Can 1-Year Smartphone and Wearable Data allow Individual-level Modeling of Depressive Symptom Severity?

Background Objective markers of depression severity in people with Major Depression (MD) would be highly relevant for monitoring disease trajectories and treatment effects. Passive sensing via mobile devices offers a means to capture continuous behavioral and physiological data patterns related to depression. Shortcomings of most prior studies include short observation periods, infrequent symptom assessments, limited data sources, or group-level analyses that obscure within-person variation. Therefore, this study tested whether individual-level models built from 1-year daily mobile sensing and self-report data can provide markers of daily self-rated depression severity (Patient Health Questionnaire-2, PHQ-2) in patients with MD, and whether non-linear ensemble models outperform linear models in detecting such markers. Methods Daily PHQ-2 ratings over up to one year, together with continuous smartphone and smartwatch data capturing daily physical activity, sleep–wake patterns, heart-rate, phone-based social interaction, app usage, and brief prompted speech samples, were collected from 15 adults with recurrent MD. Four patients were excluded due to excessive missing PHQ-2 scores. Linear elastic net (EN) regression was applied on the remaining 11 patients to estimate PHQ-2 scores from the mobile sensing data. For each individual, the first 70% of the time series was used to train the model, and the later 30% was kept separate and used only for testing. Model performance on the unseen test period was evaluated using predefined criteria (R² > 0.30 or mean absolute error (MAE) < 1). In addition, non-linear random forest models were tested to determine whether they increased the number of participants for whom PHQ-2 severity could be estimated using the same performance criteria. Results Using individual-level elastic net models, none of the 11 participants met the predefined performance criteria on unseen test data. Non-linear random forest models identified one participant (9.1%) meeting the performance threshold based on MAE. No participant reached the R²-based performance threshold under either modeling method. Conclusions 1-year mobile sensing does not allow the identification of valid indicators of depression severity to support MD patients with self-management of depression. A 1-year observation period might not provide enough variance in depression severity to identify such indicators at the individual level. Trial registration: German Clinical Trials Register (DRKS), DRKS00032618. Registered 12 September 2023.