A novel multimodal self-supervised framework for ECG arrhythmia classification.

The electrocardiogram (ECG) has emerged as a primary tool in clinical practice for identifying cardiovascular diseases, owing to its low cost, simplicity, and non-invasiveness. Given the high cost associated with acquiring a substantial amount of ECG signals that require annotation by medical professionals, advanced self-supervised learning (SSL) techniques can effectively leverage abundant unlabeled data for learning, mitigating the performance impact of insufficient ECG classification labels. Contrastive learning has been successful as a self-supervised pre-training approach in image and time series domains. Inspired by this success, a novel pre-training technique, i.e., a simple multimodal self-supervised framework for ECG arrhythmia classification, is proposed in this paper by utilizing multi-modal data from ECG signals to enhance model initialization. Compared to other modalities, the expectation is that representations based on time and frequency for the same example should be brought as close together as possible. The pre-training is achieved through self-supervision by constructing time-domain contrastive learning loss and time-frequency loss, effectively learning features of ECG signals. The proposed method evaluates datasets containing both multi-lead and single-lead ECG data. Experimental results demonstrate that, by applying the pre-training method followed by fine-tuning for downstream tasks, the proposed algorithm outperforms standard contrastive learning paradigms on ACC and AUC, respectively, and even outperforms supervised learning.