent vector embeddings). The first health data 102 may include all available health data of the individual or any suitable portion thereof (e.g., one or more types of the individual’s health data). In some examples, the first health data 102 include EHR of the individual or structured data obtained from EHR of the individual. Other non-limiting examples of types of health data are described herein. In some examples, the encoder 110 is or includes a neural network. In some examples, the encoder 110 is part of an autoencoder. In some examples, the autoencoder is configured to embed portions of the first health data 102, which may include numeric biomarker values, lab results, lifestyle factors, environmental factors, and / or the polygenic risk score (PRS) for the subject, etc.
[0129] In some examples, the encoder 110 includes a large language model (LLM) configured to process unstructured clinical notes associated with a subject, including progress notes, encounter summaries, discharge summaries, radiology reports, pathology reports, medication notes, procedure notes, and / or clinician-entered free-text fields. The LLM can tokenize theAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTunstructured clinical notes and generate contextual vector representations of words, phrases, sentences, note sections, or complete notes based on semantic and temporal relationships within the unstructured clinical notes. The contextual vector representations can be aggregated to generate one or more note-level embeddings corresponding to respective clinical notes, encounter-level embeddings corresponding to respective clinical encounters including multiple clinical notes or note sections, and / or patient-level embeddings corresponding to a longitudinal record of the subject across multiple encounters. The aggregation may include pooling hidden states output by the LLM, applying an attention mechanism to token-level or section-level representations, combining embeddings generated for temporally ordered note chunks, and / or projecting the contextual vector representations into a predetermined embedding dimension. The note-level embeddings, encounter-level embeddings, or patient-level embeddings can be supplied to a merger layer 140 of the model 100, and the merge layer 140 can combine the embeddings generated from the unstructured clinical notes with embeddings generated from structured health data, genetic data, polygenic risk score data, biomarker data, imaging data, lifestyle data, environmental data, temporal data, and / or any other suitable data to generate a fused representation for classifying disease risk of the subject.
[0130] In some embodiments, the classifier 120 provides disease risk data 122 indicating the risk of the individual having one or more diseases based, at least in part, on the feature encodings 112. In some examples, the disease risk data 122 provided by the classifier 120 are further based on feature encodings 132 provided by one or more subnetworks 130. Some embodiments of such subnetworks 130 and feature encodings 132 are described below. In some examples, the disease risk data 122 include raw or normalized risk scores of the individual for one or more diseases. In some examples, the disease risk data 122 include categorizations of the individual’s risk levels for one or more diseases (e.g., low, normal, elevated, etc.).
[0131] Any suitable type of classifier 120 may be used. In some examples, the classifier is or includes one or more decision trees, naive Bayes classifiers, k-nearest neighbors classifiers, logistic regression models, and / or support vector machines (SVM). In some examples, the classifier is or includes a neural network. In some examples, the classifier is or includes a generative model (e.g., LLM) trained to generate the disease risk data 122 based on the feature encodings 112 and / or the feature encodings 132.
[0132] In some examples, the model 100 includes the encoder 110, the classifier 120, one or more subnetworks 130, and a merge layer 140 (or “fusion layer”) organized as a fusion network. In some examples, the subnetworks 130 provide feature encodings 132 of one or more features based on second health data 104 of the individual. In some examples, the feature encodings 132Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTare embeddings (e.g., vector embeddings, latent embeddings, latent vector embeddings, etc.). The second health data 104 may include all available health data of the individual or any suitable portion thereof (e.g., one or more types of the individual’s health data). In some examples, the second health data 104 include EHR of the individual or unstructured data obtained from EHR of the individual, PRS of the individual for one or more diseases, principal components of ethnicity of the individual, temporal data, and / or imaging data (e.g., medical imaging data). Medical imaging data may include medical images of the individual (e.g., medical images obtained using any suitable imaging modality, for example, radiography (e.g., X-ray, positron emission tomography (PET) scan), computed tomography (CT) (e.g., CT scan), magnetic resonance imaging (MRI), sonography (e.g., ultrasound), nuclear medicine, hybrid modalities (e.g., PET / CT, MRI / CT), etc. In some examples, medical imaging data include expert interpretations of the medical images (e.g., notes provided by a radiologist or other clinician).
[0133] In some examples, each of the subnetworks 130 is or includes a neural network (e.g., an encoder). In some examples, each of the subnetworks 130 provides encodings 132 of a corresponding subset of the second health data 104 (e.g., unstructured data from EHR, PRS, PC-ethnicity, temporal data, imaging data, etc.). In some examples, at least one subnetwork 130d includes a generative model (e.g., LLM) trained to extract structured health data from unstructured portions of the second health data 104 (e.g., unstructured portions of EHR) and an encoder configured to encode the extracted health data. In some examples, at least one subnetwork includes an image feature extraction model (e.g., a convolutional neural network (CNN), a vision transformer (ViT), a pretrained image foundation model, etc.) configured to extract features from medical imaging data (e.g., medical images).
[0134] In some examples, at least one subnetwork 130 includes a large language model (LLM) configured to extract structured health data from unstructured health data of the subject. The unstructured health data may include progress notes, encounter summaries, discharge summaries, radiology reports, pathology reports, medication notes, procedure notes, and / or clinician-entered free-text fields obtained from one or more electronic health records (EHRs) of the subject. The LLM may receive the unstructured health data as context and receive a prompt that specifies one or more clinical features to be extracted from the unstructured health data. The prompt may include a system message defining an extraction task, a prompt template identifying feature names, expected output formats, units, time windows, confidence indicators, or missing-data indicators, and one or more portions of the unstructured health data. In response to the prompt, the LLM can generate structured health data including one or more extracted feature values of suitable features (e.g., diagnoses, symptoms, medications, medication start or stop dates,Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTprocedures, laboratory findings, imaging findings, family history, social history, lifestyle factors, environmental exposures, and / or clinician-assessed disease indicators). The extracted feature values may be normalized to a predetermined vocabulary, code set, numerical range, categorical value set, or temporal representation and may be associated with respective timestamps, encounter identifiers, source-note identifiers, or confidence values. An encoder of the subnetwork 130 can encode the structured health data extracted by the LLM to generate feature encodings 132, and the feature encodings 132 can be supplied to the merge layer 140 for combination with encodings 112 generated from first health data 102 and with other feature encodings generated by one or more additional subnetworks 130.
[0135] In some examples, one of the subnetworks 130 is an image-processing subnetwork configured to process medical image data associated with the subject. The medical image data may include one or more radiographic images, computed tomography images, magnetic resonance images, ultrasound images, mammography images, retinal images, fundus images, optical coherence tomography images, dermatology images, endoscopy images, colonoscopy images, pathology slide images, histology images, or other medical images generated during clinical evaluation of the subject. The image-processing subnetwork may include an image encoder configured to extract image features from the medical image data and to generate one or more image embeddings. The image encoder may include a convolutional neural network, a vision transformer, a U-Net encoder, a residual neural network, a three-dimensional convolutional neural network, or another neural network architecture configured to generate a latent representation of image content relevant to disease risk classification. In some examples, the image encoder generates image embeddings that represent anatomical structures, lesions, tissue morphology, image-derived measurements, radiological findings, pathological findings, ophthalmic findings, or other image-derived features.
[0136] In some examples, the image-processing subnetwork preprocesses the medical image data before generating the image embeddings. The preprocessing may include resizing, cropping, normalization, intensity scaling, color normalization, registration, segmentation, artifact removal, slice selection, patch extraction, study-level aggregation, temporal ordering of serial images, conversion of an image series into a subject-level image representation, or any other suitable image processing task. For example, a plurality of image slices from a computed tomography or magnetic resonance imaging study may be encoded and aggregated to generate a study-level embedding, a plurality of pathology image patches may be encoded and pooled to generate a slide-level embedding, or a plurality of retinal images may be encoded and combined to generate an ophthalmic embedding. The image-processing subnetwork may further associate the imageAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTembeddings with metadata indicating an imaging modality, acquisition time, anatomical region, laterality, clinical indication, image quality measure, source identifier, or other suitable information.
