Multi-Label Cardiovascular Disease Classification from 12-Lead ECG Using Enhanced DenseNet-121 with CBAM and Grad-CAM Explainability

Abstract

Early and accurate diagnosis of cardiovascular diseases (CVDs) is critical, as they are one of the most prevalent causes of death across the globe. We propose a multi-label classification model consisting of an Enhanced DenseNet-121 (EDN-121) deep learning framework for 12-lead ECG signals obtained initially from the CPSC 2018 dataset (in total 6,877 samples) for eight cardiovascular conditions. Data Preprocessing: A Butterworth high-pass filter was employed to remove noise from the ECG signals, followed by normalization to ensure the data was within a consistent range. Then, the Synthetic Minority Oversampling Technique (SMOTE) was applied to fix the class imbalance problem. EDN-121 architecture incorporates a Convolutional Block Attention Module (CBAM), which contributes to superior spatial and channel-wise representation features. Additionally, Gradient-weighted Class Activation Mapping (Grad-CAM) ensures the explainability of the prediction by identifying the clinically relevant region on the ECG. Experimental results based on five-fold cross-validation show that it outperforms the baseline VGG-16 with 97.90% accuracy, 97.75% F1-score, and 97.34% recall. Conclusion: This study demonstrates the performance and interpretability of a machine learning framework that has the potential to be implemented in clinical settings for the real-time, automated ECG-based CVD screening.