Digital phenotyping of affect and stress in emerging adults.

Background: Depression is highly heterogeneous and difficult to monitor or predict in daily life. One strategy for monitoring depressive symptoms is digital phenotyping, the real-time tracking of behaviors via personal devices. Digital phenotyping may be especially useful for predicting mood in emerging adults, a developmental period characterized by heightened rates of depression and smartphone use. However, prior research lacks long-term data and rigorous comparison of modeling approaches. Objective: The present pre-registered study addresses these gaps by modeling smartphone sensor-based behavioral markers to predict affect and stress over a one-year period in emerging adults ( Methods: Measures included daily ecological momentary assessment and continuous collection of sensor data from smartphones. Behaviors were estimated as features reflecting sleep, activity, mobility, and phone use, and tested as predictors of daily affect and stress in supervised machine learning models. Comparisons were performed across idiographic and nomothetic XGBoost models: Model 1 Group General that estimated behaviors that predict affect/stress across people, Model 2 Group Personalized that allowed for person-level estimation of behaviors that predict affect/stress, and Model 3 Within-Person Personalized which fit models independently for each participant. Results: Models provided complementary insights and showed different strengths and weaknesses. Model 1 Group General identified behaviors that predicted daily affect/stress across people, but showed poor generalizability. The personalized models (Model 2-3) outperformed general models, and Model 2 Group Personalized offered the best balance of accuracy and stability across evaluation metrics. However, participant ID contributed most of that model's predictive power, suggesting that the model primarily captured stable individual differences in affect/stress. In contrast, Model 3 Within-Person Personalized revealed person-specific patterns of behaviors predicted daily affect, but model reliability was limited. Conclusions: Findings reveal complementary strengths and weaknesses from machine learning models spanning the idiographic-nomothetic spectrum, for predicting affect and stress from passively sensed behavioral features. These results highlight the need for future research to rigorously compare, and quantify strengths and weaknesses, of personalized and hybrid modeling strategies that predict affective and stress outcomes. Insights from this study can guide future digital phenotyping research, which is crucial for exploring translational applications.