[0137] In some examples, the image embeddings generated by the image-processing subnetwork are provided to the merge layer 140 as feature encodings 132. The merge layer 140 may combine the image embeddings with structured-data embeddings generated by the encoder 110 or another subnetwork 130 and / or with unstructured-text embeddings or extracted-text-feature encodings generated by the encoder 110. The merge layer 140 may concatenate, weight, gate, attend to, or otherwise combine the image embeddings, structured-data embeddings, and unstructured-text embeddings to generate the unified representation 142. The classifier 120 may process the unified representation 142 to generate disease risk data 122 for the subject, such that image-derived information may contribute to disease risk classification together with electronic health record data, polygenic risk score data, biomarker data, lifestyle data, environmental data, temporal data, or other health data associated with the subject.
[0138] In some examples, the merge layer 140 provides a unified representation 142 of the features based on the encodings 112 and 132 (e.g., combines the encodings 112 and 132 into a unified representation 142). In some examples, the merge layer 140 includes one or more neural network layers. In some examples, the merge layer applies a gating mechanism or attention mechanism that weights contributions from the embeddings provided by the encoder 110 and contributions from the embeddings provided by the subnetworks 130 differently based on context information (e.g., domain context).
[0139] In some examples, the merge layer 140 includes an attention facility configured to generate modality weights for respective embeddings received from the encoder 110 and the one or more subnetworks 130. For example, the attention facility may receive a set of modalityspecific embeddings (e.g., one or more structured-data embeddings, LLM-derived embeddings, PRS embeddings, biomarker embeddings, imaging embeddings, lifestyle embeddings, environmental embeddings, and / or temporal embeddings), and may compute an attention score for each modality-specific embedding based on a learned query vector, one or more learned reference vectors, or a domain-context vector. The attention scores may be normalized, for example using a softmax function, to produce modality weights, and the merge layer 140 may generate the unified representation 142 (e.g., as a weighted combination, concatenation of weighted embeddings, or projection of weighted embeddings). In some examples, the attention facility includes multi-head attention that learns different weighting patterns for different disease domains, prediction horizons, clinical contexts, or combinations thereof.Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT
[0140] In some examples, the merge layer 140 includes a gating facility configured to selectively increase or decrease contributions of modality-specific embeddings to the unified representation 142. The gating facility may generate a gate vector based on the modality-specific embeddings and domain context associated with the disease risk classification task. The domain context may include one or more of a target disease, disease category, prediction time horizon, subject age, sex, ancestry information, principal-component ethnicity information, clinical specialty, encounter type, data-source identifier, imaging modality, availability or absence of a data modality, missing-data pattern, confidence value associated with an extracted feature, recency of clinical information, reliability measure associated with a data source, etc. The gate vector may include respective gate values for the modality-specific embeddings, and the merge layer 140 may multiply each modality-specific embedding by a corresponding gate value before combining the gated embeddings. In this manner, the merge layer 140 may learn, for example, to assign a greater contribution to PRS embeddings for diseases having a strong genetic component, to assign a greater contribution to imaging embeddings for diseases associated with diagnostic imaging findings, or to reduce a contribution of an embedding generated from a missing, stale, low-confidence, or less reliable data source.
[0141] The disease risk phenotyping model 100 and / or portions of the model 100 (e.g., the encoder 110, the classifier 120, the subnetworks 130, the merge layer 140) may be trained using any suitable training data and training techniques. For example, the model 100 (or one or more portions thereof) may be trained using a supervised learning algorithm, an unsupervised learning algorithm, a semi-supervised learning algorithm, a deep learning algorithm, or any combination thereof. In some examples, the training data includes health data of a set of individuals. In some examples, the training process involves fitting the model 100 to temporal slices of the health data. For example, an individual’s health data can be partitioned into temporal windows (or slices), and the model 100 can be trained to predict the risk of the individual having the disease in a later temporal window based on the health-related information available in an earlier temporal window.
[0142] In some examples, the encoder 110 and the classifier 120 are jointly trained, such that the encoder 110 learns to produce encodings 112 that are well-suited for predicting an individual’s risk scores. In some examples, the encoder 110 and the classifier 120 are trained separately. Likewise, one or more of the subnetworks 130, the merge layer 140, and the classifier 120 can be trained jointly or separately. In some examples, the encoder 110, subnetworks 130, merge layer 140, and classifier 120 are all jointly trained. In some examples, the generative model (e.g., LLM) of a subnetwork 130 is fine-tuned or trained jointly with the training of one orAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTmore other components of the model 100. The generative model (LLM) may be fine-tuned or trained to embed free-text clinical notes.
[0143] In some examples, the model 100 is trained end-to-end using training records for a plurality of subjects, where each training record includes unstructured health data, structured health data, and one or more labels indicating an observed disease status, disease onset, time-to-event outcome, risk category, or other clinical outcome for a corresponding subject. During training, an LLM of the encoder 110 can process the unstructured health data to generate text-derived embeddings or structured features extracted from the unstructured health data, and a structured-data encoder (e.g., of the encoder 110 or of a sub-network 130) can process structured health data to generate structured-data embeddings. In some examples, the model 100 further includes a polygenic risk score (PRS) subnetwork that processes PRS data, genotype data, principal-component ethnicity data, or other genetic data to generate genetic-risk embeddings, and a temporal subnetwork that processes temporally ordered health data, encounter data, laboratory values, medication data, diagnosis data, procedure data, or clinical-event sequences to generate temporal embeddings. The merge layer 140 can receive the text-derived embeddings or encodings generated from the LLM branch, the structured-data embeddings, and, when present, the genetic-risk embeddings and / or the temporal embeddings, and combine the received embeddings or encodings to generate a unified representation 142 for each subject. The classifier 120 can process the unified representation 142 to generate disease risk data 122, including a disease risk score, a risk category, a probability of disease onset within a predetermined time horizon, or another risk-indicative output.
[0144] In some examples, the encoder 110, the structured-data encoder, the PRS subnetwork, the temporal subnetwork, the merge layer 140, and the classifier 120 are trained jointly by updating parameters of two or more of the encoder 110, the structured-data encoder, the PRS subnetwork, the temporal subnetwork, the merge layer 140, and the classifier 120 based on a common objective function. The common objective function may include a classification loss based on a difference between the disease risk data 122 generated by the classifier 120 and the one or more labels of the training records. The classification loss may include a cross-entropy loss, binary cross-entropy loss, focal loss, survival loss, ranking loss, regression loss, or another supervised learning loss selected based on the type of clinical outcome being predicted. In some embodiments, the common objective function further includes a reconstruction loss associated with reconstructing at least a portion of the structured health data, unstructured-health-data-derived features, genetic data, temporal data, or intermediate embeddings from a latent representation generated by an encoder of the model 100. The common objective function mayAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTalso include one or more regularization terms, missing-data consistency terms, modality-dropout terms, calibration terms, or contrastive terms that promote robustness of the unified representation 142 across different combinations of available health data modalities.
[0145] In some examples, end-to-end training includes backpropagating gradients from the classifier 120 through the merge layer 140 and into one or more upstream branches of the model 100. For example, the gradients may update a projection layer coupled to an output of the encoder 110, one or more trainable layers of the encoder 110, the structured-data encoder, the PRS subnetwork, the temporal subnetwork, or combinations thereof. In some examples, parameters of the encoder 110 are frozen during at least part of training and the end-to-end training updates a trainable projection head that maps LLM outputs to a predetermined embedding dimension used by the merge layer 140. In other examples, at least a portion of the encoder 110 is fine-tuned jointly with the structured-data encoder, the PRS subnetwork, the temporal subnetwork, the merge layer 140, and the classifier 120 using the common objective function. The model 100 may be trained using mini -batches of subject records, temporally segmented records corresponding to respective observation windows, or disease-specific training cohorts, and the trained model 100 may thereafter generate disease risk data 122 for a subject based on one or more available health data modalities.
[0146] Referring to FIG. 2A, in some examples a disease risk phenotyping model 200 includes an encoder (e.g., an encoder 215 of an autoencoder 210), subnetworks 230a-c, a merge layer 240, and a classifier 220. In the example of FIG. 2A, the encoder 215 generates latent embeddings 212 of first health data 202 obtained from an individual’s EHR. The first health data 202 may be pre-processed (e.g., one-hot encoded). In some examples, subnetwork 230a encodes second health data 204 (e.g., an individual’s PC-ethni city data or features derived therefrom, or any other suitable health data) in encodings 232a, subnetwork 230b encodes third health data 206 (e.g., an individual’s PRS scores for one or more health conditions or features derived therefrom, or any other suitable health data) in encodings 232b, and subnetwork 230c encodes fourth health data 208 (e.g., temporal information associated with the first, second, and / or third health data, or any other suitable health data) in encodings 232c. In some examples, the disease risk phenotyping model 200 includes fewer than 3 subnetworks or more than 3 subnetworks. In some examples, the disease risk model 200 includes a subnetwork that encodes medical imaging data (e.g., extract features from medical images and encodes those features). In some examples, the latent embeddings 212 generated by the encoder 215 and the encodings provided by the subnetworks are combined in the merge layer 240. In some examples, the classifier 220 providesAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTrisk data 222 (e.g., the individual’s integrated risk scores or risk levels derived therefrom) for one or more health conditions.
