A new study presented at ICONIP 2024 in Auckland, New Zealand, introduces a domain knowledge-based temporal-spatial graph convolution network for ECG recognition. This innovative approach addresses challenges in AI interpretability within healthcare, specifically targeting the accurate classification of electrocardiogram data.
Innovative Approaches to ECG Recognition
In recent years, advancements in artificial intelligence have transformed various fields, including healthcare. However, challenges in interpretability remain, particularly in specialized areas such as ECG recognition. The study, led by Wenting Ma and co-authored by eight others, emphasizes a novel method that integrates key landmarks from ECG data as domain knowledge, thus enhancing the model's effectiveness.
Unlike traditional end-to-end convolutional neural networks, this research proposes a double-stream directed graph model that captures both intra and inter ECG cycles. By employing spatial directed graphs to represent positional relationships among key points and temporal directed graphs to illustrate dependencies between adjacent cycles, the model demonstrates improved classification accuracy.
Experimental Results and Performance Metrics
The researchers conducted experiments using the First Chinese ECG Intelligent Competition dataset, which categorizes ECG readings into nine distinct classes. The results revealed that the proposed model achieved an impressive overall average F1 score of 88.1%. Notably, the average F1 score for rare categories reached 76.3%, significantly outperforming existing state-of-the-art models.
