Smart devices and stress level detection: Balancing machine learning model complexity and device intrusiveness

Stress is a common mental health problem that profoundly affects people’s daily lives. In this study, we propose a novel method aimed at classifying subjects’ stress levels using data sourced from smart devices, including mobile phones and wearable devices. Our approach outlines a comprehensive methodology for the classification of stress levels in multiple classes employing advanced machine learning (ML) techniques. Central to our investigation is the exploration of two critical aspects: first, the degree of intrusion required in individuals’ daily routines to accurately assess stress levels using data from smart devices. Second, we explore the trade-off between model complexity and detection accuracy, comparing ML and deep learning (DL) techniques, including the effectiveness of temporal components. We conducted experiments using publicly available datasets that contain physiological and behavioral data. Our results show that transformers and ML techniques obtain the best stress level classification across all data types. Specifically, we achieved high performance using data from wearable devices (98–100% F1-score for wrist and chest bands, respectively) transformed into simple aggregated features. Even with mobile phone data, classical models attained competitive performance (77% F1-score) without relying on complex architectures. Finally, our findings indicate that satisfactory accuracy in stress-level prediction with wearable devices necessitates a moderate level of intrusiveness. This represents an optimal balance, enabling the use of devices that are both practical and comfortable for everyday wear by the general population. This ensures reliable stress detection without compromising user compliance, thereby making these wearables viable for large-scale, real-world applications.