Don’t Sleep on Sleep Data: Influence of Sleep Physiological Signals on Stress Detection

Soundarya Ramesh, Takahiro Masuda, Lee Hyung Woon, Yongquan Hu, Suranga Chandima Nanayakkara

Keywords: stress detection, sleep, physiological data, wearables, machine learning

TL;DR: We utilize fine-grained previous night's sleep physiological data, together with the current physiological context, to detect stress.

Abstract: Stress is a critical determinant of both short-term well-being and long-term health. While wearable sensors have enabled continuous monitoring of stress through physiological signals, existing approaches that rely only on current physiology have shown limited success. Prior work suggests that the previous night's sleep is predictive of stress, yet current methods typically use only coarse sleep summaries (e.g., duration, resting heart rate). In this paper, we argue that fine-grained sleep physiological data can provide richer insights for stress detection. We collect a month-long smartwatch dataset comprising both day-time and night-time physiological signals, including detailed sleep-derived features, and trained two models -- XGBoost and a custom multi-modal neural network. Our results provide initial evidence that incorporating fine-grained sleep features significantly improves stress detection, opening up several promising directions for future research.

Submission Number: 39