面向不平衡心电信号分类的类别感知混合Mixup方法-《中国医学物理学杂志》

文章信息/Info

Title:: Class-aware hybrid Mixup method for imbalanced electrocardiogram signal classification

作者:: 田悦芃1; 2; 徐琳2; 李钰雯1; 1.东南大学仪器科学与工程学院，江苏南京 210096； 2.信息系统工程全国重点实验室，江苏南京 210096

Author(s):: TIAN Yuepeng1; 2; XU Lin2; LI Yuwen1; 1. School of Instrument Science and Engineering, Southeast University, Nanjing 210096, China 2. National Key Laboratory of Information Systems Engineering, Nanjing 210096, China

Keywords:: Keywords: electrocardiogram signal classification class imbalance data augmentation Mixup CutMix1d Manifold Mixup

摘要:: 针对心电信号分类中的类别不平衡问题，提出一种面向心电分类的类别敏感融合Mixup方法。该方法结合CutMix1d与Manifold Mixup的优势：前者在原始空间生成更贴近真实分布的少数类样本，后者通过特征空间插值平滑决策边界，从而兼顾少数类增强与整体判别能力提升。同时，引入少数类感知机制，使模型在稀有类别上的优化更加针对性。在两个公开心电数据集上开展实验，包括PTB-XL的3个不平衡任务以及CPSC2018任务。结果表明，本文方法在4个任务上的平均Macro-AUC较基线整体提升接近0.01，表现出稳定一致的优势；在Sub-diagnostic任务上进一步引入Macro-F1评估时，平均值较基线提升0.075，类别层面的改进更为显著。上述结果表明，该方法能够有效缓解心电数据的不平衡问题，为临床心电自动分析提供一种高效的数据增强方案。

Abstract:: Abstract: A class-aware hybrid Mixup method customized for electrocardiogram (ECG) analysis is proposed to alleviate class imbalance in ECG signal classification. The proposed method leverages the strengths of CutMix1d and Manifold Mixup: the former generates minority-class samples in the original space that better preserve realistic signal structures, while the latter smoothes decision boundaries through feature-space interpolation, thereby simultaneously enhancing minority-class representation and overall discriminative capability. In addition, a minority-class-aware mechanism is introduced to enable more targeted optimization for rare classes. Experiments are conducted on two public ECG datasets, including 3 imbalanced tasks from PTB-XL and a CPSC2018 task. Results show that the proposed method achieves an average Macro-AUC improvement of nearly 0.01 over the baseline across 4 tasks, demonstrating consistent and stable gains. On the Sub-diagnostic task, the Macro-F1 score improves by 0.075 compared with the baseline, indicating more pronounced improvements at the class level. These findings confirm that the proposed approach can effectively alleviate class imbalance in ECG data, providing an efficient data augmentation solution for clinical ECG analysis.