Interpretable and Scalable Multi-Lead ECG Arrhythmia Detection Using a Selective Multi-Feature Branch Network

Abstract

Precise analysis of cardiac arrhythmias as influenced by multi-lead electrocardiogram (ECG) signals for early diagnosis and clinical intervention is important. However, available deep learning techniques overlook lead-dependent morphological changes or overcompensate with high cost and complexity hence less ideal for real-time and resource-constrained environments. This work introduces SMFB-Net (Selective Multi-Feature Branch Network), an efficient, accurate, and robust deep learning framework for multi-class arrhythmia classification based on 12-lead ECG signals. The architecture proposed is designed to embed an Adaptive Lead Selection Module (LSM) for prioritizing diagnostically informative leads with minimum redundant computations, which resulted in floating point operations (FLOPs) being 30.92% lower than 12-lead processing. For spatial and multi-scale features, convolution and inception blocks are employed and temporal dependencies are modeled with the help of Bidirectional Gated Recurrent Units (Bi-GRUs). The obtained representations are combined, and global average pooling is performed to classify them into nine arrhythmia subgroups. Experimental tests on the publicly available China Physiological Signal Challenge (CPSC) 2018 dataset reveal that SMFB-Net achieves a macro F1-score of 0.864. The model performed particularly well for difficult arrhythmic classes such as first-degree atrioventricular block (I-AVB) and ST-segment deviation abnormalities. Furthermore, SHAP interpretability analysis on patient and population data also confirmed clinically relevant ECG leads. In general, SMFB-Net shows a balance between classification performance, interpretability, and computational efficiency, giving it better overall performance compared to other deep learning techniques. This makes our proposed multi branch framework capable of deployment in real-time and resource-limited arrhythmia monitoring applications.