Graph-based mental health condition prediction

Abstract

Aspects of the present disclosure relate to graph-based mental health condition prediction. In examples, a combination of clinical data and neuroimaging outcomes for a given subject are used to populate a graph data store (e.g., comprising a set of nodes and associated edges corresponding to the subject). A graph neural network (GNN) is used to evaluate associations within the graph data store and / or changes to such associations over time to predict the incidence of BD, MDD, and / or DMDD, among other mental health conditions. Such a GNN thus combines the power of defined data relationships within the graph data store with the pattern detection capabilities of deep learning. In this way, the GNN identifies relationships within the data (e.g., as are embodied in within the graph data store) and feature combinations, thus using clinical data and neuroimaging outcomes to discover relationships that are predictive of such health mental conditions.

Description

GRAPH-BASED MENTAL HEALTH CONDITION PREDICTION CROSS-REFERENCE TO RELATED APPLICATION

[0001] This application claims priority to U.S. Provisional Application No. 63 / 560,374, titled “Graph-Based Mental Health Condition Prediction,” filed on March 1, 2024, the entire disclosure of which is hereby incorporated by reference in its entirety.BACKGROUND

[0002] Bipolar disorder (BD) is among the leading causes of disability worldwide. However, the initial diagnosis of BD is often delayed, which may lead to detrimental effects on its course and increase the burden on afflicted individuals, families, and society. Thus, early detection of BD (e.g., prior to the onset of mania) may have a positive effect on patients' lives and may yield novel insights into underlying neurobiological mechanisms. Moreover, identifying those at very high risk of developing BD will pave the road for developing innovative prevention strategies. Similar benefits may be observed for the early detection of major depressive disorder (MDD) and / or disruptive mood dysregulation disorder (DMDD), among other examples.

[0003] It is with respect to these and other general considerations that embodiments have been described. Also, although relatively specific problems have been discussed, it should be understood that the embodiments should not be limited to solving the specific problems identified in the background.SUMMARY

[0004] Aspects of the present disclosure relate to graph-based mental health condition prediction. In examples, a combination of clinical data and neuroimaging outcomes for a given subject are used to populate a graph data store (e.g., comprising a set of nodes and associated edges corresponding to the subject). A graph neural network (GNN) is used to evaluate associations within the graph data store and / or changes to such associations over time to predict the incidence of BD, MDD, and / or DMDD, among other mental health conditions. Such a GNN thus combines the power of defined data relationships within the graph data store with the pattern detection capabilities of deep learning. In this way, the GNN identifies relationships within the data (e.g., as are embodied in within the graph data store) and feature combinations, thus using clinical data andneuroimaging outcomes to discover relationships that are predictive of such mental health conditions.

[0005] This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Non-limiting and non-exhaustive examples are described with reference to the following Figures.

[0007] Figure 1 illustrates an overview of an example system in accordance with aspects described herein.

[0008] Figure 2 illustrates an example graph subpart that may be used to train a condition prediction model and / or to generate a prediction according to aspects described herein.

[0009] Figure 3 illustrates an overview of an example method for generating a set of nodes and edges for a subject within a graph data store according to aspects described herein.

[0010] Figure 4A illustrates an overview of example method for training a machine learning model to predict a mental health condition according to aspects described herein.

[0011] Figure 4B illustrates an overview of an example method for generating a prediction for a subject according to aspects described herein.

[0012] Figure 5 illustrates an example of a suitable operating environment in which one or more aspects of the present application may be implemented.DETAILED DESCRIPTION

[0013] In the following detailed description, references are made to the accompanying drawings that form a part hereof, and in which are shown by way of illustrations specific embodiments or examples. These aspects may be combined, other aspects may be utilized, and structural changes may be made without departing from the present disclosure. Embodiments may be practiced as methods, systems or devices. Accordingly, embodiments may take the form of a hardware implementation, an entirely software implementation, or an implementation combining software and hardware aspects. The following detailed description is therefore not to be taken in a limiting sense, and the scope of the present disclosure is defined by the appended claims and their equivalents.

[0014] The early identification of youth who will later develop bipolar disorder (BD), major depressive disorder (MDD), disruptive mood dysregulation disorder (DMDD), and / or other mental health conditions remains a critical public health goal, for example to allow intervention earlier in the progression of such conditions. While combinations of clinical history and symptoms, or clinical-risk status, may offer good predictive value in certain circumstances (e.g., in research samples enriched for risk of BD), the predictive validity of clinical risk status for future conditions does not generalize well. In parallel, magnetic resonance (MR) brain imaging may permit the identification of functional and / or structural signatures for certain conditions (e.g., BD) in those with the condition as compared to those without any mental health condition. However, like the limitations of clinical assessment, current neuroimaging prediction models are good only when discriminating between individuals with a condition and individuals with no mental health condition. Thus, like clinical-only prediction algorithms, such techniques may not adequately distinguish between individuals with BD or other health conditions from individuals without any mental health conditions.

[0015] Accordingly, aspects of the present disclosure relate to graph-based mental health condition prediction. In examples, a combination of clinical data (e.g., clinical history, symptoms, and / or clinical-risk status) and neuroimaging outcomes (e.g., structural and / or resting-state functional imaging outcomes) for a given subject are used to populate a graph data store (e.g., comprising a set of nodes and associated edges corresponding to the subject). A graph neural network (GNN) is used to evaluate associations within the graph data store and / or changes to suchassociations over time to predict the incidence of BD, MDD, and / or DMDD, among other mental health conditions. Such a GNN thus combines the power of defined data relationships within the graph data store with the pattern detection capabilities of deep learning. In this way, the GNN identifies relationships within the data (e.g., as are embodied in within the graph data store) and feature combinations, thus using clinical data and neuroimaging outcomes to discover relationships that are predictive of such health mental conditions.

