Abstract
This research presents a semi-supervised hybrid Temporal Convolutional Network-Long Short-Term Memory (TCN-LSTM) model for interpretable and data-efficient ElectroCardioGram (ECG) heartbeat classification. ECG signals from the MIT-BIH and INCART databases were resampled at 125 Hz, 4th order Butterworth filtered (0.5-20 Hz), and segmented into 0.8 s (188-sample) windows (279,641 beats). The architecture integrates two Temporal Convolutional Network (TCN) blocks (kernel = 3, receptive field = 63) with parallel 64-unit Long Short-Term Memory (LSTM) layers fused via element-wise maximum to capture both local and global temporal dynamics. Data were split beat-wise (60/20/20 for SL; 80/20 for SSL), with 10-30% labeled beats and pseudo-labels generated using adaptive thresholding. The model achieved 0.980 accuracy and an F1-score of 0.870 in supervised learning and 0.979 accuracy and an F1-score of 0.850 in semi-supervised mode using 30% labeled data, outperforming comparable deep learning architectures. Guided Grad-CAM visualizations highlighted activations over QRS and R-peak regions, validating the physiological interpretability and diagnostic potential.