Prospective Practice-Oriented Research Project: Predicting hospitalization and disengagement with automated speech and language analysis

Feinstein Institute For Medical Research, Manhasset NY

Investigators

Abstract

Abstract The goal of this prospective practice-oriented project is to develop a means for predicting rehospitalization and disengagement in coordinated specialty care (CSC) for psychosis. CSC has demonstrated efficacy in improving outcomes, yet challenges persist with rehospitalization and treatment disengagement. There is an urgent need for predictive tools that are scalable and efficient. Automated speech and language analysis holds several advantages: they are accurate markers of mental states related to psychosis, assessments can be conducted repeatedly and longitudinally, and little specialized equipment or training is required to collect necessary speech samples. The study will be conducted within the ESPRITO hub of the EPINET consortium, drawing upon the expertise of a multidisciplinary team in psychosis, early intervention, and natural language processing. Feature extraction will be approached through data-driven (Aim 1A/2A) and insight-driven (Aim 1B/2B) methods based on existing datasets. The existing data for Aim 1 will include 200 clinical notes describing individual participants’ early warning signs for psychosis relapse, and relevant data for Aim 2 will include 400 psychotherapy progress notes from the health records. Themes related to the outcomes of interest will be automatically identified using topic modeling. Features will be derived by calculating the semantic distance between the identified themes and the content of prospectively collected transcripts. In the insight- driven approach, the study team will review and interpret existing data in the context of prior work and clinical expertise; hypothesis-driven acoustic and NLP features will be identified and calculated. Prospectively collected speech data from psychotherapy sessions (n=150 participants; ~1200-1500 sessions) will then be used to develop and test predictive algorithms for hospitalization (Aim 1C) and disengagement (Aim 2C). The outcomes will be considered in a time-to-event setting as well as in a binary variable setting (hospitalization in 30 days & 90 days, disengagement in 90 days & 6 months). The primary analytic strategy for time-to-event outcomes will be based on regularized Cox regression, which allows for the selection of features significantly associated with hazard function for hospitalization or disengagement. Additionally, with a focus on improving prediction accuracy, machine learning-based survival and classification models will be considered (e.g. Deep Convolutional and Deep Recurrent Neural Network models: CNN-Surv & RNN-Surv). Speech and language features recorded from all sessions across all participants will be used. Training and testing for all models will be based on the standard 60%/20%/20% random split into training, validation, and test datasets. Key deliverables include the development of speech-based tools capable of identifying individuals at risk for rehospitalization and treatment disengagement in real-time, which enables timely interventions to mitigate adverse outcomes. Successful completion of this project could pave the way for future randomized clinical trials to evaluate the effectiveness of real-time implementation in improving outcomes within CSC.