Robust multi-label ECG classification: a deep learning framework for concurrent arrhythmias with clinical generalizability

Abstract

Electrocardiogram (ECG) signals play a critical role in the clinical screening and diagnosis of various cardiovascular diseases. Although computer-aided diagnostic technologies have alleviated the burden on clinicians, the variability, complexity, and class imbalance of ECG data still pose significant challenges to multi-label arrhythmia classification. This study proposes a cascaded deep learning model based on a residual structure, named ResCGT-Net, for multi-label one-dimensional ECG signal recognition and classification. By integrating CNN-extracted waveform features, BiGRU-modeled temporal dependencies, and a Transformer encoder for long-range contextual representation through a residual connection, ResCGT-Net effectively captures multi-level features from ECG signals. Experimental validation on the CPSC 2018 multi-label classification dataset demonstrated that the model achieved an average F1 score of 0.8226, showcasing excellent classification performance. Notably, the Transformer encoder’s multi-head self-attention mechanism provides inherent interpretability by highlighting diagnostically relevant ECG segments, enhancing clinical credibility. Based on single-lead training and testing results, the study provides clinicians with critical lead references for the diagnosis of specific diseases, identifying leads I, II, and AVR as consistently informative across multiple arrhythmia types. Furthermore, gender-stratified analysis revealed that females exhibit more pronounced arrhythmic features compared to males (average F1: 0.8209 vs. 0.7265, t-test, p < 0.01), corroborating clinical findings and highlighting the need for sex-aware diagnostic considerations. The proposed method offers an efficient and reliable solution for automated multi-label arrhythmia analysis, with significant clinical application value.