Lead-Aware Hierarchical Transformer and Convolution Fusion Network for ECG Classification

ECG classification is a typical and practical multivariate time series classification task. Recently, ECG classification with deep networks has been widely researched and achieved promising results. However, most of existing works focus on multi-class rather than multi-label ECG classification, while the latter is more clinically practical. Moreover, information within and between specific ECG leads can be further mined and utilized by deep learning models. Therefore, we develop Lead-aware Hierarchical Transformer and Convolution fusion Network (LHTC-Net). It integrates an Attention Convolution module and a Hierarchical Transformer module to extract both local and long-term dependency in ECG signals. In constructing the hierarchical transformer, we design three novel Window-based Transformer Blocks dedicated to local, global, and lead-specific information respectively. Additionally, a lead-aware mechanism is proposed to capture lead-specific information. Experiments show that LHTC-Net outperforms five SOTA methods with 82.67% and 78.53% micro F \(_{1}\) scores on two real-world datasets. Extensive ablation studies demonstrate the roles of different modules of LHTC-Net, providing insights into our algorithm and ECG classification task.