[0016] In addition to generating a concurrent and / or future prediction for the incidence of a mental health condition, aspects described herein may generate an “explainability indication” that corresponds to a concurrent and / or future prediction. As an example, the explainability indication includes one or more features that explain a generated prediction for a subject and / or for a group, thereby indicating to a user (e.g., to a clinician) that the features were determined to be indicative of or otherwise relevant to the generated prediction. In examples, the explainability indication includes one or more weights that were applied to the indicated features, among other alternative or additional information to provide insight as to how the prediction was generated (e.g., from clinical data and / or neuroimaging outcomes) according to aspects described herein. As a result of providing an explainability metric, aspects of the present disclosure may be used as a clinical decision support tool and may thus provide tailored guidance (e g., for a given subject) for the diagnosing clinician and / or may provide more generalizable indicators of pathophysiology (e.g., for a group).

[0017] Thus, aspects described herein identify predictive structural connectivity (diffusion weighted imaging metrics) and functional connectivity (global connectivity and regional homogeneity) from neuroimaging outcomes, as well as predictive item-wise responses from parent and child standardized interviews and behavioral scales clinical data, that have concurrent and future predictive associations to BD, MDD, and / or DMDD to generate predictions for a given subject or group of subjects.

[0018] Specific clinical phenotypes may have predictive validity for the emergence of such mental health conditions. For example, subthreshold manic symptoms and a family history of mania may be used, where the latter is consistent with the high heritability of BD. However, the predictive power of the risk phenotype may degrade when applied to a mixed clinical sample. Moreover, these indicators may not allow for the identification of high-risk states specific to BD relative toother mood and psychotic disorders. Even so, aspects described herein use the predictive power of self- and / or caregiver-reported manic symptoms, especially subthreshold manic symptoms that arise over a sufficiently long period (e.g., on the scale of years), for early detection (e.g., before meeting full criteria for BD).

[0019] With respect to neuroimaging outcomes, connectivity metrics derived from structural and resting-state functional imaging outcomes may similarly have predictive validity. For example, current etiological theories of BD note the role of atypical developmental trajectories of white matter microstructure proposing both atypical axonal pruning and myelination as key mechanisms. Such atypical white matter microstructure may affect the function of neural networks, as evidenced by more than 35 reports of widespread aberrancies in intrinsic resting-state connectivity in BD. Indeed, emerging empirical evidence supports the predictive value of features derived from both structural and functional connectivity outcomes for concurrent BD, such that these metrics may thus be leveraged to predict the emergence of BD and / or other mental health conditions.

[0020] The disclosed aspects may be especially useful for younger subjects (e.g., adolescents), where diagnosis of such mental health conditions is challenging due to the limited time frame for which clinical data is available for such subjects. Indeed, as noted above, generating a prediction for a future incidence of a mental health condition may enable a subject to receive improved treatment earlier in the progression of the condition (e.g., prior to the onset of clinical symptoms). Further, in addition to such longitudinal predictions, concurrent predictions are also valuable in both the clinic and research contexts. For example, most affective episodes in BD, including the first one, are depressive, leading to frequent misdiagnosis of unipolar depression. Such confusion may thus lead to misaligned clinical care, which may result in negative long-term consequences and / or the further potential to confound research findings.

[0021] It will be appreciated that classification according to a graph neural network is provided as an example classifier according to the disclosed aspects. In other examples, a graph data store may be processed using any of a variety of additional or alternative classifiers (also referred to herein as a “condition prediction model”). For instance, a regression model is fit to a training dataset, such that the regression model is usable to process a graph data store to generate a risk score accordingly. Thus, a risk score generated by a regression model may be indicative of a probability (e.g., a concurrent prediction and / or a future prediction) for a given diagnosis.

[0022] Figure 1 illustrates an overview of an example system 100 in accordance with aspects described herein. As illustrated, system 100 comprises condition prediction platform 102, computing device 104, imaging device 106, and network 108. In examples, condition prediction platform 102, computing device 104, and imaging device 106 communicate via network 108. For example, network 108 may comprise a local area network, a wireless network, or the Internet, or any combination thereof, among other examples.

[0023] In examples, imaging device 106 is a magnetic resonance imaging (MRI) device that is used to acquire neuroimaging data of a subject, which may be ingested (e.g., by condition prediction platform 102) according to aspects described herein. It will be appreciated that condition prediction platform 102 may ingest neuroimaging outcomes from any of a variety of other sources, including, but not limited to, computing device 104 (e.g., as may be the case when computing device 104 acquires data from an imaging device connected thereto) or as part of a third-party dataset, among other examples.

[0024] Computing device 104 communicates with condition prediction platform 102, for example to provide clinical data and / or neuroimaging outcomes to be ingested by condition prediction platform 102. For example, a subject may use application 118 of computing device 104 to complete a survey and / or provide any of a variety of clinical data. In other examples, application 118 obtains a prediction generated by condition prediction platform 102 and presents the generated prediction to a user (e.g., a clinician or) accordingly.

[0025] Condition prediction platform 102 may include one or more server computing devices and is illustrated as comprising graph manager 110, graph data store 112, machine learning engine 114, and condition prediction model 116. In examples, graph manager 110 manages the content of graph data store 112. For example, graph manager 110 processes data associated with a subject (e.g., neuroimaging outcomes and / or clinical data) and generates a set of nodes and associated edges corresponding to the data within graph data store 112. An example graph subpart corresponding to a subject is discussed below with respect to Figure 2.

