A Shallow Deep Learning Model for Stress Monitoring from Photoplethysmography Signals

Chronic stress poses significant risks to physical and mental health, emphasizing the need for continuous and unobtrusive monitoring using wearable devices. Blood volume pulse (BVP) and electrodermal activity (EDA) are cost-effective, noninvasive signals that capture complementary cardiovascular and autonomic responses. In this study, we propose a lightweight multimodal deep learning framework that integrates convolu- tional neural networks (CNNs) and bidirectional gated recurrent units (BiGRUs) to jointly learn spatial and temporal features. BVP and EDA signals are processed through separate pipelines, and their fused representations are refined through convolution, pooling, and global average pooling layers. To address class imbalance, a sliding-window augmentation strategy is employed. The model is trained and evaluated on the publicly available WESAD dataset using a subject-independent leave-one-subject- out (LOSO) cross-validation protocol. It achieves 98.56% ac- curacy, 99.19% AUC, and 96.87% Cohen’s κ, outperforming existing baselines. With only 0.48M parameters, the architecture is computationally efficient and suitable for real-time stress monitoring on resource-constrained wearable devices.