Cardiac Abnormality Detection from 12-Lead ECG Using Multi-Branch CNN and Attention-Based Feature Fusion

The automatic classification of 12-lead electrocardiogram (ECG) signals plays an important role in diagnosing cardiac abnormalities, providing a faster and more reliable alternative to manual interpretation. In this work, we present a novel deep learning framework for multi-label classification of cardiac conditions directly from 12-lead ECG data. The model utilizes a multi-branch convolutional neural network, where each lead is processed independently through dedicated branches, preserving lead-specific information. The features from all 12 leads are then fused through an attention mechanism, allowing the model to capture inter-lead dependencies and enhance the classification performance. To address class imbalance and the multi-label nature of the task, we employ a weighted binary cross-entropy loss function, which ensures that both rare and co-occurring conditions are accurately classified. The proposed multi-branch convolutional neural network with attention-based feature fusion (MB-CNN-ATT) is validated on the PTB-XL and CPSC datasets, comprehensive collections of 12-lead ECGs from 18869 and 6877 subjects, respectively. The datasets are annotated with five cardiac condition superclasses and nine diagnostic classes, respectively. Our approach achieves macro-average area under the receiver operating characteristic curve (AUC) scores of 93.3% and 96.1% on the PTB-XL and CPSC datasets, respectively, outperforming existing methods such as fully convolutional networks, ResNet variants, and other state-of-the-art architectures. Our work contributes to the field of instrumentation and measurement by advancing the methodology for conditioning and processing physiological electrical signals by integrating lead-wise feature extraction with attention-driven fusion, thereby improving the precision and interpretability of automated cardiac monitoring systems.