Relapse prediction using wearable data through convolutional autoencoders and clustering for patients with psychotic disorders.

Relapse of psychotic disorders occurs commonly even after appropriate treatment. Digital phenotyping becomes essential to achieve remote monitoring for mental conditions. We applied a personalized approach using neural-network-based anomaly detection and clustering to predict relapse for patients with psychotic disorders. We used a dataset provided by e-Prevention grand challenge (SPGC), containing physiological signals for 10 patients monitored over 2.5 years (relapse events: 560 vs. non-relapse events: 2139). We created 2-dimensional multivariate time-series profiles containing activity and heart rate variability metrics, extracted latent features via convolutional autoencoders, and identified relapse clusters. Our model showed promising results compared to the 1^st place of SPGC (area under precision-recall curve = 0.711 vs. 0.651, area under receiver operating curve = 0.633 vs. 0.647, harmonic mean = 0.672 vs. 0.649) and added to existing evidence of data collected during sleep being more informative in detecting relapse. Our study demonstrates the potential of unsupervised learning in identifying abnormal behavioral changes in patients with psychotic disorders using objective measures derived from granular, long-term biosignals collected by unobstructive wearables. It contributes to the first step towards determining relapse-related biomarkers that could improve predictions and enable timely interventions to enhance patients' quality of life.