3/5 Clinical Outcome Prediction of Psychosis from Electronic Health Records (COPPER) - PROJECT SUMMARY Clinical predictors are now firmly incorporated into routine standard-of-care in many fields of medicine, in contrast with Psychiatry where quantitative predictors that guide clinical decision-making remain extremely limited. Psychosis-related disorders are responsible for a substantial public health burden, for which there are significant unmet needs that would be subserved by clinical predictors. For example, long-term outcomes vary widely and identifying individuals with poor or advantageous future outcomes would help to optimize treatment planning and resource allocation. Furthermore, antipsychotics are associated with adverse side effects, such as increased risk of diabetes. In this application, we propose to use machine learning approaches to build predictors and identify subtypes of clinical outcomes among individuals with schizophrenia, through integration of longitudinal electronic health records (EHRs), dimensional phenotyping, and genetic analyses. We will also explore the psychosocial and ethical implications of psychiatric clinical predictors. Our long-term objective is to advance the goals of Precision Psychiatry to achieve individualized treatment planning, outcome monitoring, and preventive interventions. We propose the following specific aims: Aim 1: Leverage two independent EHR databases for outcome prediction and sub-classification of psychosis-related disorders. (a) We will use the longitudinal PSYCKES and MarketScan databases to build machine learning-based individual-level prediction models to forecast the onset of four major prognostic outcomes: treatment response (antipsychotic resistance), illness severity (long-term hospitalization), medical comorbidity (diabetes), and diagnostic transition from a psychosis-related disorder to schizophrenia. (b) We will perform cohort-level analyses using unsupervised methods to discover novel psychosis-related diagnosis and prognosis subtypes. Aim 2: Enhance predictive modeling through dimensional phenotyping and whole genome sequencing. (a) We will recruit n = 10,000 patients with schizophrenia from the PSYCKES database population for enriched data collection: 1) dimensional phenotyping (cognition, exposome, and social determinants of health), and 2) whole genome sequencing to enable calling of rare variants, structural variants, and common variants (polygenic risk). (b) We will investigate the extent to which dimensional phenotypes and genomic data can improve the models developed in Aim 1. Aim 3: Explore the psychosocial and ethical implications of psychiatric clinical predictors. (a) We will survey a subset of patients and their clinicians regarding their attitudes towards implementation of clinical outcome predictors. (b) We will return pathogenic findings to patients through genetic counseling and survey the experience of patients and their clinicians on their emotional reactions and perceptions of impairment, treatability, and life-planning.