[0147] From a more rigorous theoretical perspective, in some examples the encoder 215 learns a low-dimensional code ZEHR from the high-dimensional first health data 202. In some examples, the subnetworks 230a-c (components of a fusion network) project features derived from the second, third, and fourth health data into embeddings (e.g., ethnicity embeddings ZETH, PRS embeddings ZPRS, and time embeddings ZTIME). In some examples, the merge layer 240 (another component of the fusion network) concatenates the embeddings as shown in Equation (1) or in any suitable manner. In some examples, the fusion network operates as a universal approximator having feedforward subnetworks with non-linear activations.|| |l ^eth II %») * — > Sitrvivafflea<:l(-) € R"Eq- W
[0148] In some examples, the autoencoder and the other components of the model 100 are jointly trained to minimize total loss, which is the sum of autoencoder loss and survival loss as shown in Eq. (2). In some examples, the autoencoder loss Xrecon (e.g., MSE / BCE) enforces reconstruction as shown in Eq (3). In some examples, the survival loss Xsurv (e.g., cross-entropy or Cox partial likelihood) enforces risk ranking and / or prediction as shown in Eq. (4). With sufficient training data, the empirical minimizer converges to the population minimizer as shown in Eq. (5). With sufficient network capacity, the minimization of joint loss can enable both accurate reconstruction and survival modeling, approximating the true functions.(’f • / Riuv (1 Rk. (AClll) Eq. (2) X .5k, , j.with #~ j. Eq (3)2, q,.. s, > ■> risk logits Eq (4) lim •••■ min £^(0)1 > d ~ 0 Eq. (5)
[0149] Some examples have been described in which a disease risk phenotyping model (e.g., model 100 or 200) includes an encoder (e.g., encoder 110 or 215) configured to encode structured data (e.g., structured data extracted from electronic health records). In some examples, that structured data encoder is the encoder portion of an autoencoder. However, any suitable structured data encoder can be used. In some examples, the structured data encoder is an electronic health record (EHR) encoder configured to produce an encoding (e.g., embedding) of one or more EHRs (e.g., one or more EHRs of an individual or group).
[0150] An example of an EHR encoder 250 is shown in FIG. 2B. In some examples, the EHR encoder 250 encodes structured data extracted from one or more EHRs and produces anAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTembedding of the one or more EHRs (e.g., EHR embedding 268). In some examples, the structured data extracted from the EHR(s) includes one or more features, and each feature includes a feature descriptor and a feature value. In some examples, the feature descriptor includes a description (e.g., semantic description) of the feature. Some examples of feature descriptors may include “weight” (describing the individual’s body weight), “smoking” (describing the individual’s status as a smoker or a person who lives with smokers), “diastolic blood pressure” (describing a measurement of the individual’s diastolic blood pressure), etc. Any suitable representation 252 of the feature descriptor may be provided as input to the EHR encoder. In some examples, the feature descriptor representation 252 is plain text. In some examples, the feature descriptor representation 252 is an embedding of the feature descriptor. Such an embedding may be generated by an LLM or any other suitable embedding model.
[0151] In some examples, the feature value indicates a value of the feature. Any suitable representation 254 of the feature value may be provided as input to the EHR encoder. In some examples, the feature value representation 254 is a categorical value, binary value, or numeric value. In some examples, the feature value representation 254 is an embedding of the feature value. Such an embedding may be generated by any suitable embedding model.
[0152] In some examples, the EHR encoder 250 includes one or more feature descriptor encoding layers 256, one or more feature value encoding layers 258, and an embedding combiner 264. In some examples, the feature descriptor encoding layer(s) 256 perform dimension reduction and feature extraction on the feature descriptor representation 252, thereby producing an embedding 260 of the feature descriptor. In some examples, the feature value encoding layer(s) 258 perform dimension reduction and feature extraction on the feature value representation 254, thereby producing an embedding 262 of the feature value. In some examples, the embedding combiner 264 combines the feature descriptor embedding 260 and the feature value embedding 262 to produce an embedding 266 of the feature. The embedding combiner 264 can perform any suitable operation to combine the feature descriptor embedding 260 and the feature value embedding 262 (e.g., summation, concatenation, etc.).
[0153] In some examples, the EHR encoder 250 encodes two or more (e.g., all) features of one or more EHRs to produce two or more feature embeddings 266, and then combines (e.g., concatenates) those feature embeddings 266 to produce an embedding 268 of the EHR(s).
[0154] Another example of an EHR encoder 270 is shown in FIG. 2C. The EHR encoder 270 includes an encoder 272 and an embedder 280. In some examples, the encoder 272 encodes structured data extracted from one or more EHRs and produces a representation 274 of the EHR(s) (e.g., an embedding of the EHR(s)). In some examples, the encoder 272 is EHR encoderAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT250, and the EHR representation 274 is EHR embedding 268. However, any suitable encoder 272 and any suitable EHR representation 274 may be used.
[0155] In some examples, the embedder 280 generates an embedding 282 of the EHR(s) based on the EHR representation 274. In some examples, the embedder 280 includes one or more transformers. In some examples, the transformer’s attention mechanism helps the transformer learn interactions between two or more features of the EHR.
[0156] In some examples, one or more operations of a transformer (e.g., the transformer of the embedder 280) may be performed, approximated, optimized, or regularized using a quantum-assisted computational process. For example, the quantum-assisted computational process may be used to compute or approximate one or more query projections, key projections, value projections, attention scores, attention-weight matrices, positional encodings, temporal encodings, feed-forward layer parameters, or latent clinical embeddings of the transformer. In some examples, the quantum-assisted computational process includes execution of one or more quantum circuits, variational quantum circuits, quantum kernel estimation routines, quantum annealing operations, quantum approximate optimization routines, or hybrid quantum-classical optimization routines.
[0157] In some examples, one or more quantum-assisted models may be used to process health data (e.g., genetic-risk data) associated with a subject. The genetic-risk data may include genotype data, genetic variants, single nucleotide polymorphisms, variant effect sizes, disease-associated loci, ancestry-adjusted genetic-risk features, one or more polygenic risk scores, principal components of ethnicity, or combinations thereof. In some examples, the quantum-assisted model generates a genetic-risk embedding representing inherited risk for one or more diseases. In some examples, the quantum-assisted model estimates nonlinear interactions among genetic variants, interactions among polygenic risk components, interactions between genetic-risk features and clinical features, and / or interactions between genetic-risk features and environmental or lifestyle features.
[0158] In some examples, embeddings generated by the transformer and embeddings generated by a quantum-assisted genetic-risk model are combined by a merge layer or fusion layer, such as merge layer 140 or merge layer 240. The merge layer or fusion layer may learn disease-specific interactions between longitudinal clinical features and inherited genetic-risk features. In some examples, a hybrid quantum-classical optimization process is used to train, fine-tune, calibrate, or otherwise optimize one or more components of the disease risk phenotyping model, including the encoder 110, encoder 215, EHR encoder 250, EHR encoderAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT270, embedder 280, one or more subnetworks 130a-d or 230a-c, merge layer 140 or 240, classifier 120 or 220, or combinations thereof.
[0159] Referring to FIG. 3, a disease risk phenotyping method 300 may include steps 302-306. The method 300 may be performed by any suitable device or system (e.g., a computer device 400 as shown in FIG. 4.).
[0160] In step 302, one or more risk scores are generated for a subject. Each risk score indicates an amount or level of disease risk for a subject with respect to a particular disease. Generating a risk score may include extracting values of features from health data of the subject and classifying the disease risk for the subject based on the values of the features. In some embodiments, the feature values are extracted by one or more neural networks (e.g., an encoder and optionally one or more subnetworks of a fusion neural network). In some embodiments, the classifying is performed by a classifier.
[0161] In step 304, one or more recommended actions for the subject are determined based on the risk score(s). Some non-limiting examples of recommended actions are described herein.
