Attention-Enhanced Multi-Scale Feature Optimization for Silent Myocardial Infarction and Early Atrial Fibrillation Detection in ECG Signals

Abstract

Cardiovascular diseases remain the leading cause of mortality worldwide, with silent myocardial infarction and paroxysmal atrial fibrillation presenting significant diagnostic challenges due to their subtle electrocardiographic manifestations. This study proposes a novel multi-scale feature extraction method combining residual U-blocks with depthwise separable convolutions and attention mechanisms for enhanced detection of these conditions. The approach integrates clinical domain knowledge through guided spatial attention, emphasizing ST-T segments and high-frequency QRS components while preserving prognostic signals during preprocessing. A comprehensive evaluation of the MIT-BIH Arrhythmia Database, PTB-XL Database, and PhysioNet Challenge 2017 demonstrates 98.7% accuracy for arrhythmia classification, 96.4% sensitivity for myocardial infarction detection, and 97.9% specificity for atrial fibrillation identification, representing improvements of 3.2%, 4.1%, and 2.8% over baseline CNN-LSTM architectures. The optimized network achieves 58% parameter reduction through efficient convolution strategies, enabling deployment on resource-constrained devices. Grad-CAM visualization validates clinical relevance by demonstrating focused attention on diagnostically significant regions of the ECG.