[0026] For instance, predictive features for a given subject may be organized within one or more nodes within graph data store 112 that are linked within a subject node, such that they may be used to generate a concurrent and / or future prediction according to aspects described herein. Example features include, but are not limited to demographics (e.g., age, sex, race, ethnicity, householdincome, highest parent education, and / or parental marital status), social determinants of health, clinical history (e.g., neurocognitive picture vocabulary score, family history of mania symptoms, anxiety disorder diagnosis, 10-item parent-report mania scale, seven-item youth-report mania scale, and / or history of trauma), symptoms, neurobiology, and / or neuroimaging outcomes (e.g., diffusion tensor imaging (DTI) and / or echoplanar imaging (EPI) metrics, as may be generated from MRI), among other examples. Such nodes may be associated with a subject node for the subject to which the data relates, thereby forming a subpart of the graph for the subject. As an example, a subject node includes demographic information and, in some examples, a neurocognitive picture vocabulary score for a given subject. The subject node has a set of associated nodes, for example including a family history node, a trauma / anxiety node, a mania scale node (e.g., including information relating to a parent- and / or youth-report mania scale), and a neuroimaging metrics node.

[0027] Additionally, or alternatively, graph manager 110 may nest nodes and graph types of graph data store 112 within superordinate nodes. For example, the site of participation (e.g., at which a clinical data is obtained for a given subject) may be represented as a nested feature within the demographic node for the subject within graph data store 112. This may thus enable the site to interact with multiple associated features within graph data store 112, such that features may be adjusted as a result of such an association. Similarly, data within graph data store 112 may be nested by time, which may thus be superordinate to site, while site may be superordinate to subjectlevel nodes.

[0028] In some examples, relationships between nodes may be represented as one or more features within edges of graph data store 112. For example, a relationship between one node and another node may be weighted according to a predefined structure. As an example, rather than representing time as a superordinate node, time may instead by represented as one or more edge features between the same set of nodes. Thus, time may be constrained and ordered for an estimate of the amplification or decay of a specific feature’s predictive power over time. It will be appreciated that additional or alternative features may be defined in other examples. For instance, in instances where a large, site-specific effect is identified on neuroimaging outcomes, a site edge feature may be defined between a subject and one or more associated neuroimaging nodes. Thus, it may be possible to better account for subject-site specific influence on the predictive value of neuroimaging outcomes.

[0029] Condition prediction platform 102 is further illustrated as comprising machine learning engine 114, which trains condition prediction model 116 according to data within graph data store 112. As another example, machine learning engine 114 uses condition prediction model 116 to process a graph subpart of graph data store 112 (e.g., that corresponds to a given subject) to generate a prediction for the subject according to aspects described herein. For example, machine learning engine 114 uses clinical and neuroimaging outcomes (e.g., as may be stored by graph data store 112) to predict the emergence of one or more mental conditions. In examples, the generated prediction may instead indicate no diagnosis.

[0030] Data used to train condition prediction model 116 may exclude subjects with either positive drug screens or those failing a DTI quality control check. For instance, graph manager 110 may omit such data, such that graph data store 112 does not include data associated with such subjects. Further, a group of subjects diagnosed with a mental condition may be identified and used to train condition prediction model 116 accordingly. For example, a BD group may include individuals that were diagnosed with a bipolar spectrum disorder at the two-year visit, but not the baseline visit. As another example, an MDD group may include individuals that were diagnosed with MDD at the two-year visit, but not the baseline visit. Similarly, a DMDD group may include individuals that were diagnosed with DMDD at the two-year visit, but not the baseline visit. In some examples, an exception may be made to include youth with a diagnosis of generalized, social, or separation anxiety disorders, as such anxiety disorders may have a common etiological origin and may thus be highly comorbid in BD or other mental health conditions.

[0031] In some examples, data used to train condition prediction model 116 and / or that is processed using condition prediction model 116 is normalized by machine learning engine 114 (e g., assessing the distribution and applying a data transformation as appropriate) and features are extracted from graph data store 112 (e.g., as a “data frame”) and structured into appropriate data types for subsequent processing.

[0032] As noted above, graph data store 112 stores features associated with a subject (e.g., which are clustered into nodes) and relationships between the nodes are defined within the edges. The nodes and edges that are extracted from graph data store 112 (e.g., as a data frame) may then be used to train condition prediction model 116. For example, condition prediction model 116 may be a multi-layer neural network that is used to determine a pattern of features and their organization(e g., including accounting for the neighborhood within the nodes and their edge relationships) that corresponds with no diagnosis and / or one or more mental health conditions.

[0033] Condition prediction model 116 is fit by minimizing an objective function. As an example, a cross entropy loss function is used. It will be appreciated that cross entropy loss is an objective function that is used in deep learning, which allows for binomial or multinomial classification. Condition prediction model 116 may be trained to minimize cross entropy loss, for example using an adaptive gradient descent algorithm. In examples, condition prediction model 116 is determined from a set of models that are compared using n-fold cross validation, which may improve out-of- sample prediction. Regularization of condition prediction model 116 may be improved through optimizing a set of explicit hyperparameters (e.g., the number of hidden layers in the neural network, which are automatically set through an automated tuning process) using the training dataset (e.g., as may be extracted from graph data store 112 as a data frame).

[0034] In some examples, there is risk of imbalance in outcomes (e.g., for diagnostic groups or on final fit). For example, some groups may have more constituents than others (e.g., those with no diagnosis as compared to a BD, MDD, or DMDD group). As such, fit may be biased toward accuracy in the no-diagnosis group as compared to the others. Parallel, balanced matched sample prediction may be used for the training dataset in other examples to assess balance effects. Thus, one or more GNN balance correction techniques may be used, such as ensemble learning, when determining an initial fit.

[0035] Similarly, overfit may occur in instances where condition prediction model 116 is fit to noise or idiosyncratic aspects of the training dataset, which thus impede generalization. To reduce overfit, a 10-fold cross validation procedure in the training dataset may be used in some examples. Model selection among competing folds may be determined by best Fl score in the training set, which may be selected as it balances precision and recall. A holdout sample of subjects may be used in some examples, which may be used to evaluate the generalizability of condition prediction model 116 after training. In instances where data is clustered by site or any of a variety of other clustering criteria, a sample corresponding to one such cluster may similarly be used for validation to verify no specific cluster overly influences predictions generated by condition prediction model 116.

