Abstract

Deep learning approaches have boosted automated electrocardiogram (ECG) abnormality diagnosis, yet they still suffer from limited labeled datasets, semantic discrepancies between multi-modal data, and class imbalance. To address these issues, this paper proposes a multi-label ECG classification algorithm based on self-supervised pretraining and multi-modal semantic
alignment. First, a unimodal contrastive enhancement network based on self-supervised learning is introduced, which leverages contrastive learning on an unlabeled dataset for self-supervised pretraining to enhance unimodal feature extraction. Then, in the semantic-guided multi-modal fusion module, label semantics are utilized to align cross-modal semantics, and a crossmodal
attention network is employed for feature fusion. Finally, in the multilabel classification module, a multi-label contrastive loss based on disease co-occurrence relationships is proposed to mitigate the class imbalance problem and improve the recognition accuracy of minority classes. Evaluated on the public 12-lead CPSC2018 benchmark, the proposed algorithm attains an F1 of 85.1%, outperforming baseline algorithms and setting a new state of the art for multi-label arrhythmia classification. The results confirm the effectiveness of self-supervised feature enhancement, semantic alignment, and class-aware contrastive learning in real-world ECG analysis.