ResNet-TCN: A Joint Model for ECG Heartbeat Classification with High Accuracy

Shaogang Hu, Renwei Li, Qiang Hu, Guanchao Qiao

Abstract: The electrocardiogram (ECG) is a reliable indicator of heart health and is widely used to diagnose arrhythmias. In this work, we propose ResNet-TCN, a joint model based on residual network (ResNet) and temporal convolutional network (TCN) for more accurate ECG classification. Specifically, the ResNet model extracts the spatial features of the signal, the TCN model extracts the temporal features of the signal, and the linear layer combines the features for classification. On the MIT-BIH public dataset, the proposed ResNet-TCN reaches a high accuracy (Acc) of 98.84%, a positive predictive value (P <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">+</sup> ) of 93.34%, and a sensitivity (Se) of 91.69%. Ablation studies show that compared with single ResNet or TCN, the Acc and P <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">+</sup> of ResNet-TCN improved by 0.72% and 5.13%, 0.48% and 5.59%, respectively, and the Se is comparable. Moreover, compared with previous works, the proposed model achieves a 1%~4.8% accuracy improvement. The proposed model is pretty suitable for clinical ECG detection that requires high accuracy.