Automated mood disorder symptoms monitoring from multivariate time-series sensory data: getting the full picture beyond a single number

This paper proposes a new task in mood disorder (MD) monitoring: inferring all items in the HDRS and YMRS scales using wearable sensor data. A deep learning pipeline was developed to score MD symptoms in a large cohort of patients, demonstrating clinically significant agreement between predictions and expert clinician assessments (quadratic Cohen’s κ and macro-average F1 score both 0.609). The study investigated solutions to challenges such as multi-task learning, class imbalance, ordinal target variables, and subject-invariant representations. The importance of out-of-distribution testing was also highlighted.

Filippo Corponi, Bryan M. Li, Gerard Anmella, Ariadna Mas, Isabella Pacchiarotti, Marc Valenti, Iria Grande, Antoni Benabarre, Marina Garriga, Eduard Vieta, Stephen M. Lawrie, Heather C. Whalley, Diego Hidalgo-Mazzei, Antonio Vergari