Abstract / Summary
Electrocardiogram (ECG) signals are non-stationary and exhibit complex temporal dependencies, making automated analysis challenging. Current deep learning methods tend to either extract spatial or temporal features separately, which do not provide a complete signal representation. In this paper, we propose KT-CBHNET, an interpretable spatio-temporal deep learning framework that combines staged reuse of learned ECG representations with CNN-based morphological feature extraction and BiLSTM-based temporal modeling for cardiovascular disease classification. This model combines a 1D CNN to extract local morphological features, with a bidirectional long short-term memory (BiLSTM) network to capture long-range temporal dynamics.An in-domain transfer-learning strategy initializes the hybrid architecture using CNN representations pretrained on PTB-XL, followed by joint optimization of the transferred CNN and newly introduced BiLSTM layers. To achieve interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM) is used to indicate the clinically relevant waveform portions that impact the classification decisions, showing that the model’s focus is on clinically relevant portions. The results show that the accuracy of the proposed model (KT-CBHNET) is 94.61% on the PTB-XL dataset, the precision is 91.4%, the recall is 97.79%, the F1 score is 94.49%, and the AUC is 98.02%, which indicates that the model is accurate and interpretable for the ECG signal classification problem.