Lee Myeonghun, Lim Jiwoo, Kim JinKook
HUINNO Co., Ltd., Seoul, Republic of Korea.
School of Systems Biomedical Science, Soongsil University, Seoul, Republic of Korea.
Heart Rhythm O2. 2025 May 19;6(8):1199-1211. doi: 10.1016/j.hroo.2025.05.012. eCollection 2025 Aug.
Deep learning has significantly improved medical diagnostics, particularly in electrocardiogram (ECG) analysis, yet accurate classification of arrhythmias remains challenging.
We propose Electrocardiogram Graph Convolutional Network (ECG-GraphNet), a graph convolutional network designed to classify arrhythmias into 3 types: normal (N), supraventricular ectopic (S), and ventricular ectopic (V) beats.
ECG-GraphNet utilizes a novel graph representation of ECG data in which the P wave, QRS complex, and T wave are modeled as individual nodes. A unique QRS-centered weighted average pooling method is employed to enhance beat-specific feature extraction. We systematically explored various aspects including node features, edge definitions, a data augmentation method, and architecture configuration to determine the optimal model design. Experiments were conducted on 10-second ECG recordings from 328 patients using a single-lead device.
The optimized ECG-GraphNet achieved a Macro F1 score of 88.61% in 5-fold cross-validation. Scalability experiments further demonstrated its robustness, with Macro F1 scores of 85.21% and 87.03% across diverse ECG patterns and sizes.
Our novel approach and comprehensive analysis underscore the potential advantages of ECG-GraphNet in clinical diagnosis and monitoring.
深度学习显著改善了医学诊断，尤其是在心电图（ECG）分析方面，然而心律失常的准确分类仍然具有挑战性。
我们提出了心电图图卷积网络（ECG-GraphNet），这是一种图卷积网络，旨在将心律失常分为三种类型：正常（N）、室上性异位（S）和室性异位（V）搏动。
ECG-GraphNet利用一种新颖的心电图数据图表示，其中P波、QRS波群和T波被建模为单独的节点。采用独特的以QRS为中心的加权平均池化方法来增强特定搏动的特征提取。我们系统地探索了各个方面，包括节点特征、边的定义、数据增强方法和架构配置，以确定最佳模型设计。使用单导联设备对328例患者的10秒心电图记录进行了实验。
优化后的ECG-GraphNet在5折交叉验证中实现了88.61%的宏F1分数。扩展性实验进一步证明了其稳健性，在不同的心电图模式和大小下，宏F1分数分别为85.21%和87.03%。
我们的新方法和全面分析强调了ECG-GraphNet在临床诊断和监测中的潜在优势。
