Individual-Level Modeling of Depressive Symptom Severity Using Smartphone and Wearable Data: Time-Aware 1-Year Study...

Smartphones and wearables can continuously capture behavioral and physiological data in everyday life. Such mobile-sensing data may help track depressive symptoms more closely than occasional retrospective questionnaires, but prior findings have been mixed. Inconsistent findings do not preclude the presence of predictive relationships in specific individuals or time periods. In addition, it remains unclear which broader sensor domains, rather than single features, contribute most to prediction at the individual level.

The objective of this study is to evaluate whether smartphone and wearable data improve prospective prediction of daily depressive symptom severity beyond a person-specific baseline and, where improvement is observed, to identify contributing sensor feature groups.

Data from the MONDY (secure and open platform for AI-based health care apps) n-of-1 study were analyzed, including adults with recurrent major depressive disorder who provided up to 12 months of daily self-reported depressive symptom ratings, together with continuous smartphone and smartwatch data. To capture individual-specific patterns, separate participant-specific (idiographic) linear and nonlinear models were developed for each individual. Performance was evaluated using repeated temporally separated training and test periods against a person-specific baseline model predicting the mean symptom score from the training period. Explainability analyses were restricted to models showing improvement over baseline, and summarized at the level of sensor feature groups.

Among 11 participants with sufficient data, linear models improved predictive performance over the person-specific baseline in 8 (73%) participants, and nonlinear models did so in a different set of 8 (73%) participants. Improvements over the person-specific baseline, expressed as reductions in mean absolute error (change in mean absolute error relative to baseline), ranged from 0.05 to 1.23 points for linear models and from 0.12 to 1.07 points for nonlinear models on the adapted Patient Health Questionnaire-2 scale. Predictive performance varied substantially between individuals, and although nonlinear models exhibited greater aggregate feature-group contributions, this did not translate into a consistent overall predictive advantage over linear models. Feature-group importance profiles were heterogeneous, and no single sensor domain dominated across all participants.

Multimodal smartphone and wearable data can provide individual-specific predictive signals for daily depressive symptom severity in a subset of patients when evaluated under prospective, time-aware conditions. Both predictive performance and contributing sensor domains vary markedly across individuals, and no single modeling approach is uniformly superior. These findings suggest that previously reported associations in mobile-sensing research, often arising from heterogeneous study designs and feature representations, may not consistently translate into prospective prediction, and highlight the importance of evaluation frameworks that assess predictive signal at the temporal individual-participant level.