Uncertainty-Aware Prediction of Differential Responses to Antidepressants: Leveraging EHR and Genomics - Background: Depression is a serious mental disorder, with treatment selection largely relying on trial and error, often prolonging patients' suffering. The increased availability of electronic health records (EHRs) and advancements in AI offer new opportunities to address this clinical challenge. However, current EHR-based approaches have shortcomings: a. they underutilize information in unstructured data that could be important for outcome prediction and confounding adjustments; b. they lack accuracy in cohort definition and treatment response assessments; c. they omit genomic information, which is known to affect treatment response; and d. they are not aware of uncertainties arising from the fitness of assumptions required to produce reliable predictions, potentially providing misleading estimates. In addition, genetic tests currently available are limited to select genetic variations, failing to utilize information from the full genome. Research: We propose to address these limitations by crafting advanced AI models for predicting differential antidepressant treatment responses, leveraging the latest developments in natural language processing (NLP), predictive modeling, causal inference, and the inclusion of both EHR and genomic data. Aim 1 will involve developing a large language model-based, human-in-the-loop active learning framework to identify an incident-user cohort started on antidepressants for depression, assess treatment responses, and extract key depression-related information from clinical notes. Aim 2 will develop uncertainty-aware, EHR-based prediction models for differential antidepressant responses, accounting for cases where a patient-antidepressant pairing falls outside the training data and for residual confounding. Aim 3 will combine EHR and three classes of genomic predictors for response prediction: genome-wide and pathway-specific polygenic risk scores, and variations associated with cytochrome P450 enzymes. This effort will enhance our understanding of integrating EHR and genomic data to predict personalized treatment responses, paving the way for future comprehensive systems. Candidate's Career Development, Goals, and Environment: The research objectives and the candidate's career development will be facilitated by the abundant resources at Massachusetts General Hospital and Harvard Medical School, as well as formal training and mentorship in (G1) advanced clinical NLP, (G2) integration and analysis of large-scale EHR and genomic data, (G3) ‘causal machine learning’ and its uncertainty assessments, and (G4) grantsmanship, leadership, effective collaborations, and research management. The mentorship team comprises Mentor Dr. Jordan Smoller, a leader in precision psychiatry and clinical predictive analytics; Co-Mentor Dr. Tianxi Cai, an authority in bioinformatics and healthcare predictive modeling; and Consultants Dr. Timothy Miller, an expert in NLP and AI, Dr. Issa Dahabreh, a specialist in causal inference, and Dr. Tian Ge, a renowned statistician and geneticist. This award will equip the candidate with the advanced skillset to become an independent researcher in precision psychiatry.
