Explainable attention-guided deep learning for physiological signal processing in body area networks

The incorporation of Body Area Networks (BANs) supports wearable device-based health monitoring and real-time physiological signal processing, where Human Activity Recognition (HAR) acts as a significant medical application. As the demand for smart health monitoring and personalized wellness increases, the utilization of multimodal physiological signals to identify human activities has become a major field of study. Although information from various sensors offers valuable insights on recognizing human activity, such datasets frequently suffer from problems like noise, missing data, redundant data, complex outliers, and irregular relations that can affect the risk of precisely detecting actions. Moreover, although conventional artificial intelligence (AI) and deep learning (DL) methodologies frequently generate better outcomes, they function as “black boxes” and lack intelligibility, which restrains their reliability for real-time clinical applications. Therefore, this study presents an Explainable Artificial Intelligence Driven Multimodal Physiological Signal Analysis framework (XAI-MPSA) for Human Activity Recognition. The proposed model applies the minimum redundancy maximum relevance method to select informative features while reducing redundancy. To further enhance feature representation, an attention-based feature refinement mechanism is introduced to capture complex relationships among multimodal physiological signals. A convolutional neural network-transformer-attention-based classification approach is then employed to learn both local feature patterns and long-range temporal dependencies in the signal sequences. Finally, Bayesian Optimization and HyperBand are applied to tune the model parameters for better performance, while SHapley Additive exPlanations are utilized to interpret the model predictions by identifying the contribution of individual physiological features in the final decision. Comprehensive experimental analysis demonstrates that the proposed XAI-MPSA achieves robust and interpretable human activity recognition performance with an accuracy of 95.43%, highlighting its applicability in intelligent healthcare and physiological monitoring systems.