Predicting adverse events in schizophrenia using digital phenotyping and machine learning: results from the HOPE-S observational study

Abstract Digital phenotyping has emerged as a promising tool in psychiatry to monitor patient status and to predict clinically significant events such as relapse and hospitalization in schizophrenia. Digital data from a smartphone and wrist wearable were collected to derive 540,744 daily measurements of behavior and health from 99 patients with schizophrenia spectrum disorders over a six-month period. We found that machine learning algorithms based on gradient boosting trees can achieve a sensitivity of 91.4% and a specificity of 95.3% when predicting hospital readmissions, visits to the emergency room, and deterioration in clinical status from changes in a patient’s digital data alone. Our results provide strong evidence to the nascent body of literature demonstrating that digital phenotyping can be reliably employed to predict relapse and other adverse events in psychosis. *Creighton Heaukulani and Amelia Sim are joint first author.