An explainable deep learning framework for biosensing data interpretation in biomedical engineering and real-time health diagnostics.

Introduction: This work proposes an explainable deep learning framework to transform complex biosignal dynamics into interpretable health assessments. The core of our approach is the PhysioGraph Inference Network (PGIN), which combines temporal graph reasoning with probabilistic modeling to capture dynamic inter sensor dependencies under physiological priors. Methods: To further enhance adaptability, an Adaptive Health State Inference Mechanism (AHSIM) is introduced to adjust diagnostic granularity based on uncertainty and signal entropy. Results and discussion: Evaluations on four biosensing datasets show that our framework achieves superior diagnostic accuracy (up to 92.48%) and AUC (up to 93.65%), outperforming several transformer-based baselines such as RoBERTa and T5. Furthermore, the model provides transparent uncertainty estimates, making it suitable for deployment in clinical and wearable scenarios. By integrating physiological semantics and model interpretability, our framework bridges the gap between black-box AI and trustworthy biomedical intelligence.