[0036] Additionally, or alternatively, condition prediction model 116 comprises a regression model that is used to process a graph subpart of graph data store 112 to generate a risk score for a given diagnosis according to aspects described herein. In such an example, condition prediction model 116 is fit to a training data set using a least squares approach, minimizing a cost function, and / or according to any of a variety of other regression techniques. Thus, the regression model includes a plurality of parameters that each correspond to one or more features of the graph data store, such that the regression model is fit according to a training data set, such that the resulting regression model is usable to classify a graph subpart for a given subject (e.g., to generate a concurrent and / or future prediction for the subject).

[0037] Once machine learning engine 114 has trained condition prediction model 116 (e.g., based on a training dataset from graph data store 112), condition prediction model 116 may be used to generate concurrent and / or future predictions for a given subject. As an example, a future prediction that is generated according to aspects described herein may comprise an indication as to whether a given subject is likely to exhibit or otherwise be diagnosed with a given mental health condition after a predetermined amount of time has elapsed (e.g., one or two years). In examples, neuroimaging outcomes and / or clinical data for the subject may be stored by graph data store 112, such that a graph subpart is identified and processed using condition prediction model 116. As another example, neuroimaging outcomes and / or clinical data may be received (e.g., from computing device 104), which may be ingested by graph manager 110 to add, remove, or update one or more nodes and / or edges within graph data store 112 accordingly.

[0038] Condition prediction model 116 may generate a concurrent prediction and / or a future prediction based on a graph subpart representation of neuroimaging outcomes and / or clinical data for a given subject. A prediction may comprise a binary indication (e.g., no diagnosis versus a diagnosis of BD / MDD / DMDD), may comprise a probability (e.g., a subject has a determined probability of having a given mental health condition), may comprise a confidence score, or any combination thereof, among other examples. As an example, a prediction generated by condition prediction model 116 may comprise a set of probabilities that each correspond to the probability of no diagnosis, the probability of BD, the probability of MDD, and / or the probability of DMDD.

[0039] In examples, multiple models may be used, where a first model is used to generate a concurrent prediction, while a second model is used to generate a future prediction. As notedabove, condition prediction model 1 16 may generate an explainability indication associated with a generated prediction. For example, the explainability indication may comprise one or more associated confidence scores and / or weights corresponding to features that had an effect on the generated prediction. The generated prediction(s) (and, in some examples, an associated explainability indication) may be provided to computing device 104 for display to a user by application 118.

[0040] It will be appreciated that while system 100 is illustrated as comprising one condition prediction platform 102, one computing device 104, and one imaging device 106, any number of such elements may be used in other examples. For example, a subject may use a first computing device (e.g., to provide clinical data), while a clinician may use a second computing device.

[0041] Further, while application 118 has been described as both providing data to be ingested by condition prediction platform 102 and displaying a prediction that was generated by condition prediction platform 102, it will be appreciated that any number of applications may be used, each of which may include at least a subset of the functionality described herein. Additionally, the functionality described herein may be distributed among or otherwise implemented on any number of different devices in any of a variety of other configurations in other examples.

[0042] As an example, computing device 104 may implement aspects associated with graph manager 110, graph data store 112, and / or machine learning engine 114, such that at least a part of the data associated with a subject may be stored and / or processed local to computing device 104 in addition to or as an alternative to the above-described processing by condition prediction platform 102, which may thus maintain or improve user privacy in some examples.

[0043] Figure 2 illustrates an example graph subpart 200 that may be used to train a condition prediction model (e.g., condition prediction model 116 in Figure 1) and / or to generate a prediction according to aspects described herein. In examples, graph subpart 200 is stored in a graph data store (e.g., graph data store 112) and may have been generated by a graph manager (e.g., graph manager 110) based on neuroimaging outcomes and clinical data associated with a subject. For example, the graph subpart may be extracted from the graph data store and used to generate a data frame, which may be processed according to aspects described herein.

[0044] As illustrated, graph subpart 200 includes subject node 202, family history node 204, mania scale node 206, image data node 208, and trauma / anxiety node 210. It will be appreciated that theillustrated nodes are provided as an example and, in other examples, any of a variety of additional or alternative nodes may be used. For instance, the illustrated example includes first-level nodes associated with subject node 202. In other examples, additional levels of nodes (e.g., nodes having one or more intermediate nodes in relation to subject node 202) may be processed according to aspects herein. Each node 202, 204, 206, 208, and 210 is further illustrated as comprising a set of properties stored therein.

[0045] Additionally, graph subpart 200 includes edges 214, 216, 218, and 220 that relate nodes 204, 206, 208, and 210, respectively, with subject node 202. Similar to nodes 202, 204, 206, 208, and 210, edges 214, 216, 218, and 220 include a set of associated properties. As illustrated, edges 214, 216, 218, and 220 include, among other things, an “Interview_Date” property, thereby enabling longitudinal feature extraction in some examples.

[0046] Thus, features may be extracted from graph subpart 200 according to aspects described herein (e.g., relating to a subject to which subject node 202 corresponds) and used to train a machine learning model, fit a regression model, and / or to generate a prediction accordingly, according to aspects described herein.

[0047] Figure 3 illustrates an overview of an example method 300 for generating a set of nodes and edges for a subject within a graph data store according to aspects described herein. In examples, aspects of method 300 are performed by a condition prediction platform, such as condition prediction platform 102 in Figure 1. For example, graph manager 110 may perform aspects of method 300.

[0048] Method 300 begins at operation 302, where one or more neuroimaging outcomes are obtained for a subject. For example, the neuroimaging outcomes may include DTI and / or EPI outcomes, as noted above. In examples, operation 302 comprises acquiring imaging data from an imaging device (e.g., imaging device 106 in Figure 1). As another example, operation 302 comprises obtaining pre-existing imaging data from a data store, for example of a remote computing device or from a third party (e.g., as may be the case when a pre-existing dataset is used). It will therefore be appreciated that neuroimaging outcomes may be obtained from any of a variety of sources.

