StressBVPNet: An Explainable Deep Learning Model for Stress Monitoring from BVP Signals

Mental stress significantly impacts both physical and psychological health, with prolonged exposure leading to severe complications. Blood Volume Pulse (BVP) sensors, widely integrated into wearable devices, offer a non-invasive, cost-effective, and practical solution for continuous stress monitoring. This paper proposes StressBVPNet, a lightweight hybrid deep learning architecture that integrates convolutional neural networks (CNN), bidirectional long short-term memory (BiLSTM), and bidirectional gated recurrent units (BiGRU) to enable real-time stress detection using only unimodal BVP signals. The model was evaluated on two datasets: a self-collected RUET SPML dataset and a publicly available WESAD dataset. It achieved 92.09% accuracy, 84.21% specificity, 94.72% sensitivity, 94.72% F1-score, 93.96% AUC, and 78.83% Cohen's kappa on the RUET SPML dataset, and 97.90% accuracy, 98.30% specificity, 97.20% sensitivity, 97.06% F1-score, 99.12% AUC, and 95.42% Cohen's kappa on the WESAD dataset, outperforming existing approaches. To enhance model transparency, we introduce a novel explainability technique that analyzes band-specific dependencies in the BVP signal based on cosine similarity, revealing physiologically meaningful insights into the model's decision-making. With only 0.182M parameters, StressBVPNet is highly efficient and suitable for real-time deployment on resource-constrained wearable platforms, offering strong potential for practical, continuous stress monitoring in daily life.