Relapse Detection in Patients with Psychotic Disorders Using Unsupervised Learning on Smartwatch Signals

Relapse detection is a crucial component of mental disorders treatment and management. In this paper, we present our solution for the ICASSP Signal Processing Grand Challenge e-Prevention track 2 Relapse Detection. We propose an unsupervised learning approach to detect relapse in patients with mental health disorders using anomaly detection with an autoencoder. To this end, we extract features from sleep, physical activity, and physiological data recordings. We train an autoencoder exclusively on non-relapse data. To detect relapses, we calculate the reconstruction error of the autoencoder and use it as an anomaly score. Our team, Emotion, ranked second in the e-prevention challenge with a ROC-AUC score of 0.6 and a PR-AUC score of 0.63 on the testing dataset.