XMedFuse: An Explainable Multimodal Feature Fusion Framework for Healthcare Diagnostics

Area of Science:

• Artificial Intelligence in Medicine
• Biomedical Signal Processing
• Machine Learning for Healthcare

Background:

• AI enhances clinical decision support but often lacks medical reasoning and efficiency for edge devices.
• Existing AI models struggle with interpretability and computational demands in healthcare 4.0.
• Need for explainable, efficient AI solutions for real-time medical diagnostics on resource-constrained devices.

Purpose of the Study:

• To develop XMedFuse, an explainable multimodal feature fusion framework for AI-driven clinical decision support.
• To integrate morphological and temporal features from physiological signals using a lightweight architecture.
• To ensure AI recommendations are interpretable and align with medical reasoning for clinical validation.

Main Methods:

• Proposed XMedFuse framework using depthwise separable convolutions and dense connectivity for efficient feature extraction.
• Integrated morphological (waveform) and temporal (rhythm) signal pathways with adaptive fusion.
• Employed gradient-weighted attribution mapping for interpretable decision visualizations.

Main Results:

• Achieved 98.4% internal and 96.7% external testing accuracy on cardiovascular diagnostic tasks.
• Demonstrated high efficiency (48.2K parameters) and robustness in noisy conditions (93% accuracy at 15dB SNR).
• Interpretability analysis showed 94% alignment between model attention and clinical criteria, confirming medical relevance.

Conclusions:

• XMedFuse offers a highly accurate, efficient, and interpretable AI solution for clinical decision support on edge devices.
• The framework's explainability enables clinicians to validate AI-driven diagnostic recommendations against medical knowledge.
• XMedFuse represents a significant advancement towards transparent and reliable AI deployment in modern healthcare systems.