| 摘 要: 心电图(Electrocardiogram, ECG)的时域与频域信息具有显著的互补性，多模态融合是提升多标签ECG诊断性能的有效途径。针对时频特征分布差异大，多模态特征融合不充分的问题，本文提出一种多模态多任务联合学习网络(Multimodal and Multitask Joint Learning Network, M2JLNet)用于多标签ECG分类。首先，在多模态特征提取阶段，引入通道–空间交叉注意力(Channel-Spatial Cross-Attention, CSCA)，以增强模型对关键特征的建模能力；其次，构建监督对比学习任务，在特征空间中对共享标签样本进行表示对齐，从而增强时频特征一致性并提升类别可分性。实验结果表明，所提方法在CPSC2018数据集上的F1分数达到84.88%，并在PTB-XL数据集的超类任务上取得75.27%的F1分数，综合性能优于现有主流方法。 |
| 关键词: 心电图分类 多模态 多任务学习 多标签 监督对比学习 注意力机制 |
|
中图分类号:
文献标识码:
|
| 基金项目: 国家自然科学基金项目（面上项目，重点项目，重大项目） |
|
| Joint Supervised Contrastive Learning for Multi-Modal Multi-Label ECG Classification |
|
Hu Xinping, Sun Zhanquan, Qian Chenglong
|
School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology
|
| Abstract: The time-domain and frequency-domain information of electrocardiogram (ECG) signals exhibit significant complementarity, and multimodal fusion is an effective approach to improving multi-label ECG diagnosis performance. However, large dis-tribution discrepancies between time-frequency features and insufficient multimodal integration may limit representation quality and diagnostic accuracy. To address these challenges, we propose a Multimodal and Multitask Joint Learning Network (M2JLNet) for multi-label ECG classification. First, in the multimodal feature ex-traction stage, a Channel-Spatial Cross-Attention (CSCA) mechanism is introduced to enhance the modeling capability of critical features. Second, a supervised contras-tive learning task is constructed to align time-frequency representations in the feature space, thereby improving cross-modal consistency and enhancing class separability. Experimental results demonstrate that the proposed method achieves an F1 score of 84.88% on the CPSC2018 dataset and 75.27% on the superclass task of the PTB-XL dataset, outperforming existing state-of-the-art methods in terms of overall performance. |
| Keywords: ecg classification multimodal multitask learning multi-label supervised contras-tive learning attention mechanism |
