Identifying Depression Symptom Severity and Heterogeneity via Unsupervised Clustering of Wearable and Smartphone Sensor Dat
Annotatsiya
Depression is a prevalent and debilitating mental health condition marked not only by its severity but also by substantial heterogeneity and symptom fluctuation over time. Conventional assessment methods, which rely on sporadic self-reports and clinical evaluations, often fail to capture these temporal dynamics. Continuous monitoring through wearable devices and smartphones offers a promising avenue to collect real-time, multi-modal data that may reveal underlying symptom patterns and subtypes of depression. Objective: This study investigates the use of wearable electrocardiogram (ECG), accelerometer (ACC), and smartphone sensor data to monitor depressive symptoms in both patients and non-depressed participants. We apply unsupervised clustering on top of deep representations of the multivariate time series to uncover data-driven patterns and examine their associations with self-reported measures of depression severity. Methods: Over 35 days, participants wore a patch to continuously record ECG and ACC signals, while passive smartphone sensing captured behavioral and contextual information. Participants completed weekly Beck Depression Inventory-II (BDI-II) and daily ecological momentary assessments (EMA) via a mobile app. A clinician-rated Hamilton Depression Rating Scale (HAMD-17) was administered at baseline and predefined timepoints. Deep embedding features were extracted using various autoencoder architectures, and unsupervised clustering was applied to identify distinct participant groups. The resulting clusters were then statistically analyzed to discover significant differences between them. Results: Six distinct clusters emerged, corresponding to varying levels of depressive symptom severity as measured by BDI-II. Clusters differed significantly across specific symptoms and multimodal digital features, including physiological patterns, social interaction metrics, and affective states. Conclusion: Unsupervised clustering of deep representations from wearable and smartphone sensor data can reveal clinically meaningful subgroups within depression. These findings support the potential of digital health technologies for continuous, personalized monitoring and suggest new directions for tailoring interventions to individual symptom trajectories.