ConvResNet: Towards an Explainable Deep Learning Architecture for Mental Stress Detection from ECG Signal

Stress is a natural human reaction when people face any threat, whether from physical, emotional, or environmental sources. This reaction triggers physiological changes that are detectable through different cardiac biomarkers. The electrocardiogram (ECG) signal represents the electrical activity of the heart, and an automated technique for stress monitoring using ECG could serve as an objective alternative to traditional subjective stress assessment methods. In this study, we propose a hybrid residual convolutional and transformer-based architecture, ConvResNet, for mental stress detection from ECG signals. The proposed architecture achieved 94% and 84.7% accuracy, 90.1% and 93.9% sensitivity, 95.4% and 75.6% specificity, 92.8% and 82.8% F1-score, and 0.96 and 0.84 AUC on the WESAD and Vollmer datasets, respectively, using leave-one-out crossvalidation. Deep learning models often lack transparency in their decision-making process; however, explainability is essential in healthcare applications. We have also proposed two approaches based on perturbation and integrated gradients to explain which temporal and spectral regions of the ECG signals are responsible for the model's decision-making. Therefore, the model's transparency in decision-making and the state-of-the-art performance across different datasets highlight its robustness and generalizability. With the potential applicability in wearable devices, the model can help reduce stress-related adversities and improve overall human well-being. The code is available at the following link: https://github.com/SapnilSarker/ConvResNet-WI_IAT.