ReTeNet: A Residual Encoder and Transformer Encoders Network for Stress Monitoring From Wearable Device

Abstract

Mental stress adversely impacts overall well-being, and chronic stress poses serious risks to both physical and mental health, necessitating the development of wearable-based stress monitoring tools. Blood volume pulse (BVP) sensors, widely integrated into commercially available wearable devices, offer a cost-effective and convenient solution for stress detection, but existing methods face challenges such as data imbalance, complex models unsuitable for real-time use, and limited generalizability. This study presents a lightweight residual-encoder and transformer-encoders network (ReTeNet) for stress monitoring using BVP signals. The proposed model is trained and evaluated on two datasets using a subject-independent, leave-one-subject-out (LOSO) cross-validation strategy: the private RUET SPML dataset with 26 healthy subjects and the publicly available WESAD dataset with 15 healthy subjects. It achieves 93.59% accuracy, 95.60% F1-score, and 0.9569 AUC on the RUET SPML dataset, while attaining 98.23% accuracy, 97.58% F1-score, and 0.9953 AUC on the WESAD dataset. The model effectively distinguishes mental stress with high accuracy while maintaining a lightweight architecture, making it well-suited for wearable devices. Furthermore, its capability to maintain balanced performance across imbalanced datasets highlights its potential for reliable real-time stress monitoring.