Modern healthcare systems increasingly rely on artificial intelligence for clinical decision support. While existing approaches achieve high diagnostic accuracy, they often fail to provide clinically meaningful explanations that align with medical reasoning processes, and their computational demands exceed the capabilities of resource-constrained edge devices essential for healthcare 4. 0 applications. This study proposes XMedFuse, an explainable multimodal feature fusion framework that integrates morphological and temporal features from physiological signals through a lightweight architecture combining depthwise separable convolutions for efficiency with dense connectivity for feature reuse. The morphological pathway extracts hierarchical waveform characteristics while a parallel temporal dynamic branch captures rhythm patterns, with adaptive fusion mechanisms dynamically weighting each modality's contribution. Gradient-weighted attribution mapping provides interpretable visualizations revealing which signal regions drive diagnostic decisions, enabling clinicians to validate recommendations against medical knowledge. Validation on cardiovascular diagnostic tasks using two datasets demonstrated 98. 4% internal testing accuracy and 96. 7% external testing accuracy with only 1. 7-point generalization degradation, confirmed through both hold-out evaluation and 5-fold cross-validation (98. 2±0. 8% mean accuracy). Statistical significance testing via paired t-tests confirmed genuine improvements over all baseline methods (p< 0. 001). The framework operates with 48. 2K parameters and 0. 42 billion operations, substantially more efficient than baseline methods while achieving superior diagnostic performance even under noisy conditions (maintaining above 93% accuracy at SNR=15dB). Interpretability analysis showed 94% alignment between model attention patterns and clinical diagnostic criteria, confirming that XMedFuse learns medically meaningful decision strategies suitable for transparent clinical deployment.
Li et al. (Thu,) studied this question.