ECG-STAR: Spatio-temporal attention residual networks for multi-label ECG abnormality classification

Abstract

Accurate and timely diagnosis of cardiovascular diseases (CVDs) through automated electrocardiogram (ECG) interpretation is crucial for facilitating early clinical interventions, reducing mortality rates, and improving post-treatment patient outcomes. This paper introduces ECG-STAR (Spatio-Temporal Attention Residual) Net, a novel deep-learning model designed for multi-label classification of various ECG abnormalities, including arrhythmias and other cardiac conditions. ECG-STAR Net integrates linear layers, long short-term memory (LSTM) networks, spatial-temporal convolutions, and efficient channel attention mechanisms, thereby effectively capturing complex patterns and dependencies in ECG signals. Evaluations on three benchmark ECG datasets, PTB-XL, CPSC 2018, and G12EC, demonstrate the proposed model's high performance, achieving AUC scores of 0.9345, 0.9692, and 0.9117, respectively. Furthermore, ECG-STAR Net achieves notable F1 scores of 0.7556, 0.817, and 0.4429, coupled with low Hamming-loss values of 0.1049, 0.035, and 0.0517 across these datasets. A comprehensive ablation study further underscores the contributions of individual model components, providing additional insights into its architecture and effectiveness. These results highlight the ECG-STAR Net's potential for advancing automated ECG diagnosis and enhancing clinical decision-making.