[0049] At operation 304, clinical data is obtained for the subject. For example, the clinical data may be obtained from an electronic health record (EHR) system. As another example, clinical datamay be obtained from a computing device (e.g., computing device 104 in Figure 1), as may be the case when the user completes a survey, questionnaire, or test, among other examples. In some instances, the clinical data obtained at operation 304 may correspond to a different individual than the subject, as may relate to the subject’s family history. Thus, the clinical data obtained at operation 304 may be pre-existing and / or may be newly generated clinical data, among other examples. It will therefore be appreciated that clinical data may be obtained from any of a variety of sources.

[0050] Flow progresses to operation 306, where a node is generated for the subject in a graph data store (e.g., graph data store 112 in Figure 1). Aspects of the generated subject node may be similar to subject node 202 discussed above with respect to graph subpart 200 in Figure 2. In other examples, a node for the subject may be pre-existing, as may be the case when the graph data store is updated (e.g., to include new data, to update existing data, and / or to remove data). In such examples, operation 306 may instead comprise identifying an existing node within the graph data store that corresponds to the subject.

[0051] At operation 308, nodes and / or edges associated with the subject node (e.g., subject node 202) are generated based on the data that was obtained at operation 302 and 304. For example, properties may be extracted from the data that was obtained at operation 302 and 304, which may be stored as properties in the generated nodes and / or edges according to aspects described herein. With reference to graph subpart 200 in Figure 2, an image data node similar to image data node 208 may be generated based on the neuroimaging outcomes that was obtained at operation 302, while a family history node, a mania scale node, and / or a trauma / anxiety node similar to nodes 204, 206, and 210, respectively, may be generated based on the clinical data that was obtained at operation 304. It will be appreciated that example nodes, edges, and properties are described but that, in other examples, any of a variety of additional or alternative nodes, edges, and / or properties may be used.

[0052] Method 300 continues to operation 310, where subsequent processing is performed for the subject based on the updated graph data store. Example subsequent processing includes, but is not limited to, training or updating a condition prediction model (e.g., condition prediction model 116 in Figure 1) and / or processing one or more nodes and / or edges that were generated or updated atoperations 306 and 308 to generate a prediction according to aspects described herein. Method 300 terminates at operation 310.

[0053] Figure 4A illustrates an overview of example method 400 for training a machine learning model to predict a mental health condition according to aspects described herein. In examples, aspects of method 400 are performed by a condition prediction platform, such as condition prediction platform 102 in Figure 1. For example, machine learning engine 114 may perform aspects of method 400 to generate a condition prediction model, such as condition prediction model 116.

[0054] Method 400 begins at operation 402, where a set of nodes and / or edges associated with a subject are identified from a graph data store (e.g., graph data store 112 in Figure 1). In examples, the nodes and / or edges are identified based on an association with a subject node (e.g., subject node 202 in Figure 2). Operation 402 may comprise identifying first-level nodes associated with a subject node (e.g., those that are directly connected with the subject node by an edge) or may comprise identifying first- and second-level nodes (e.g., those that are directly connected and those that have one intermediate node), among other examples. In other examples, nodes and / or edges may be filtered according to associated properties. For example, only nodes / edges having properties that are within a predetermined date range may be identified at operation 402. It will therefore be appreciated that nodes and / or edges may be identified according to any of a variety of criteria. The identified nodes / edges may also be referred to herein as a graph subpart (e.g., graph subpart 200 in Figure 2).

[0055] Flow progresses to operation 404, where a data frame is generated based on the nodes / edges that were identified at operation 402. In examples, the data frame includes features that are extracted from the graph subpart that was identified at operation 402. With reference to graph subpart 200 in Figure 2, an example data frame may include features relating to family history, neuroimaging outcomes, trauma / anxiety experience, and / or a mania scale, among other examples.

[0056] In some instances, additional processing is performed based on data that was identified from the graph data store, as may be the case when one or more additional properties of nodes and / or edges are programmatically generated. For example, one or more properties may be processed using a machine learning model to generate model output that is used as a feature or, asanother example, linear regression or other processing may be used to programmatically enrich the extracted properties accordingly, such that one or more features may be generated therefrom.

[0057] Flow progresses to operation 406, where the data frame is preprocessed. For example, feature normalization and / or balancing techniques may be used to account for data issues that may potentially affect the generalizability of the resulting model. While example preprocessing techniques are described, it will be appreciated that any of a variety of additional or alternative techniques may be used in other examples.

[0058] Operation 406 is illustrated as using a dashed line to indicate that, in some examples, operation 406 may be omitted. Arrow 412 is illustrated from operation 406 to operation 402 to indicate that, in some examples, method 400 loops between operations 402, 404, and 406, as may be the case when multiple data frames are generated (e.g., for multiple subjects). For example, method 400 loops between operations 402, 404, and 406 to generate data frames corresponding to subjects that were not diagnosed with a mental health condition and subjects that were diagnosed with BD, MDD, and / or DMDD, thereby generating a training dataset with which to train a machine learning model (e.g., condition prediction model 116 in Figure 1). In some examples, preprocessing at operation 406 is performed after multiple data frames have been generated, such that flow loops between operations 402 and 404, after which it arrives at operation 406.

[0059] Eventually, flow arrives at operation 408, where a machine learning model is trained according to the data frame(s) that were generated as a result of operations 402, 404, and 406. As noted above, the machine learning model may be iteratively trained according to the generated training dataset with the objective to minimize a cross entropy loss function exhibited by the machine learning model. While example model structures and associated training techniques are described, it will be appreciated that any of a variety of other structures and / or training techniques may be used in other examples.

[0060] At operation 410, the machine learning model is stored for subsequent condition prediction. For example, the machine learning model may be stored at a condition prediction platform (e.g., condition prediction platform 102 in Figure 1). As another example, the machine learning model may be provided to a computing device (e.g., computing device 104), such that the machine learning model may be used to evaluate clinical data and / or neuroimaging outcomes local to the computing device. Method 400 terminates at operation 410.

