Leveraging Smartphone Human Interaction Routine Behavior Task Mining and Modeling for Daily Stress Monitoring

With advancements in mobile sensing technologies, there is a growing need for scalable and interpretable stress monitoring solutions that remain robust over time. Existing smartphone passive sensing approaches rely on statistical app usage features or multi-modal sensor data, making them susceptible to distribution shift and feature evolution (e.g., schema drift), and adding model complexity. To address these challenges, we propose the Smartphone Human Interaction-based Routine Behavior Task Mining, Modeling, and Feature Extraction (SHIRBT-MMF) framework, which models stress-related behaviors by mining finegrained interaction routine tasks rather than aggregated app usage patterns. SHIRBT-MMF leverages multi-level sequential pattern mining and large language model-based automated task modeling to extract interpretable and stable features from within-app UI state transitions. Unlike traditional methods that require hundreds of apps, SHIRBT focuses on a small, consistent set of routine-based tasks, mitigating covariate shift and feature evolution while improving model robustness. We validated SHIRBT-MMF through one- and four-month in-the-wild datasets with 26 participants, demonstrating that the SHIRBT-based personalized model achieves an average accuracy of 75%, outperforming baseline models by 5% while using only 3-6% of app types. Additionally, SHIRBT features remain stable over time, reducing covariate shift and ensuring reliable performance. With its expandability to other mental health, interpretability, and privacy-conscious design, the SHIRBT-MMF framework lays the foundation for personalized digital mental health monitoring.