A deep learning framework for multi-lead ECG arrhythmia classification with CAM interpretability and LLM-driven clinical guidance.

Cardiac arrhythmias remain a major cause of morbidity and mortality, requiring accurate and interpretable automated diagnosis. This reserach presents a lightweight deep learning framework for multi-lead ECG arrhythmia classification. A Winograd-scaled 1D MobileNet efficiently extracts discriminative features, while the Lead-aware Skip Weighting Residual ConvNeXt Attention with Builder Optimizer (LSW-RCABO) enhances classification through adaptive residual weighting and lead-aware attention. Grad-CAM provides visual explanations by highlighting waveform regions, and BioGPT generates patient-specific clinical recommendations using classification results and metadata. Experimental evaluation on the MIT-BIH, 12-lead ECG, and PhysioNet 2020 datasets demonstrates superior performance, achieving classification accuracies of 99.98%, 99.98%, and 99.97%, respectively.