An Attention Driven Deep Learning Architecture for Stress Monitoring Using Single Channel Electrocardiogram Signal

Stress is a physiological response to mental, physical, and emotional challenges, often reflected by alterations in cardiac rhythm measurable from electrocardiogram (ECG) signals. Existing machine learning and deep learning approaches for ECG-based stress detection often trade performance for computational cost, particularly when relying on multimodal sensing or 2D signal transformations. This study proposes a novel attention-based convolutional neural network with bidirectional long short-term memory (CNN-BiLSTM) for automatic stress detection from single-channel raw ECG. The model was evaluated on two public datasets, WESAD (WD) and Vollmer (VD), and achieved strong performance across multiple metrics, including F1 score (WD: 99.6%, VD: 93%), accuracy (WD: 99.7%, VD: 93.1%), sensitivity (WD: 99.2%, VD: 91.4%), specificity (WD: 100%, VD: 94.8%), and precision (WD: 100%, VD: 94.6%). Feature visualisation and integrated gradients were used to interpret the model and explain its predictions. The proposed method outperforms existing raw single-channel ECG-based models and is competitive with many multimodal approaches for binary stress classification. Deployment on a Raspberry Pi further demonstrates suitability for real-time wearable stress monitoring.