[0162] In step 306, an indication of the one or more recommended actions is provided. The indication may be provided, for example, to a display screen of a user’s computing device via a communication network.Computer-Based Implementations
[0163] Techniques operating according to the principles described herein may be implemented in any suitable manner. Included in the discussion above block diagram(s) and flowchart(s) illustrating components of models and steps of various processes that perform disease risk phenotyping. The processing and decision blocks of the flowchart(s) above represent steps and acts that may be included in algorithms that carry out these various processes. Algorithms derived from these processes may be implemented as software integrated with and directing the operation of one or more single- or multi-purpose processors, may be implemented as functionally-equivalent circuits such as a Digital Signal Processing (DSP) circuit, Field Programmable Gate Array (FPGA), or an Application-Specific Integrated Circuit (ASIC), or may be implemented in any other suitable manner. It should be appreciated that the flowchart(s) included herein do not depict the syntax or operation of any particular circuit or of any particular programming language or type of programming language. Rather, the flowchart(s) illustrate the functional information one of ordinary skill in the art may use to fabricate circuits or to implement computer software algorithms to perform the processing of a particular apparatus carrying out the types of techniques described herein. It should also be appreciated that, unlessAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTotherwise indicated herein, the particular sequence of steps and / or acts described in each flowchart is merely illustrative of the algorithms that may be implemented and can be varied in implementations and embodiments of the principles described herein.
[0164] Accordingly, in some embodiments, the techniques described herein may be embodied in computer-executable instructions implemented as software, including as application software, system software, firmware, middleware, embedded code, or any other suitable type of software. Such computer-executable instructions may be written using any of a number of suitable programming languages and / or programming or scripting tools, and also may be compiled as executable machine language code or intermediate code that is executed on a framework or virtual machine.
[0165] When techniques described herein are embodied as computer-executable instructions, these computer-executable instructions may be implemented in any suitable manner, including as a number of functional facilities, each providing one or more operations to complete execution of algorithms operating according to these techniques. A “functional facility,” however instantiated, is a structural component of a computer system that, when integrated with and executed by one or more computers, causes the one or more computers to perform a specific operational role. A functional facility may be a portion of or an entire software element. For example, a functional facility may be implemented as a function of a process, or as a discrete process, or as any other suitable unit of processing. If techniques described herein are implemented as multiple functional facilities, each functional facility may be implemented in its own way; all need not be implemented the same way. Additionally, these functional facilities may be executed in parallel and / or serially, as appropriate, and may pass information between one another using a shared memory on the computer(s) on which they are executing, using a message passing protocol, or in any other suitable way.
[0166] Generally, functional facilities include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types.Typically, the functionality of the functional facilities may be combined or distributed as desired in the systems in which they operate. In some implementations, one or more functional facilities carrying out techniques herein may together form a complete software package. These functional facilities may, in alternative embodiments, be adapted to interact with other, unrelated functional facilities and / or processes, to implement a software program application.
[0167] Some exemplary functional facilities are described herein for carrying out one or more tasks. It should be appreciated, though, that the functional facilities and division of tasks described is merely illustrative of the type of functional facilities that may implement theAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTexemplary techniques described herein, and that embodiments are not limited to being implemented in any specific number, division, or type of functional facilities. In some implementations, all functionality may be implemented in a single functional facility. It should also be appreciated that, in some implementations, some of the functional facilities described herein may be implemented together with or separately from others (i.e., as a single unit or separate units), or some of these functional facilities may not be implemented.
[0168] Computer-executable instructions implementing the techniques described herein (when implemented as one or more functional facilities or in any other manner) may, in some embodiments, be encoded on one or more computer-readable media to provide functionality to the media. Computer-readable media include magnetic media such as a hard disk drive, optical media such as a Compact Disk (CD) or a Digital Versatile Disk (DVD), a persistent or non-persistent solid-state memory (e.g., Flash memory, Magnetic RAM, etc.), or any other suitable storage media. Such a computer-readable medium may be implemented in any suitable manner, including as computer-readable storage media 406 of FIG. 4 described below (i.e., as a portion of a computing device 400) or as a stand-alone, separate storage medium. As used herein, “computer-readable media” (also called “computer-readable storage media”) refers to tangible storage media. Tangible storage media are non-transitory and have at least one physical, structural component. In a “computer-readable medium,” as used herein, at least one physical, structural component has at least one physical property that may be altered in some way during a process of creating the medium with embedded information, a process of recording information thereon, or any other process of encoding the medium with information. For example, a magnetization state of a portion of a physical structure of a computer-readable medium may be altered during a recording process.
[0169] In some, but not all, implementations in which the techniques may be embodied as computer-executable instructions, these instructions may be executed on one or more suitable computing device(s) operating in any suitable computer system, or one or more computing devices (or one or more processors of one or more computing devices) may be programmed to execute the computer-executable instructions. A computing device or processor may be programmed to execute instructions when the instructions are stored in a manner accessible to the computing device / processor, such as in a local memory (e.g., an on-chip cache or instruction register, a computer-readable storage medium accessible via a bus, a computer-readable storage medium accessible via one or more networks and accessible by the device / processor, etc.).Functional facilities that comprise these computer-executable instructions may be integrated with and direct the operation of a single multi-purpose programmable digital computer apparatus, aAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTcoordinated system of two or more multi-purpose computer apparatuses sharing processing power and jointly carrying out the techniques described herein, a single computer apparatus or coordinated system of computer apparatuses (co-located or geographically distributed) dedicated to executing the techniques described herein, one or more Field-Programmable Gate Arrays (FPGAs) for carrying out the techniques described herein, or any other suitable system.
[0170] FIG. 4 illustrates one exemplary implementation of a computing device in the form of a computing device 400 that may be used in a system implementing the techniques described herein, although others are possible. It should be appreciated that FIG. 4 is intended neither to be a depiction of necessary components for a computing device to operate in accordance with the principles described herein, nor a comprehensive depiction.
[0171] Computing device 400 may comprise at least one processor 402, a network adapter 404, and computer-readable storage media 406. Computing device 400 may be, for example, a desktop or laptop personal computer, a personal digital assistant (PDA), a smart mobile phone, a server, a wireless access point or other networking element, or any other suitable computing device. Network adapter 404 may be any suitable hardware and / or software to enable the computing device 400 to communicate wired and / or wirelessly with any other suitable computing device over any suitable computing network. The computing network may include wireless access points, switches, routers, gateways, and / or other networking equipment as well as any suitable wired and / or wireless communication medium or media for exchanging data between two or more computers, including the Internet. Computer-readable media 406 may be adapted to store data to be processed and / or instructions to be executed by one or more processors 402. Processor 402 enables processing of data and execution of instructions. The data and instructions may be stored on the computer-readable storage media 406.
[0172] The data and instructions stored on computer-readable storage media 406 may comprise computer-executable instructions implementing techniques which operate according to the principles described herein. In the example of FIG. 4, computer-readable storage media 406 stores computer-executable instructions implementing various facilities and storing various information as described above. Computer-readable storage media 406 may store, for example, health data 412 (e.g., first health data 102, second health data 104, first health data 202, second health data 204, third health data 206, fourth health data 208, etc.), a modeling facility 414 (e.g., a model 100 or any of its components, a model 200 or any of its components, other deep learning models for disease risk phenotyping, a training pipeline for training such models, a scoring pipeline for using such models to perform disease risk phenotyping, etc.), and a recommendation facility 416 (e.g., for recommending actions for an individual based on risk scores). In someAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTexamples, the modeling facility 414 uses different models and optionally different health data of an individual to determine risk scores for the individual for different diseases. In some examples, the modeling facility 414 uses the same model and the same or different health data of an individual to determine risk scores for the individual for different diseases.
[0173] While not illustrated in FIG. 4, a computing device may additionally have one or more components and peripherals, including input and output devices. These devices can be used, among other things, to present a user interface. Examples of output devices that can be used to provide a user interface include printers or display screens for visual presentation of output and speakers or other sound generating devices for audible presentation of output. Examples of input devices that can be used for a user interface include keyboards, and pointing devices, such as mice, touch pads, and digitizing tablets. As another example, a computing device may receive input information through speech recognition or in other audible format.
[0174] Embodiments have been described where the techniques are implemented in circuitry and / or computer-executable instructions. It should be appreciated that some embodiments may be in the form of a method, of which at least one example has been provided. The acts performed as part of the method may be ordered in any suitable way. Accordingly, embodiments may be constructed in which acts are performed in an order different than illustrated, which may include performing some acts simultaneously, even though shown as sequential acts in illustrative embodiments.
