Improving emergency psychiatric care through machine learning, causal inference, and psychometrics - Background: The United States is experiencing a crisis in mental health care, in which long-term shifts towards the emergency department (ED) for psychiatric treatment have been exacerbated by the COVID-19 pandemic – reduced inpatient hospitalization capacity, delayed outpatient appointments, extended ED boarding, and staffing shortages, combined with substantial growth in demand for mental health services (due in part to pandemic-related increases in stress, depression, social isolation, and anxiety, and with differential impacts on vulnerable populations). Data-driven approaches have not been widely applied to understand mechanisms and predictors of ED utilization for psychiatric patients, or to optimize clinical decision-making to improve patient outcomes. Research: In this study, I will conduct secondary data analysis of 141,431 ED visits for psychiatric care in a large regional health system from 2017-2023, extracting and preparing over 5,000 patient characteristics from the electronic health record (diagnoses, labs, medications, procedures, visits, and notes). Aim 1: I will create a transdiagnostic subtyping system, using latent class analysis, hierarchical density- based clustering, and neural network autoencoders, to reliably categorize and track major trends in patient populations receiving emergency psychiatric care. Aim 2: I will develop a suite of ensemble machine learning models to predict ED length of stay, ED boarding duration, short-term ED re-admission, transfer to inpatient hospitalization, and post-discharge adverse events (suicide attempt, overdose, or psychotic episode). Aim 3: I will estimate a precision treatment rule, using targeted causal inference and ensemble machine learning, to capture heterogeneous treatment effects for primary ED disposition decisions (inpatient hospitalization, partial hospitalization / intensive outpatient program, or outpatient monitoring) on adverse psychiatric events. Candidate's Career Development, Goals, and Environment: This proposal's research aims and the candidate's career development will be supported by the extensive resources available at Massachusetts General Hospital (MGH) and Harvard Medical School, as well as formal training, coursework, and mentorship in (T1) mental disorders and psychopathology, (T2) precision treatment optimization, (T3) hybrid pragmatic effectiveness trials, and (T4) professional development in preparation for a future R01 submission. The mentorship team includes Primary Mentor Dr. Jordan Smoller, a leading expert in precision psychiatry; Co- Mentors Dr. Matthew Nock, a leader in mental disorders and suicide prevention; Dr. Susan Murphy, a leader in precision treatment optimization; and Dr. Stephen Bartels, a leader in implementation science and hybrid pragmatic trials; and Consultants Dr. Suzanne Bird, expert in emergency psychiatry and director of MGH Acute Psychiatry Services; Dr. Edwin Boudreaux, expert in hybrid pragmatic trials in emergency psychiatry; and Dr. Soroush Saghafian, expert in reinforcement learning to optimize hospital decision-making. This award will enable the candidate to develop an independent, rigorous research program in computational mental health.
