ChillMate Real Time Wrist Based Stress Detection with Motion Aware Personalization on Smartwatches

Abstract

Chronic stress is a growing health concern, and consumer smartwatches offer a practical way to monitor it continuously. Moving laboratory-grade stress detection onto wearable devices, however, raises challenges that are rarely addressed together: real-time feature extraction from limited sensors, motion artifact rejection, and inter-individual physiological variability. We present ChillMate, an end-to-end stress detection system for smartwatches that combines heart rate variability, motion, and rolling personalization features in a single deployed machine-learning pipeline. Unlike prior wrist studies that rely on additional biosignals such as electrodermal activity, the system uses only signals available on a mass-market watch: heart rate, heart rate variability and accelerometer data. An Extra Trees classifier trained on the WESAD wrist subset reaches F1 score of 0.781 and ROC-AUC of 0.948 under subject-independent cross-validation. An ablation study shows that motion and personalization features each contribute measurably, and a cross-dataset analysis on SWELL-KW quantifies the difficulty of domain transfer. Together, these results show that heart rate variability and motion based stress detection can run end-to-end on unmodified consumer hardware, while the cross-dataset analysis makes explicit the limits of such models under domain shift.