[0175] Various aspects of the embodiments described above may be used alone, in combination, or in a variety of arrangements not specifically discussed in the embodiments described in the foregoing and is therefore not limited in its application to the details and arrangement of components set forth in the foregoing description or illustrated in the drawings. For example, aspects described in one embodiment may be combined in any manner with aspects described in other embodiments.Examples
[0176] FIG. 5 shows an ideal Receiver Operating Characteristic (ROC) curve and an ideal Precision-Recall curve, each of which represents the performance of an ideal classifier. The ideal classifier perfectly discriminates between two classes (e.g., individuals for whom the risk of having a particular disease exceeds a threshold risk level, and individuals for whom the risk of having the particular disease does not exceed the threshold risk level). Thus, for the ideal classifier, the number of false positives is zero and the number of false negatives is also zero, yielding a true positive rate (TPR or sensitivity) of 1, a false positive rate (FPR or specificity) ofAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT0, a precision of 1, and a recall of 1. Accordingly, in the example of FIG. 5, the area under the ROC curve (AUROC) is 1.00, and the area under the PRC curve (AUPRC) is 1.00.
[0177] In a first example, FIGS. 6A and 6B show precision vs. recall curves and specificity vs. recall curves, respectively, of a set of risk models for coronary artery disease. The curves 610a and 610b correspond to pooled cohort equations (PCE), the curves 620a and 620b correspond to a first disease risk phenotyping model (e.g., model 100 or 200) that classifies an individual’s risk for coronary artery disease based on data derived from electronic health records (EHR), the curves 630a and 630b correspond to a second disease risk phenotyping model that classifies an individual’s risk for coronary artery disease based on data derived from EHR and data indicative of the individual’s lifestyle and environment, and the curves 640a and 640b correspond to a third disease risk phenotyping model that classifies an individual’s risk for coronary artery disease based on data derived from EHR, data indicative of the individual’s lifestyle and environment, and genetic data (e.g., one or more polygenic risk scores). The models corresponding to curves 620a, 620b, 630a, 630b, 640a, and 640b include generative models (e.g., LLMs) that extract or derive structured health data from unstructured portions of EHRs. As FIGS. 6A and 6B show, each of the disease risk phenotyping models for CAD generally exhibits superior precision and specificity relative to PCE over a wide range of recall values. In addition, the third disease risk phenotyping model for CAD generally exhibits superior precision and specificity relative to the first and second risk models for CAD (and relative to PCE) over the same wide range of recall values.
[0178] In a second example, FIGS. 7A and 7B show precision vs. recall curves and specificity vs. recall curves, respectively, of a set of risk models for breast cancer. The curves 710a and 710b correspond to the Tyrer-Cuzick model, the curves 720a and 720b correspond to a first disease risk phenotyping model (e.g., model 100 or 200) that classifies an individual’s risk for breast cancer based on data derived from electronic health records (EHR), the curves 730a and 730b correspond to a second disease risk phenotyping model that classifies an individual’s risk for breast cancer based on data derived from EHR and data indicative of the individual’s lifestyle and environment, and the curves 740a and 740b correspond to a third disease risk phenotyping model that classifies an individual’s risk for breast cancer based on data derived from EHR, data indicative of the individual’s lifestyle and environment, and genetic data (e.g., one or more polygenic risk scores). The models corresponding to curves 720a, 720b, 730a, 730b, 740a, and 740b include generative models (e.g., LLMs) that extract or derive structured health data from unstructured portions of EHRs. As FIGS. 7A and 7B show, each of the disease risk phenotyping models for breast cancer generally provides superior precision and specificity relative to theAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTTyrer-Cuzick model over a wide range of recall values. In addition, the third disease risk phenotyping model for breast cancer generally exhibits superior precision and specificity relative to the first and second risk models for breast cancer (and relative to Tyrer-Cuzick) over the same wide range of recall values.
[0179] In a third example, FIGS. 8A, 8B, 8C, and 8D illustrate ROC curves for sets of risk models for Alzheimer’s disease, glaucoma, colorectal cancer, and prostate cancer, respectively. Referring to FIG. 8 A, in a third example, the curve 810a corresponds to a random classifier, the curve 820a corresponds to a first disease risk phenotyping model (e.g., model 100 or 200) that classifies an individual’s risk for Alzheimer’s Disease based on data derived from electronic health records (EHR), the curve 830a corresponds to a second disease risk phenotyping model that classifies an individual’s risk for Alzheimer’s Disease based on data derived from EHR and data indicative of the individual’s lifestyle and environment, and the curve 840a corresponds to a third disease risk phenotyping model that classifies an individual’s risk for Alzheimer’s Disease based on data derived from EHR, data indicative of the individual’s lifestyle and environment, and genetic data (e.g., one or more polygenic risk scores). The models corresponding to curves 820a, 830a, and 840a include generative models (e.g., LLMs) that extract or derive structured health data from unstructured portions of EHRs. As FIG. 8 A shows, the AUC values for the first, second, and third risk models for Alzheimer’s Disease are 0.792, 0.826, and 0.846, respectively.
[0180] Referring to FIG. 8B, in a fourth example, the curve 810b corresponds to a random classifier, the curve 820b corresponds to a first disease risk phenotyping model (e.g., model 100 or 200) that classifies an individual’s risk for glaucoma based on data derived from electronic health records (EHR), the curve 830b corresponds to a second disease risk phenotyping model that classifies an individual’s risk for glaucoma based on data derived from EHR and data indicative of the individual’s lifestyle and environment, and the curve 840b corresponds to a third disease risk phenotyping model that classifies an individual’s risk for glaucoma based on data derived from EHR, data indicative of the individual’s lifestyle and environment, and genetic data (e.g., one or more polygenic risk scores). The models corresponding to curves 820b, 830b, and 840b include generative models (e.g., LLMs) that extract or derive structured health data from unstructured portions of EHRs. As FIG. 8B shows, the AUC values for the first, second, and third risk models for glaucoma are 0.675, 0.687, and 0.739, respectively.
[0181] Referring to FIG. 8C, in a fifth example, the curve 810c corresponds to a random classifier, the curve 820c corresponds to a first disease risk phenotyping model (e.g., model 100 or 200) that classifies an individual’s risk for colorectal cancer based on data derived from electronic health records (EHR), the curve 830c corresponds to a second disease riskAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTphenotyping model that classifies an individual’s risk for colorectal cancer based on data derived from EHR and data indicative of the individual’s lifestyle and environment, and the curve 840c corresponds to a third disease risk phenotyping model that classifies an individual’s risk for colorectal cancer based on data derived from EHR, data indicative of the individual’s lifestyle and environment, and genetic data (e.g., one or more polygenic risk scores). The models corresponding to curves 820c, 830c, and 840c include generative models (e.g., LLMs) that extract or derive structured health data from unstructured portions of EHRs. As FIG. 8C shows, the AUC values for the first, second, and third risk models for colorectal cancer are 0.672, 0.687, and 0.732, respectively.
[0182] Referring to FIG. 8D, in a sixth example, the curve 810d corresponds to a random classifier, the curve 820d corresponds to a first disease risk phenotyping model (e.g., model 100 or 200) that operates on data derived from electronic health records (EHR), and the curve 840d corresponds to a second disease risk phenotyping model that operates on data derived from EHR and genetic data (e.g., one or more polygenic risk scores). The models corresponding to curves 820d and 840d include generative models (e.g., LLMs) that extract or derive structured health data from unstructured portions of EHRs. As FIG. 8D shows, the AUC values for the first and second models are 0.695 and 0.781, respectively.
[0183] Referring to FIG. 9, in a seventh example, AUROC and AUPRC values are shown for four sets of disease risk phenotyping models (e.g., models 100 or 200). Each set of risk models includes a risk model trained to classify an individual’s risk for a particular disease or health condition (e.g., Alzheimer’s Disease, stroke, chronic kidney disease, coronary artery disease, glaucoma, prostate cancer, colorectal cancer, ovarian cancer, or lung cancer). The first set of risk models, for which AUROC and AUPRC data are shown in column 910, classifies an individual’s risk for a disease or health condition based on structured data in EHRs and principal components of ethnicity. The second set of risk models, for which AUROC and AUPRC data are shown in column 920, classifies an individual’s risk for a disease or health condition based on structured data in EHRs, principal components of ethnicity, and polygenic risk scores.
