Mmstress
Liang Kun,Anfu Zhou,Zhan Zhang,Hao Zhou,Huadóng Ma,Chenshu Wu
DOI: https://doi.org/10.1145/3610926
2023-01-01
Proceedings of the ACM on Interactive Mobile Wearable and Ubiquitous Technologies
Abstract:Long-term exposure to stress hurts human's mental and even physical health,and stress monitoring is of increasing significance in the prevention, diagnosis, and management of mental illness and chronic disease. However, current stress monitoring methods are either burdensome or intrusive, which hinders their widespread usage in practice. In this paper, we propose mmStress, a contact-less and non-intrusive solution, which adopts a millimeter-wave radar to sense a subject's activities of daily living, from which it distills human stress. mmStress is built upon the psychologically-validated relationship between human stress and "displacement activities", i.e., subjects under stress unconsciously perform fidgeting behaviors like scratching, wandering around, tapping foot, etc. Despite the conceptual simplicity, to realize mmStress, the key challenge lies in how to identify and quantify the latent displacement activities autonomously, as they are usually transitory and submerged in normal daily activities, and also exhibit high variation across different subjects. To address these challenges, we custom-design a neural network that learns human activities from both macro and micro timescales and exploits the continuity of human activities to extract features of abnormal displacement activities accurately. Moreover, we also address the unbalance stress distribution issue by incorporating a post-hoc logit adjustment procedure during model training. We prototype, deploy and evaluate mmStress in ten volunteers' apartments for over four weeks, and the results show that mmStress achieves a promising accuracy of ~80% in classifying low, medium and high stress. In particular, mmStress manifests advantages, particularly under free human movement scenarios, which advances the state-of-the-art that focuses on stress monitoring in quasi-static scenarios.