Abstract：In electrocardiogram (ECG) signal classification, patients often have multiple coexisting diseases, making multi-label classification more challenging and clinically meaningful. In deep learning, large discrepancies in the number of samples often lead to label imbalance. The model tends to underperform on labels with fewer examples. Label correlation can help alleviate this issue. This paper compressed a Teacher Model into a Student Model. The loss function of the Student Model was based on both ground-truth labels and intermediate outputs from the Teacher Model. This paper proposed an approach to constructing a label correlation matrix. The label correlation matrix used cosine similarity to capture inter-label relationships. The correlation matrix was then used to calibrate the predicted probabilities. Hefei-Tianchi contains 18 232 raw ECG recordings,containing 38left bundle branch conduction block data. PTB_XL contains 21 836 raw ECG data,containing 51 left bundle branch conduction block raw ECG data. The new model helped to alleviate the problem of label imbalance. For left bundle branch block, F1 score on the Hefei-Tianchi dataset increased from 0.799 to 0.837. For pacing heart rhythm, it increased from 0.862 to 0.876. For First-degree atrioventricular block, it increased from 0.862 to 0.876, higher than ECGNet′s 0.535 and Acharya′s 0.810. For PTB_XL dataset, First-degree atrioventricular block, increased from 0.679 to 0.857. For Right ventricular hypertrophy, it increased from 0.698 to 0.745. For Complete left bundle branch block, it increased from 0.649 to 0.887. Meanwhile, on the Hefei-Tianchi dataset, the F1 score of the Student Model was 0.816, while that of the Teacher Model was 0.824. On the PTB_XL dataset, the F1 score of the Student Model was 0.876, while that of the Teacher Model was 0.873. The size of the Student Model was only 0.62 times of the Teacher Model. The established model addressed the long-tail distribution issue, improving predictive capability for few-shot labels, while simplifying parameters and saving computational resources.
[1] Berkaya SK, Uysal AK, Gunal ES, et al. A survey on ECG analysis[J]. Biomedical Signal Processing and Control, 2018, 43: 216-235. [2] Liu X, Wang H, Li Z, et al. Deep learning in ECG diagnosis: a review[J]. Knowledge-Based Systems, 2021, 227: 107187. [3] Mathews SM, Kambhamettu C, Barner KE. A novel application of deep learning for single-lead ECG classification[J]. Computers in Biology and Medicine, 2018, 99: 53-62. [4] Zhang ML, Li YK, Liu XY, et al. Binary relevance for multi-label learning: an overview[J]. Frontiers of Computer Science, 2018, 12: 191-202. [5] Hannun AY, Rajpurkar P, Haghpanahi M, et al. Cardiologist-level arrhythmia detection and classification in ambulatory electrocardiograms using a deep neural network [J]. Nature Medicine, 2019, 25(1): 65-69. [6] Picon A, Irusta U, Álvarez-Gila A, et al. Mixed convolutional and long short-term memory network for the detection of lethal ventricular arrhythmia[J]. PLoS ONE, 2019, 14(5): e0216756. [7] Sepahvand M, Abdali-Mohammadi F. A novel method for reducing arrhythmia classification from 12-lead ECG signals to single-lead ECG with minimal loss of accuracy through teacher-student knowledge distillation[J]. Information Sciences, 2022, 593: 64-77. [8] 黄震华, 杨顺志, 林威, 等. 知识蒸馏研究综述[J]. 计算机学报, 2022, 45(3): 624-653. [9] 邵仁荣, 刘宇昂, 张伟, 等. 深度学习中知识蒸馏研究综述[J]. 计算机学报, 2022, 45(8): 1638-1673. [10] Bogatinovski J, Todorovski L, Džeroski S, et al. Comprehensive comparative study of multi-label classification methods[J]. Expert Systems with Applications, 2022, 203: 117215. [11] Shuhong W, Runchuan L, Xu W, et al. Multiscale Residual Network Based on Channel Spatial Attention Mechanism for Multilabel ECG Classification.[J].Journal of Healthcare Engineering,2021,2021:6630643. [12] 黄浩, 朱俊江. 基于GoogLeNet-ViT模型的心律不齐多标签诊断算法[J]. 计算机应用与软件, 2025, 42(5): 247-254. [13] 陈凌志.融合导联感知与迁移机制的心电图多标签分类研究[D].桂林:广西师范大学,2025. [14] Sun Z, Wang C, Zhao Y, et al. Multi-label ECG signal classification based on ensemble classifier[J]. IEEE Access, 2020, 8: 117986-117996. [15] Zhang ML, Li YK, Liu XY, et al. Binary relevance for multi-label learning: an overview[J]. Frontiers of Computer Science, 2018, 12: 191-202. [16] Read J, Pfahringer B, Holmes G, et al. Classifier chains for multi-label classification[C]//Machine Learning and Knowledge Discovery in Databases: European Conference, ECML PKDD 2009. Bled: Springer Berlin Heidelberg, 2009: 254-269. [17] 阿里云. "合肥高新杯"心电人机智能大赛——心电异常事件预测[EB/OL]. (2019-10-8) https://tianchi.aliyun.com/competition/entrance/231754/introduction. [18] Wagner P, Strodthoff N, Bousseljot RD, et al. PTB_XL, a large publicly available electrocardiography dataset[J]. Scientific Data, 2020, 7(1): 154. [19] Cai J, Sun W, Guan J, et al. Multi-ECGNet for ECG arrythmia multi-label classification[J]. IEEE Access, 2020, 8: 110848-110858. [20] Acharya UR, Fujita H, Lih OS, et al. Automated detection of arrhythmias using different intervals of tachycardia ECG segments with convolutional neural network[J]. Information Sciences, 2017, 405: 81-90. [21] He R, Liu Y, Wang K, et al. Automatic cardiac arrhythmia classification using combination of deep residual network and bidirectional LSTM[J]. IEEE Access, 2019, 7: 102119. [22] Ge Z, Jiang X, Tong Z, et al. Multi-label correlation guided feature fusion network for abnormal ECG diagnosis[J]. Knowledge-Based Systems, 2021, 233: 107508. [23] 任烜,李思瀚,胡水清,等. HEART评分及其改良版本在急诊评估中的研究进展 [J]. 中国病案, 2025, 26 (10): 104-107. [24] 张闿艺. 冠脉CT危险分层及其对冠心病预后的影响[D]. 北京:中国人民解放军医学院, 2014.