[0184] The third set of risk models, for which AUROC and AUPRC data are shown in column 930, classifies an individual’s risk for a disease or health condition based on structured and unstructured data in EHRs and principal components of ethnicity. The fourth set of risk models, for which AUROC and AUPRC data are shown in column 940, classifies an individual’s risk for a disease or health condition based on structured and unstructured data in EHRs, principal components of ethnicity, and polygenic risk scores. As FIG. 9 shows, classifying an individual’s risk for a disease or health condition based on unstructured data from an individual’s EHRsAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTand / or based on an individual’s polygenic risk tends significantly improve AUPRC relative to models that do not use these types of information.
[0185] In a fourth example, a biological sample is obtained from a subject and analyzed to generate sample-derived health data for use by the disease risk phenotyping model. The biological sample may include blood, serum, plasma, urine, sputum, saliva, buccal cells, biopsy tissue, surgical tissue, swab material, or another biological specimen obtained from the subject using a clinically appropriate collection technique. For example, a blood sample may be obtained by venipuncture or capillary collection, a urine sample may be obtained as a clean-catch or catheterized specimen, a sputum sample may be obtained by expectoration or induction, and a tissue sample may be obtained by biopsy, excision, aspiration, or another tissue collection procedure. The biological sample may be associated with subject-identifying information, collection time, collection site, specimen type, collection method, storage condition, transport condition, chain-of-custody information, or other sample metadata, and may be processed in accordance with one or more laboratory protocols suitable for preserving analytes, microorganisms, cells, nucleic acids, proteins, metabolites, or tissue morphology present in the sample.
[0186] In some examples, analyzing the biological sample includes preparing the biological sample for one or more laboratory assays and performing the one or more laboratory assays to generate assay results. Sample preparation may include centrifugation, aliquoting, dilution, concentration, filtration, fixation, staining, embedding, sectioning, extraction of nucleic acids, extraction of proteins, extraction of metabolites, cell separation, microbial culture preparation, or other preprocessing steps selected based on the specimen type and assay to be performed. The one or more laboratory assays may include a hematology assay, chemistry assay, immunoassay, microbiology assay, culture assay, antimicrobial susceptibility assay, polymerase chain reaction assay, sequencing assay, genotyping assay, proteomic assay, metabolomic assay, cytology assay, histopathology assay, or other diagnostic or analytical assay. The assay results may include one or more measured analyte values, detected organism identifiers, microbial load values, susceptibility results, sequence alterations, expression levels, cell counts, morphological findings, pathology findings, inflammatory markers, infection markers, biomarker values, or qualitative, semi-quantitative, or quantitative findings derived from the biological sample.
[0187] In some examples, the assay results and sample metadata are converted into structured health data for input to the disease risk phenotyping model. The structured health data may include numerical values, categorical values, ordinal values, binary indicators, coded clinical concepts, time-stamped results, longitudinal sample-derived measurements, or confidence valuesAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTassociated with the assay results. In some examples, the sample-derived structured health data are normalized to one or more reference ranges, units, vocabulary codes, organism taxonomies, sequence-alteration representations, or disease-specific feature definitions. The sample-derived structured health data may be encoded by an encoder or by one of the subnetworks of the disease risk phenotyping model to generate sample-derived feature encodings, and the sample-derived feature encodings may be supplied to a merge layer for combination with embeddings or encodings generated from electronic health record data, unstructured clinical text, polygenic risk score data, biomarker data, imaging data, lifestyle data, environmental data, temporal data, or other health data of the subject.Additional Terminology and Embodiments
[0188] The terminology used herein is for the purpose of description and should not be regarded as limiting.
[0189] The term “approximately”, the phrase “approximately equal to”, and other similar phrases, as used in the specification and the claims (e.g., “X has a value of approximately Y” or “X is approximately equal to Y”), should be understood to mean that one value (X) is within a predetermined range of another value (Y). The predetermined range may be plus or minus 20%, 10%, 5%, 3%, 1%, 0.1%, or less than 0.1%, unless otherwise indicated.
[0190] Measurements, sizes, amounts, etc. may be presented herein in a range format. The description in range format is merely for convenience and brevity and should not be construed as an inflexible limitation on the scope of the invention. Accordingly, the description of a range should be considered to have specifically disclosed all the possible subranges as well as individual numerical values within that range. For example, description of a range such as 10-20 inches should be considered to have specifically disclosed subranges such as 10-11 inches, 10-12 inches, 10-13 inches, 10-14 inches, 11-12 inches, 11-13 inches, etc.
[0191] The indefinite articles “a” and “an,” as used in the specification and in the claims, unless clearly indicated to the contrary, should be understood to mean “at least one.” The phrase “and / or,” as used in the specification and in the claims, should be understood to mean “either or both” of the elements so conjoined, i.e., elements that are conjunctively present in some cases and disjunctively present in other cases. Multiple elements listed with “and / or” should be construed in the same fashion, i.e., “one or more” of the elements so conjoined. Other elements may optionally be present other than the elements specifically identified by the “and / or” clause, whether related or unrelated to those elements specifically identified. Thus, as a non-limiting example, a reference to “A and / or B”, when used in conjunction with open-ended language suchAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTas “comprising” can refer, in one embodiment, to A only (optionally including elements other than B); in another embodiment, to B only (optionally including elements other than A); in yet another embodiment, to both A and B (optionally including other elements); etc.
[0192] As used in the specification and in the claims, “or” should be understood to have the same meaning as “and / or” as defined above. For example, when separating items in a list, “or” or “and / or” shall be interpreted as being inclusive, i.e., the inclusion of at least one, but also including more than one, of a number or list of elements, and, optionally, additional unlisted items. Only terms clearly indicated to the contrary, such as “only one of or “exactly one of,” or, when used in the claims, “consisting of,” will refer to the inclusion of exactly one element of a number or list of elements. In general, the term “or” as used shall only be interpreted as indicating exclusive alternatives (i.e. “one or the other but not both”) when preceded by terms of exclusivity, such as “either,” “one of,” “only one of,” or “exactly one of.” “Consisting essentially of,” when used in the claims, shall have its ordinary meaning as used in the field of patent law.
[0193] As used in the specification and in the claims, the phrase “at least one,” in reference to a list of one or more elements, should be understood to mean at least one element selected from any one or more of the elements in the list of elements, but not necessarily including at least one of each and every element specifically listed within the list of elements and not excluding any combinations of elements in the list of elements. This definition also allows that elements may optionally be present other than the elements specifically identified within the list of elements to which the phrase “at least one” refers, whether related or unrelated to those elements specifically identified. Thus, as a non-limiting example, “at least one of A and B” (or, equivalently, “at least one of A or B,” or, equivalently “at least one of A and / or B”) can refer, in one embodiment, to at least one, optionally including more than one, A, with no B present (and optionally including elements other than B); in another embodiment, to at least one, optionally including more than one, B, with no A present (and optionally including elements other than A); in yet another embodiment, to at least one, optionally including more than one, A, and at least one, optionally including more than one, B (and optionally including other elements); etc.
[0194] The use of “including,” “comprising,” “having,” “containing,” “involving,” and variations thereof, is meant to encompass the items listed thereafter and additional items.
[0195] Use of ordinal terms such as “first,” “second,” “third,” etc., in the claims to modify a claim element does not by itself connote any priority, precedence, or order of one claim element over another or the temporal order in which acts of a method are performed. Ordinal terms are used merely as labels to distinguish one claim element having a certain name from another element having a same name (but for use of the ordinal term), to distinguish the claim elements.Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT
[0196] While this specification contains many specific implementation details, these should not be construed as limitations on the scope of what may be claimed, but rather as descriptions of features that may be specific to particular embodiments. Certain features that are described in this specification in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.
[0197] Similarly, while operations are depicted in the drawings in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. In certain circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system components in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.
[0198] Particular embodiments of the subject matter have been described. Other embodiments are within the scope of the following claims. For example, the actions recited in the claims can be performed in a different order and still achieve desirable results. As one example, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In certain implementations, multitasking and parallel processing may be advantageous. Other steps or stages may be provided, or steps or stages may be eliminated, from the described processes. Accordingly, other implementations are within the scope of the following claims.
