Predicting Psychotic Relapse in Schizophrenia With Mobile Sensor Data: Routine Cluster Analysis.

Area of Science:

• Computational psychiatry
• Digital phenotyping
• Machine learning in healthcare

Background:

• Behavioral representations from mobile sensing data aid in predicting psychotic relapse in schizophrenia.
• Timely interventions can mitigate relapse severity and frequency.

Purpose of the Study:

• Develop clustering models to derive behavioral representations from multimodal mobile sensing data.
• Utilize these representations for schizophrenia relapse prediction.

Main Methods:

• Employed Gaussian mixture model (GMM) and partition around medoids (PAM) clustering on mobile sensing data from 63 schizophrenia patients.
• Trained a personalized relapse prediction model using features derived from clustering, incorporating age-based similarity for personalization.
• Evaluated model performance using a leave-one-patient-out approach.

Main Results:

• Clustering models identified distinct behavioral patterns, differentiating routine and atypical trends.
• GMM clusters showed higher density for routine behaviors, while PAM clusters offered more homogeneous characterization.
• Relapse prediction model achieved an F2 score of 0.23, significantly exceeding the random baseline (0.042).

Conclusions:

• Clustering mobile sensing data effectively captures behavioral trends in schizophrenia patients.
• Derived features from clustering models are predictive of oncoming psychotic relapse.
• This digital phenotyping approach supports early detection and timely intervention strategies.