Machine Learning Methods for Optimizing Individualized Treatment Strategies for Precision Psychiatry

Project Summary: Mental disorders cause immense disability, accounting for 183.9 million disability-adjusted life-years world- wide. Among currently available pharmacological and behavioral interventions, no single therapy is universally ef- fective. Moreover, treatment responses are far from adequate across mental disorders. As such, there is an urgent need to optimize treatment responses. Various factors appear to be associated with positive treatment responses for mental disorders, thus providing evidence for improving response rate by incorporating patient-specific charac- teristics in treatment decisions in an effort to achieve precision psychiatry. However, existing methods to incorpo- rate patient-specific characteristics do not adequately address the unique challenges facing precision psychiatry. To point, treatment decision making for mental disorders is inevitably confronted by extensive diagnostic hetero- geneity, substantial between-patient variation in biological and clinical manifestations of disease, and mismatch between diagnostic categorization and the underlying pathophysiology. To address these emerging challenges, this proposal aims to develop novel machine learning and statistical inference methods to build individualized treat- ment rules to account for the extensive heterogeneity and between-patient variability and integrate evidence from multi-domain brain and behavioral data across several disorders. Specifically, we aim to: (1) learn optimal latent representation of patients through a probabilistic generative model that has theoretical support under the National Institute of Mental Health Strategic Plan on Research Domain Criteria (RDoC); (2) incorporate prior optimal treat- ment information from the non-randomized phase of clinical trials through targeted transfer learning; (3) synthesize individualized treatment decision rules learned from multiple studies; and (4) provide rigorous statistical inference of fitted decision rules. Following the RDoC call for centering mental health research around latent constructs shared across disorders, the methods developed here will be applied to a range of randomized controlled trials (RCTs) of patients with major depressive disorder and other co-morbid disorders, including multiple high-quality RCTs with multi-modality data (e.g., symptoms, behavioral tests, psychosocial measures, brain measures). This strategy will allow for examination of treatment strategies for constructs shared across disorders and thus will in- crease generalizability. In sum, this research will use machine learning approaches and statistical inference in an effort to better leverage the complex interplay between biomarkers and clinical manifestations in the context of precision psychiatry, with the goal of selecting the best treatments for patients with mental disorders.