[0061] In examples, aspects similar to method 400 are used to fit a regression model according to aspects described herein. For instance, similar to the above-discussed aspects of operation 408 with respect to a machine learning model, parameters of a regression model may similarly be tuned to fit a set of data frames that were generated as a result of operations 402, 404, and 406, for example according to a least squares approach and / or by minimizing a cost function, among other examples. Thus, each parameter of the regression model corresponds to one or more features of a given data frame, such that the regression model is fitted to the set of data frames by tuning each corresponding parameter accordingly. The resulting regression model may thus process features for a data frame of a given subject to generate a corresponding risk probability for a given diagnosis according to aspects described herein.

[0062] Figure 4B illustrates an overview of an example method 450 for generating a prediction for a subject according to aspects described herein. In examples, aspects of method 450 are performed by a condition prediction platform, such as condition prediction platform 102 in Figure 1. For example, machine learning engine 114 may perform aspects of method 400 to generate a condition prediction using a condition prediction model (e.g., condition prediction model 116), as may have been trained as a result of performing aspects of method 400 in Figure 4 A.

[0063] Method 450 begins at operation 452, where a data frame is generated based on data for a given subject. In examples, operation 452 comprises identifying one or more nodes and / or edges from a graph data store, such as graph data store 112 in Figure 1. Aspects of operation 452 may be similar to those discussed above with respect to operations 402 and 404 of method 400 in Figure 4A and are therefore not necessarily redescribed in detail. In other examples, operation 452 comprises obtaining clinical data and / or neuroimaging outcomes for a subject, which may be processed according to aspects described above with respect to method 300 of Figure 3 to incorporate the data into a graph data store or to otherwise generate a graph subpart for the subject.

[0064] At operation 454, the data frame that was generated at operation 452 may be preprocessed. Aspects of operation 454 may be similar to those discussed above with respect to operation 406 of method 400 in Figure 4A and are therefore not necessarily redescribed in detail. Similar to operation 406, operation 454 is illustrated using a dashed box to indicate that, in some examples, operation 454 may be omitted.

[0065] Method 450 progresses to operation 456, where the data frame is processed using a machine learning model to generate a prediction for the given subject. In examples, the machine learning model is a condition prediction model, such as condition prediction model 116. For example, the machine learning model may have been trained (e.g., by a machine learning engine, such as machine learning engine 114) as a result of performing aspects of method 400 in Figure 4A.

[0066] As noted above, a concurrent and / or a future prediction may be generated, which may include an indication of no diagnosis, a diagnosis of BD, a diagnosis of MDD, and / or a diagnosis of DMDD, among other examples. In examples, a confidence score is generated in association with such a prediction. In some instances, an explainability indication is generated for a prediction, which may comprise one or more associated confidence scores and / or weights corresponding to features that had an effect on the generated prediction (e.g., as may have been obtained from a graph data store).

[0067] Flow progresses to operation 458, where an indication of the generated prediction is provided for display to a user (e.g., a clinician). For example, the indication may be communicated to a computing device (e.g., computing device 104 in Figure 1), at which point it is displayed via an application executing thereon (e.g., application 118). As another example, operation 458 comprises updating a user interface to display the generated prediction accordingly. Method 450 terminates at operation 458.

[0068] Similar to method 400, aspects of method 450 may similarly be performed using a regression model in addition to or as alternative to data frame processing according to a trained machine learning model (e.g., as discussed above with respect to operation 456). For example, a regression model (e.g., as may have been fitted to a training dataset according to the abovedescribed aspects) is used to process a data frame (e.g., as was generated by operation 452 and, in some examples, operation 454). The resulting risk probability may thus be used to generate a prediction indication consistent with aspects of operation 458 discussed above.

[0069] Figure 5 illustrates an example of a suitable operating environment 500 in which one or more of the present embodiments may be implemented. This is only one example of a suitable operating environment and is not intended to suggest any limitation as to the scope of use or functionality. Other well-known computing systems, environments, and / or configurations that may be suitable for use include, but are not limited to, personal computers, server computers, hand-held or laptop devices, multiprocessor systems, microprocessor-based systems, programmable consumer electronics such as smart phones, network PCs, minicomputers, mainframe computers, distributed computing environments that include any of the above systems or devices, and the like.

[0070] In its most basic configuration, operating environment 500 typically may include at least one processing unit 502 and memory 504. Depending on the exact configuration and type of computing device, memory 504 (storing, among other things, APIs, programs, etc. and / or other components or instructions to implement or perform the system and methods disclosed herein, etc.) may be volatile (such as RAM), non-volatile (such as ROM, flash memory, etc.), or some combination of the two. This most basic configuration is illustrated in Figure 5 by dashed line 506. Further, environment 500 may also include storage devices (removable, 508, and / or nonremovable, 510) including, but not limited to, magnetic or optical disks or tape. Similarly, environment 500 may also have input device(s) 514 such as a keyboard, mouse, pen, voice input, etc. and / or output device(s) 516 such as a display, speakers, printer, etc. Also included in the environment may be one or more communication connections, 512, such as LAN, WAN, point to point, etc.

[0071] Operating environment 500 may include at least some form of computer readable media. The computer readable media may be any available media that can be accessed by processing unit 502 or other devices comprising the operating environment. For example, the computer readable media may include computer storage media and communication media. The computer storage media may include volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. The computer storage media may include RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transitory medium, which can be used to store the desired information. The computer storage media may not include communication media.

[0072] The communication media may embody computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism and includes any information delivery media. The term “modulated data signal” may mean a signal that has one or more of its characteristics set or changed in such a manner as toencode information in the signal. For example, the communication media may include a wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared and other wireless media. Combinations of the any of the above should also be included within the scope of computer readable media.

