Personalized Prediction of Momentary Affect with Passive Smartphone Sensing Data in a Clinically Heterogeneous Sample: An evaluation of machine learning and mixed effects approaches

Passive smartphone sensing offers a scalable, low-burden approach to continuous mental health monitoring, yet most studies which use passive sensing features to predict momentary psychological states rely on small, clinically homogeneous samples, limiting generalizability and clinical utility. This study compares methods for personalized prediction of momentary affect, stress, and fatigue using passive sensing in a large, clinically diverse sample of 308 adults (66% with a psychiatric diagnosis). Participants completed 15 days of ecological momentary assessment (23,560 total observations) with passive sensing of GPS, accelerometer, phone call, screen-on-time, and sleep data. We predicted five continuous outcomes: positive affect, negative affect, stress, energy, and a depression composite. We explored methods for balancing group-level and personalized prediction, including mixed effects regression tree (MERT) boosting, a method that blends mixed effects and machine learning approaches. We compared this to group-level pooled models, fully personalized models, and linear mixed models. A key finding was that the use of cluster-mean centering (CMC) as a preprocessing step dramatically improved pooled machine learning model performance, enabling a simple XGBoost model to match or exceed fully personalized and MERT boosting approaches. Best models compared favorably to past work predicting continuous momentary affect, achieving overall R-squared of 0.38-0.49 and average idiographic R-squared of 0.16-0.26 calculated on held out test data, with energy and the depression composite emerging as the most predictable outcomes. These results carry direct clinical implications: scalable group-level models can rival personalized models when data are appropriately preprocessed, lowering barriers to real-world deployment of passive sensing–based monitoring. The high predictability of energy and the depressive composite highlights viable targets for just-in-time adaptive interventions. This work advances digital phenotyping by demonstrating clinically relevant affect prediction at scale and identifying CMC as a critical tool for scalable personalization of mental health monitoring.