Modular deep residual network for driver stress detection based on photoplethysmography

Conventional approaches to driver stress detection frequently depend on complex multimodal sensor systems, which involve intrusive and costly data acquisition procedures. In contrast, this study explores the hypothesis that the photoplethysmography (PPG) signal can independently provide reliable indicators of stress. To evaluate this premise, we present a modular end-toend deep residual network specifically designed to balance high classification accuracy with relatively low computational complexity. The architecture incorporates several optimization techniques, starting with optimal design of input layers that reduce spatial resolution and extract key features, eliminating the need for manual feature engineering. By leveraging bottleneck residual blocks enhanced with multi-branch mechanisms, our model improves overall accuracy. Each branch within the block applies distinct convolutional operations, enabling the extraction of both global and local features across multiple hierarchical levels. A comprehensive evaluation on two public datasets demonstrates that our model achieves accuracy over 98%, outperforming state-of-the-art methods while balancing complexity and performance. These findings highlight the model’s potential for integration into real-world driving environments.