New Graph Network Boosts ECG Recognition with Domain Knowledge

A team of researchers (Ma, Zhang, Yuan, et al.) has introduced a novel graph convolution network that leverages domain knowledge for ECG recognition. Instead of relying solely on end-to-end convolutional neural networks, their approach integrates key anatomical landmarks from the PRQST complex—the P wave, QRS complex, and T wave—as explicit domain knowledge. The architecture employs a double-stream directed graph: spatial directed graphs capture positional relationships among key points within a single heartbeat, while temporal directed graphs model dependencies between adjacent cycles in extended ECG sequences. This structure enables the network to understand both the spatial layout of cardiac electrical activity and its evolution over time.

Experimental results on the First Chinese ECG Intelligent Competition dataset, which classifies ECGs into nine categories, demonstrate the model's effectiveness. It achieved an overall average F1 score of 88.1%, with a particularly notable 76.3% F1 score for rare categories—both outperforming current state-of-the-art models. The inclusion of domain knowledge significantly boosted detection performance for infrequent arrhythmias and abnormalities. Presented at ICONIP 2024 and published in Springer's LNCS, this work highlights how embedding medical expertise into deep learning architectures can yield more accurate, interpretable, and clinically useful ECG analysis tools.