Claims
Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTCLAIMSWhat is claimed is:
1. A method for determining whether a subject is at increased risk for a disease, the method comprising:generating, by one or more processors, an integrated risk score (IRS) for a subject, wherein the IRS indicates a risk of a particular disease or disease group for the subject, and wherein generating the IRS includes (1) extracting, by one or more first neural network models, values of a plurality of features from health data of the subject, and (2) classifying, by one or more second models and based on the values of the plurality of features, the risk of the particular disease or disease group for the subject;determining one or more recommended actions for the subject based at least in part on the IRS; andproviding an indication of the one or more recommended actions for the subject.
2. The method of claim 1, wherein the one or more first neural network models include an autoencoder, and wherein the one or more second models include a classifier.
3. The method of claim 2, wherein the autoencoder and the classifier are jointly trained based on training data including, for each individual of a plurality of individuals, (1) health data of the respective individual and (2) label data indicating the risk of the particular disease or disease group for the individual.
4. The method of claim 2, wherein the IRS for the subject is a first IRS for the subject, wherein the particular disease or disease group is a first particular disease or disease group, wherein the plurality of features is a plurality of first features, wherein the health data of the subject are first health data of the subject, wherein the one or more recommended actions are one or more first recommended actions, wherein the autoencoder is a first autoencoder, wherein the classifier is a first classifier, and wherein the method further includes:generating, by one or more processors, a second integrated risk score (IRS) for the subject, wherein the second IRS indicates a risk of a second particular disease or disease group for the subject, and wherein generating the second IRS includes (1) extracting, by a second autoencoder, values of a plurality of second features from second health data of the subject, and (2) classifying, by a second classifier and based on the values of the plurality of second features, the risk of the second particular disease or disease group for the subject;Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTdetermining one or more second recommended actions for the subject based at least in part on the second IRS; andproviding an indication of the one or more second recommended actions for the subject.
5. The method of claim 2, wherein the IRS for the subject is a first IRS for the subject, wherein the particular disease or disease group is a first particular disease or disease group, wherein the plurality of features is a plurality of first features, wherein the one or more recommended actions are one or more first recommended actions, and wherein the method further includes:generating, by one or more processors, a second integrated risk score (IRS) for the subject, wherein the second IRS indicates a risk of a second particular disease or disease group for the subject, and wherein generating the second IRS includes (1) extracting, by the autoencoder, values of a plurality of second features from the health data of the subject, and (2) classifying, by the classifier and based on the values of the plurality of second features, the risk of the second particular disease or disease group for the subject;determining one or more second recommended actions for the subject based at least in part on the second IRS; andproviding an indication of the one or more second recommended actions for the subject.
6. The method of claim 2, wherein an encoder portion of the autoencoder is an electronic health record (EHR) encoder.
7. The method of claim 6, wherein the EHR encoder includes a transformer.
8. The method of claim 1, wherein the risk indicated by the IRS is a first risk, and wherein the health data include a polygenic risk score (PRS) for the subject, wherein the PRS is based on a genotype of the subject and indicates a second risk of the particular disease or disease group for the subject.
9. The method of claim 8, further comprising generating the PRS for the subject, wherein generating the PRS for the subject includes applying a polygenic model to the genotype of the subject, wherein the polygenic model is associated with the particular disease or disease group.Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT10. The method of claim 1, wherein the health data include values of one or more biomarkers for the subject, wherein the one or more biomarkers are associated with the particular disease or disease group.
11. The method of claim 8, wherein the health data further include principal components of ethnicity of the subject.
12. The method of claim 8, wherein the health data further include structured data derived from one or more electronic health records of the subject.
13. The method of claim 12, wherein the structured data include a plurality of fields, and wherein values of the plurality of fields are one-hot encoded.
14. The method of claim 12, further comprising generating, by one or more generative models, a portion of the structured data based on unstructured data in the one or more electronic health records.
15. The method of claim 14, wherein the one or more generative models include a large language model (LLM).
16. The method of claim 12, wherein the health data include one or more diagnostic codes indicating diagnoses of diagnosed health conditions of the subject; treatment records of the subject; one or more procedure codes indicating procedures administered to the subject; demographic data of the subject; admission, discharge and transfer (ADT) records of the subject, results of laboratory tests administered to the subject, vital signs of the subject, diagnostic data related to diagnosed and / or undiagnosed health conditions of the subject; encounter data related to encounters between the subject and one or more healthcare providers; clinical records related to health or treatment of the subject; allergy data of the subject; surgery data of the subject; immunization data of the subject; a family medical history of the subject; a social history of the subject; one or more radiology reports corresponding to radiological scans of the subject; one or more medical images of the subject; medication data of the subject; results of one or more cardiac health tests of the subject; microbiology data obtained from analysis of one or more samples of blood, urine, sputum, or tissue of the subject; pathology data obtained from analysis of one or more samples of blood, urine, sputum, or tissue of the subject; infusion data related toAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTone or more substances infused into a bloodstream of the subject; and / or treatment data related to one or more treatments administered to the subject.
17. The method of claim 16, wherein the health data further include physiological monitoring data obtained from one or more wearable sensors, wherein the physiological monitoring data include heart rate data, blood oxygen saturation data, activity data, sleep data, electrodermal activity data, and / or continuous glucose data.
18. The method of claim 1, wherein the health data further include medical imaging data.
19. The method of claim 1, wherein the particular disease is breast cancer, lung cancer, prostate cancer, colorectal cancer, pancreatic cancer, ovarian cancer, endometrial cancer, cervical cancer, liver cancer, gastroesophageal cancer, small intestine cancer, bladder cancer, brain cancer, kidney cancer, thyroid cancer, leukemia, multiple myeloma, lymphoma (both Hodgkin and non-Hodgkin), melanoma, head and neck cancer, obesity, coronary artery disease (CAD), type 2 diabetes (T2D), stroke, Alzheimer’s disease (AD), glaucoma, or age-related macular degeneration (AMD).
20. The method of claim 1, wherein the one or more first neural network models include: an encoder configured to provide first embeddings of first features of the plurality of features based on a first portion of the health data of the subject;one or more subnetworks configured to provide second embeddings of second features of the plurality of features based on respective second portions of the health data of the subject; and a fusion layer configured to provide a unified representation of the plurality of features based on the first embeddings and the second embeddings.
21. The method of claim 20, wherein classifying of the risk by the one or more second models is based on the unified representation of the plurality of features.
22. The method of claim 20, wherein the encoder comprises a portion of an autoencoder or an EHR encoder.
23. The method of claim 20,Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTwherein the encoder, the one or more subnetworks, the fusion layer, and the one or more second models are arranged in a fusion neural network,wherein the first portion of the health data include structured data,wherein at least one second portion of the health data includes unstructured text of electronic health records of the subject,wherein the second features include one or more unstructured text features based on the unstructured text of electronic health records of the subject,wherein a first subnetwork of the one or more subnetworks includes one or more generative models,wherein the one or more second models include a classifier, andwherein the method further includes:generating, by the one or more generative models, the one or more unstructured text features; andproviding, by the first subnetwork, one or more second embeddings of the one or more unstructured text features.
24. The method of claim 23, wherein the second features further include a polygenic risk score (PRS) of the subject, a numeric biomarker value of the subject, one or more lab results of the subject, one or more lifestyle factors of the subject, and / or one or more environmental factors of the subject.
25. The method of claim 23, wherein the first features include a polygenic risk score (PRS) of the subject, a numeric biomarker value of the subject, one or more lab results of the subject, one or more lifestyle factors of the subject, and / or one or more environmental factors of the subject.
26. The method of claim 23, wherein the one or more generative models include a large language model (LLM) comprising a generative transformer-based network trained to embed free-text clinical notes.
27. The method of claim 23, wherein the encoder and the one or more subnetworks are jointly trained in a joint optimization process.
28. The method of claim 27, further comprising:Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTapplying, by the fusion layer, a gating mechanism or attention mechanism that weights contributions from the one or more second embeddings and contributions from the first embeddings differently based on domain context.
29. The method of claim 27, wherein the joint optimization process comprises a hybrid quantum-classical optimization process in which a quantum processor evaluates at least one quantum circuit associated with a parameter update, a loss function component, a feature interaction, an attention mechanism, or an embedding transformation.
30. The method of claim 29, wherein measurement results from the quantum processor are used to update one or more parameters of the encoder, the one or more subnetworks, the fusion layer, or the classifier.
31. The method of claim 29, wherein the encoder comprises a transformer configured to model temporal, semantic, or clinical relationships among features extracted from electronic health records, and wherein the hybrid quantum-classical optimization process determines one or more attention weights, query projections, key projections, value projections, temporal encodings, or feed-forward layer parameters of the transformer.
