Deep Temporal Modeling of Longitudinal Wearable Signals for Unsupervised Detection of Depressive Relapses

Abstract

Depression is a recurrent mental health disorder whose relapse episodes significantly impact long-term outcomes and quality of life. Conventional monitoring relies on clinical interviews and self-report questionnaires, which provide intermittent and often subjective assessments. Wearable devices offer a scalable alternative by enabling continuous, nonintrusive monitoring of physiological and behavioral signals in naturalistic settings. This thesis investigates the unsupervised detection of depressive relapses through deep temporal modeling of longitudinal wearable data collected within the e-Prevention framework. Multivariate time series derived from smartwatch recordings, including heart rate variability, physical activity, sleep metrics, and step counts, are analyzed to identify relapse events as anomalies in temporal dynamics. We first constructed an expanded dataset by adding extra monitoring days for the same patient cohort, enabling more comprehensive longitudinal evaluation. We then proposed transformer-based autoencoder (AE) and variational autoencoder (VAE) architectures to learn compact latent representations, incorporating multitask learning objectives to improve representation quality. We systematically evaluated multiple imputation strategies to handle missing data and integrated selected sleep and activity features alongside baseline physiological signals to provide complementary behavioral information. Results indicate that AE/VAE and multitask formulations provide improvements on the initial dataset, while on the expanded dataset data handling, particularly KNN imputation combined with temporal post-processing, plays a more critical role than architectural modifications. Incorporating selected sleep and activity features further improves performance on the expanded dataset by providing complementary behavioral context. Overall, these findings highlight the potential of transformer-based deep temporal models for continuous, unsupervised monitoring of depressive relapses using wearable devices.