[0073] The operating environment 500 may be a single computer operating in a networked environment using logical connections to one or more remote computers. The remote computer may be a personal computer, a server, a router, a network PC, a peer device or other common network node, and typically includes many or all of the elements described above as well as others not so mentioned. The logical connections may include any method supported by available communications media. Such networking environments are commonplace in offices, enterprisewide computer networks, intranets and the Internet.

[0074] The different aspects described herein may be employed using software, hardware, or a combination of software and hardware to implement and perform the systems and methods disclosed herein. Although specific devices have been recited throughout the disclosure as performing specific functions, one skilled in the art will appreciate that these devices are provided for illustrative purposes, and other devices may be employed to perform the functionality disclosed herein without departing from the scope of the disclosure.

[0075] As stated above, a number of program modules and data files may be stored in the system memory 504. While executing on the processing unit 502, program modules (e.g., applications, Input / Output (I / O) management, and other utilities) may perform processes including, but not limited to, one or more of the stages of the operational methods described herein such as the methods illustrated in Figures 3, 4A, or 4B, for example.

[0076] Furthermore, examples of the invention may be practiced in an electrical circuit comprising discrete electronic elements, packaged or integrated electronic chips containing logic gates, a circuit utilizing a microprocessor, or on a single chip containing electronic elements or microprocessors. For example, examples of the invention may be practiced via a system-on-a-chip (SOC) where each or many of the components illustrated in Figure 5 may be integrated onto a single integrated circuit. Such an SOC device may include one or more processing units, graphics units, communications units, system virtualization units and various application functionality all of which are integrated (or "burned") onto the chip substrate as a single integrated circuit. Whenoperating via an SOC, the functionality described herein may be operated via application-specific logic integrated with other components of the operating environment 500 on the single integrated circuit (chip). Examples of the present disclosure may also be practiced using other technologies capable of performing logical operations such as, for example, AND, OR, and NOT, including but not limited to mechanical, optical, fluidic, and quantum technologies. In addition, examples of the invention may be practiced within a general purpose computer or in any other circuits or systems.

[0077] As will be understood from the foregoing disclosure, one aspect of the technology relates to a system comprising: at least one processor; and memory storing instructions that, when executed by the at least one processor, cause the system to perform a set of operations. The set of operations comprises: obtaining at least one of a neuroimaging outcome or clinical data for a subject; generating, for the subject and based on the obtained data, a set of nodes and a set of edges that are associated with a subject node within a graph data store, thereby forming a graph subpart for the subject; processing the graph subpart using a condition prediction model to generate a prediction for the subject, wherein the prediction is for one of an incidence of: bipolar disorder; major depressive disorder; or disruptive mood dysregulation disorder; and providing an indication of the generated prediction for display to a user. In an example, processing the graph subpart further comprises generating an explainability indicator associated with the generated prediction; and the explainability indicator is provided in conjunction with the generated prediction for display to the user. In another example, the explainability indicator includes a set of features that were used to generate the prediction. In a further example, the prediction is one of a concurrent prediction or a future prediction. In yet another example, the graph subpart for the subject includes nodes and edges that are directly related to the subject node. In a further still example, generating the set of nodes and the set of edges comprises updating the graph data store to include the set of nodes and the set of edges in association with an existing subject node. In another example, the condition prediction model is a graph neural network that was trained using training data corresponding to: a first graph subpart for a first subject that was not diagnosed with a mental health condition; and a second graph subpart for a second subject that was diagnosed with the mental health condition. In a further example, the first subject and the second subject each did not exhibit the mental health condition at a first point in time; and the second subject exhibited the mental health condition at a second point in time that is later than the first point in time. In yet another example, the prediction is a future prediction for an incidence of the mental healthcondition in a time frame corresponding to the first point in time and the second point in time. In a further still example, the condition prediction model is a regression model fit to a plurality of graph subparts that each correspond to a subject having a corresponding diagnosis. In another example, the indication of the generated prediction is provided to a computing device from which at least a part of the data for the subject was obtained.

[0078] In another aspect, the technology relates to a method for processing a graph data store to generate a prediction for a subject. The method comprises: processing a graph subpart for the subject using a condition prediction model to generate a prediction for the subject, wherein the graph subpart comprises a set of nodes and a set of edges within the graph data store that are associated with at least one of a neuroimaging outcome or clinical data for the subject, and wherein the prediction is for one of an incidence of: bipolar disorder; major depressive disorder; or disruptive mood dysregulation disorder; and providing an indication of the generated prediction for display to a user. In an example, the condition prediction model is a graph neural network that was trained using training data corresponding to: a first graph subpart of the graph data store for a first subject that was not diagnosed with a mental health condition; and a second graph subpart of the graph data store for a second subject that was diagnosed with the mental health condition. In another example, the condition prediction model is a regression model fit to a plurality of graph subparts that each correspond to a subject having a corresponding diagnosis. In a further example, the prediction is one of a concurrent prediction or a future prediction.

[0079] In a further aspect, the technology relates to a method for processing a graph data store to generate a prediction for a subject. The method comprises: obtaining at least one of a neuroimaging outcome or clinical data for a subject; generating, for the subject and based on the obtained data, a set of nodes and a set of edges that are associated with a subject node within a graph data store, thereby forming a graph subpart for the subject; processing the graph subpart using a condition prediction model to generate a prediction for the subject, wherein the prediction is for one of an incidence of: bipolar disorder; major depressive disorder; or disruptive mood dysregulation disorder; and providing an indication of the generated prediction for display to a user. In an example, processing the graph subpart further comprises generating an explainability indicator associated with the generated prediction, wherein the explainability indicator includes a set of features that were used to generate the prediction; and the explainability indicator is provided in conjunction with the generated prediction for display to the user. In another example, the conditionprediction model is a graph neural network that was trained using training data corresponding to: a first graph subpart for a first subject that was not diagnosed with a mental health condition; and a second graph subpart for a second subject that was diagnosed with the mental health condition. In a further example, the first subject and the second subject each did not exhibit the mental health condition at a first point in time; and the second subject exhibited the mental health condition at a second point in time that is later than the first point in time. In yet another example, the condition prediction model is a regression model fit to a plurality of graph subparts that each correspond to a subject having a corresponding diagnosis.

