Detecting momentary reward and affect with real-time passive digital sensor data.

Objectives: This study explores the capability of passive digital sensor data from smartphones and smartwatches to predict self-reported ecological momentary assessments (EMA) of affect, motivation, interest, and pleasure in activities in an unseen test sample. Materials and Methods: Data were collected from 245 depressed participants with high-to-low anhedonia (195 train, 50 test) generating 23 812 EMA sessions. Machine learning models were used to assess the ability of behavioral and physiological features, aggregated over windows of 15 minutes to 3 hours, to predict momentary subjective states. Results: For 12 of 15 EMA questions asked, machine learning models exceeded random chance in the fully-held-out test sample, suggesting detectable signals between passive measures and subjective states. Dependent on the sensor type, the optimal aggregation periods ranged from 15 minutes to 3 hours, with generally at least two hours of data being required. Subgroup analyses revealed variations in model performance by demographics, depression severity, and anhedonia severity. Conclusion: This study establishes the feasibility of using passive digital sensing to detect momentary subjective states, providing a baseline for scalable, non-invasive mental health monitoring.