Vital Signs for Psychosis: Personalized Longitudinal Markers of Psychosis Severity Using Automated Speech and Language Analysis

Project Summary There is a scarcity of objective, low-cost markers for illness severity for psychosis, which hinders the implementation of measurement-based care. The goal is to develop a low-cost, low-burden approach that outputs accurate, objective markers of psychosis severity to be used in a repeated, longitudinal clinical context – serving as “vital signs” for concurrent psychosis severity. Automated speech and language analysis show promise for this application, but high accuracy and reliability is required to produce clinically actionable measures. (Aim 1) This project will implement and evaluate novel approaches for adapting predictive models to repeated measures for each individual, producing personalized models that predict concurrent psychosis severity based on speech and language features. Specifically, we will first build a static, generalizable model after harmonizing a large existing dataset of ~3,500 speech samples from individuals on the psychosis spectrum (PS), including those at clinical high risk (CHR) for psychosis drawn from the publicly available AMP Schizophrenia dataset, people with early episode psychotic disorders, and individuals with chronic illness – either schizophrenia spectrum or mood disorders with psychotic features. Then, in a prospective sample of 100 PS participants followed over 15-50 timepoints in 1 year, we will build and compare four novel deep learning and algorithmic strategies to successively adapt the models to each participant individually, one timepoint at-a- time. (Aim 2) We will then evaluate whether trajectories in predicted severity scores are able to capture concrete instances of clinical exacerbations (ER visits, hospitalizations). We will compare the sensitivity of scores derived from human raters against the static and adaptive models using an artificial intelligence and clinician interpretation approaches. (Aim 3) Individuals with lived experience with psychosis and community mental health clinicians will be engaged to contribute to study design at the beginning of the study, and then at the end to contextualize findings and develop a subsequent randomized controlled trial. We hypothesize that personalized adaptive models will significantly improve symptom prediction accuracy over static models, with severity scores performing comparably to clinical ratings in detecting exacerbations. Furthermore, engaging community members will ensure person-centered development towards implementation in real-world clinical settings. With these innovative approaches, we aim to introduce a transformative precision approach to measurement-based care for psychosis.