32. The method of claim 29, wherein at least one of the one or more subnetworks is configured to generate a polygenic risk embedding based on genetic-risk data of the subject, and wherein generating the polygenic risk embedding comprises applying a quantum-assisted model to genotype data, variant data, single nucleotide polymorphism data, variant effect sizes, ancestry-adjusted variant weights, disease-associated loci, one or more polygenic risk scores, or a combination thereof.
33. The method of claim 32, wherein the quantum-assisted model is configured to estimate nonlinear interactions among genetic variants, interactions among polygenic risk components, interactions between genetic-risk features and clinical features, or interactions between genetic-risk features and environmental or lifestyle features.
34. The method of claim 32, wherein the polygenic risk embedding is combined, by the fusion layer, with an embedding generated by a second encoder, wherein the second encoder is transformer-based and is configured to encode one or more electronic health records.Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT35. The method of claim 34, wherein the fusion layer learns disease-specific interactions between longitudinal clinical features represented by the embedding generated by the second encoder and inherited genetic-risk features represented by the polygenic risk embedding.
36. The method of claim 29, wherein the quantum processor comprises a gate-based quantum processor, a quantum annealer, a photonic quantum processor, an ion-trap quantum processor, a superconducting quantum processor, or a simulated quantum processor.
37. The method of claim 1, wherein the one or more recommended actions for the subject include undergoing a screening test, a blood test, lab work, and / or a medical imaging procedure.
38. The method of claim 1, wherein generating the integrated risk score comprises:determining a missingness characteristic of electronic health record data of the subject; determining, based at least in part on the missingness characteristic, one or more weights for genetic-risk data of the subject and electronic health record data of the subject; and generating the integrated risk score by combining a genetic-risk component derived from the genetic-risk data and a clinical-risk component derived from the electronic health record data according to the one or more weights.
39. The method of claim 38, wherein the missingness characteristic comprises an amount of missing electronic health record data, a pattern of missing electronic health record data, a type of missing electronic health record data, cohort-specific missingness, feature-specific missingness, or combinations thereof.
40. A method comprising:(a) obtaining first embeddings of unstructured health data of a subject, wherein obtaining the first embeddings includes applying a large language model (LLM) to at least a portion of the unstructured health data;(b) obtaining second embeddings of structured health data of the subject, wherein obtaining the second embeddings includes applying an encoder to at least a portion of the structured health data, and wherein the structured health data include at least one polygenic risk score (PRS) of the subject, wherein the encoder comprises (1) a portion of an autoencoder or (2) a transformer;Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT(c) combining the first embeddings and the second embeddings in a fusion layer of a deep neural network, thereby producing a unified representation indicative of a risk of a particular disease or disease group;(d) classifying, based on the unified representation, the risk of the particular disease or disease group for the subject;(e) identifying a therapy for the subject based at least in part on the classified risk; and (f) providing a recommendation that the therapy be administered to the subject to reduce the risk of the particular disease or disease group.
41. A computer-implemented method for training a disease risk model, the method comprising:obtaining training data comprising health data for a plurality of individuals and label data indicating disease labels for the plurality of individuals;generating, by one or more neural network models, a plurality of embeddings of the health data, wherein the health data comprise at least two types of health data, wherein the two types of health data include electronic health record data, unstructured clinical text, polygenic risk score data, biomarker data, medical imaging data, lifestyle data, environmental data, and / or principal components of ethnicity;combining the plurality of embeddings using a fusion layer to produce a plurality of unified representations for the plurality of individuals;evaluating, using a hybrid quantum-classical optimization process, a loss function associated with predictions generated based on the plurality of unified representations;updating one or more parameters of the one or more neural network models based at least in part on output of the hybrid quantum-classical optimization process; andstoring the updated one or more parameters for subsequent use in generating a disease risk score for a subject.
42. The method of claim 41, wherein the disease labels indicate disease status, disease risk, and / or disease outcome for one or more diseases for the plurality of individuals.
43. A system for determining whether a subject is at increased risk for a disease, the system comprising:one or more processors; andAttorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCTa computer-readable storage medium storing instructions that, when executed by the one or more processors, cause the system to perform operations including:(a) obtaining first embeddings of unstructured health data of the subject, wherein obtaining the first embeddings includes applying a large language model (LLM) to at least a portion of the unstructured health data;(b) obtaining second embeddings of structured health data of the subject, wherein obtaining the second embeddings includes applying one or more encoders to one or more portions of the structured health data;(c) obtaining third embeddings of medical image data of the subject, wherein obtaining the third embeddings includes extracting, by a first neural network, one or more features from the medical image data and encoding, by an encoder, the one or more features, thereby producing the third embeddings;(d) combining the first embeddings, the second embeddings, and the third embeddings, in a fusion layer of a deep neural network, thereby producing a unified representation indicative of a risk of a particular disease or disease group;(e) classifying, based on the unified representation, the risk of the particular disease or disease group for the subject;(f) determining one or more recommended actions for the subject based at least in part on the classified risk; and(g) providing an indication of the one or more recommended actions.
44. The system of claim 43, wherein the structured health data include a polygenic risk score (PRS) of the subject, a numeric biomarker value of the subject, one or more lab results of the subject, one or more lifestyle factors of the subject, and / or one or more environmental factors of the subject.
45. The system of claim 44, wherein the one or more lab results include microbiology data obtained from analysis of a sample of blood, urine, sputum, or tissue of the subject, and wherein obtaining the microbiology data includes:obtaining the sample of blood, urine, sputum, or tissue of the subject; andanalyzing the sample of blood, urine, sputum, or tissue of the subject.Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT46. The system of claim 44, wherein providing the indication of the one or more recommended actions includes providing a recommendation that a therapy be administered to the subject to reduce the risk of the particular disease or disease group.
47. The system of claim 43, wherein the first neural network comprises a convolutional neural network (CNN), a vision transformer (ViT), or a pretrained image foundation model.
48. A method comprising:(a) obtaining first embeddings of unstructured health data of a subject, wherein obtaining the first embeddings includes applying a large language model (LLM) to at least a portion of the unstructured health data;(b) obtaining second embeddings of second health data of the subject, wherein obtaining the second embeddings includes applying an encoder to at least a portion of the second health data, wherein the second health data include structured health data, and wherein the structured health data include at least one polygenic risk score (PRS) of the subject;(c) combining the first embeddings and the second embeddings in a fusion layer of a deep neural network, thereby producing a unified representation indicative of a risk of a particular disease or disease group;(d) classifying, based on the unified representation, the risk of the particular disease or disease group for the subject;(e) determining one or more recommended actions for the subject based at least in part on the classified risk; and(f) providing an indication of the one or more recommended actions,wherein the encoder comprises (1) a portion of an autoencoder or (2) a transformer.
49. The method of claim 48, wherein applying the LLM to at least the portion of the unstructured health data comprises prompting the LLM to extract one or more first features from at least the portion of the unstructured health data.
50. The method of claim 48, wherein the LLM, encoder, and one or more subnetworks of the deep neural network are jointly trained or fine-tuned in a joint optimization process.
51. The method of claim 50, wherein the joint optimization process minimizes both reconstruction error and classification loss.Attorney Docket No. 206979-720601 / PCTClient Ref. No. MYOM-020PCT52. The method of claim 50, wherein the fusion layer applies a gating mechanism or attention mechanism that weights contributions from the first embeddings and the second embeddings differently based on domain context.
53. The method of claim 48, wherein the LLM is a generative transformer-based network configured to embed free-text clinical notes, and wherein the encoder is configured to embed one or more numeric biomarker values of the subject, one or more lab results of the subject, one or more lifestyle factors of the subject, one or more environmental factors of the subject, and / or the at least one polygenic risk score (PRS) for the subject.
54. The method of claim 48, wherein the encoder comprises a portion of an autoencoder.
55. The method of claim 48, wherein the encoder is an electronic health record (EHR) encoder.
56. The method of claim 48, wherein the encoder includes a transformer.
57. The method of claim 48, wherein the second health data further include medical imaging data.
58. A system comprising:one or more processors; anda computer-readable storage medium storing instructions that, when executed by the one or more processors, cause the system to perform the method of any of claims 1-57.
59. A computer program product comprising at least one non-transitory computer-readable storage medium storing instructions that, when executed, cause a system to perform the method of any of claims 1-57.
