Predicting Relapse Risk in First Episode Psychosis using Machine Learning and Sleep Data

TY - JOUR

T1 - Predicting Relapse Risk in First Episode Psychosis using Machine Learning and Sleep Data

AU - Georgiades, Anna

PY - 2026/9/2

Y1 - 2026/9/2

N2 - Sleep disturbance is common in First Episode Psychosis (FEP) and is associated with psychosis symptom exacerbation and functional decline, yet it remains unclear which sleep features predict relapse. In this longitudinal study, 269 FEP participants reported daily sleep during the first 30 days of the 12-month follow-up period via a smartphone app, including difficulty falling asleep, wake-up frequency, insufficient sleep, and sleep duration. These sleep characteristics were used to predict relapse and positive symptoms over the 12-month follow-up period, with outcomes assessed using structured interviews and clinical records. A Random Forest classifier was developed and validated using stratified nested cross-validation, achieving an AUC of 0.75 ± 0.13 and accuracy of 0.72 ± 0.16. Insufficient sleep and irregular sleep duration were identified as the strongest predictors of relapse risk. These findings were used to develop an open-access interactive web-based platform (a research prototype not intended for clinical use) to predict risk of relapse based on sleep disturbance, demonstrating the potential for real-time monitoring and personalised relapse prevention.

AB - Sleep disturbance is common in First Episode Psychosis (FEP) and is associated with psychosis symptom exacerbation and functional decline, yet it remains unclear which sleep features predict relapse. In this longitudinal study, 269 FEP participants reported daily sleep during the first 30 days of the 12-month follow-up period via a smartphone app, including difficulty falling asleep, wake-up frequency, insufficient sleep, and sleep duration. These sleep characteristics were used to predict relapse and positive symptoms over the 12-month follow-up period, with outcomes assessed using structured interviews and clinical records. A Random Forest classifier was developed and validated using stratified nested cross-validation, achieving an AUC of 0.75 ± 0.13 and accuracy of 0.72 ± 0.16. Insufficient sleep and irregular sleep duration were identified as the strongest predictors of relapse risk. These findings were used to develop an open-access interactive web-based platform (a research prototype not intended for clinical use) to predict risk of relapse based on sleep disturbance, demonstrating the potential for real-time monitoring and personalised relapse prevention.

KW - First Episode Psychosis

KW - Machine Learning

KW - Positive Symptoms

KW - Relapse

KW - Sleep

M3 - Article

SN - 2158-3188

JO - Translational Psychiatry

JF - Translational Psychiatry

ER -