[0080] Aspects of the present disclosure, for example, are described above with reference to block diagrams and / or operational illustrations of methods, systems, and computer program products according to aspects of the disclosure. The functions / acts noted in the blocks may occur out of the order as shown in any flowchart. For example, two blocks shown in succession may in fact be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality / acts involved.

[0081] The description and illustration of one or more aspects provided in this application are not intended to limit or restrict the scope of the disclosure as claimed in any way. The aspects, examples, and details provided in this application are considered sufficient to convey possession and enable others to make and use the best mode of claimed disclosure. The claimed disclosure should not be construed as being limited to any aspect, example, or detail provided in this application. Regardless of whether shown and described in combination or separately, the various features (both structural and methodological) are intended to be selectively included or omitted to produce an embodiment with a particular set of features. Having been provided with the description and illustration of the present application, one skilled in the art may envision variations, modifications, and alternate aspects falling within the spirit of the broader aspects of the general inventive concept embodied in this application that do not depart from the broader scope of the claimed disclosure.

Claims

CLAIMSWhat is claimed is:

1. A system comprising: at least one processor; and memory storing instructions that, when executed by the at least one processor, cause the system to perform a set of operations, the set of operations comprising: obtaining at least one of a neuroimaging outcome or clinical data for a subject; generating, for the subject and based on the obtained data, a set of nodes and a set of edges that are associated with a subject node within a graph data store, thereby forming a graph subpart for the subject; processing the graph subpart using a condition prediction model to generate a prediction for the subject, wherein the prediction is for one of an incidence of: bipolar disorder; major depressive disorder; or disruptive mood dysregulation disorder; and providing an indication of the generated prediction for display to a user.

2. The system of claim 1, wherein: processing the graph subpart further comprises generating an explainability indicator associated with the generated prediction; and the explainability indicator is provided in conjunction with the generated prediction for display to the user.

3. The system of claim 2, wherein the explainability indicator includes a set of features that were used to generate the prediction.

4. The system of claim 1, wherein the prediction is one of a concurrent prediction or a future prediction.

5. The system of claim 1, wherein the graph subpart for the subject includes nodes and edges that are directly related to the subject node.

6. The system of claim 1, wherein generating the set of nodes and the set of edges comprises updating the graph data store to include the set of nodes and the set of edges in association with an existing subject node.

7. The system of claim 1, wherein the condition prediction model is a graph neural network that was trained using training data corresponding to: a first graph subpart for a first subject that was not diagnosed with a mental health condition; and a second graph subpart for a second subject that was diagnosed with the mental health condition.

8. The system of claim 7, wherein: the first subject and the second subject each did not exhibit the mental health condition at a first point in time; and the second subject exhibited the mental health condition at a second point in time that is later than the first point in time.

9. The system of claim 8, wherein the prediction is a future prediction for an incidence of the mental health condition in a time frame corresponding to the first point in time and the second point in time.

10. The system of claim 1, wherein the condition prediction model is a regression model fit to a plurality of graph subparts that each correspond to a subject having a corresponding diagnosis.

11. The system of claim 1, wherein the indication of the generated prediction is provided to a computing device from which at least a part of the data for the subject was obtained.

12. A method for processing a graph data store to generate a prediction for a subject, the method comprising: processing a graph subpart for the subject using a condition prediction model to generate a prediction for the subject, wherein the graph subpart comprises a set of nodes and a set of edges within the graph data store that are associated with at least one of a neuroimaging outcome or clinical data for the subject, and wherein the prediction is for one of an incidence of: bipolar disorder; major depressive disorder; or disruptive mood dysregulation disorder; and providing an indication of the generated prediction for display to a user.

13. The method of claim 12, wherein the condition prediction model is a graph neural network that was trained using training data corresponding to: a first graph subpart of the graph data store for a first subject that was not diagnosed with a mental health condition; and a second graph subpart of the graph data store for a second subject that was diagnosed with the mental health condition.

14. The method of claim 12, wherein the condition prediction model is a regression model fit to a plurality of graph subparts that each correspond to a subject having a corresponding diagnosis.

15. The method of claim 12, wherein the prediction is one of a concurrent prediction or a future prediction.

16. A method for processing a graph data store to generate a prediction for a subject, the method comprising: obtaining at least one of a neuroimaging outcome or clinical data for a subject; generating, for the subject and based on the obtained data, a set of nodes and a set of edges that are associated with a subject node within a graph data store, thereby forming a graph subpart for the subject;processing the graph subpart using a condition prediction model to generate a prediction for the subject, wherein the prediction is for one of an incidence of: bipolar disorder; major depressive disorder; or disruptive mood dysregulation disorder; and providing an indication of the generated prediction for display to a user.

17. The method of claim 16, wherein: processing the graph subpart further comprises generating an explainability indicator associated with the generated prediction, wherein the explainability indicator includes a set of features that were used to generate the prediction; and the explainability indicator is provided in conjunction with the generated prediction for display to the user.

18. The method of claim 16, wherein the condition prediction model is a graph neural network that was trained using training data corresponding to: a first graph subpart for a first subject that was not diagnosed with a mental health condition; and a second graph subpart for a second subject that was diagnosed with the mental health condition.

19. The method of claim 18, wherein: the first subject and the second subject each did not exhibit the mental health condition at a first point in time; and the second subject exhibited the mental health condition at a second point in time that is later than the first point in time.

20. The method of claim 16, wherein the condition prediction model is a regression model fit to a plurality of graph subparts that each correspond to a subject having a corresponding diagnosis.

Images