The morphology-rhythm contrast (MRC) framework yielded better ECG classification performance than supervised models using only 10% of the training data.
The MRC contrastive learning framework for multilead ECGs achieves high performance in downstream classification tasks using only 10% of training data, potentially reducing the manual labeling burden for cardiologists.
This article proposes a novel contrastive learning framework to learn high-quality representations for multilead electrocardiograms (ECGs). It is termed morphology-rhythm contrast (MRC) since it jointly considers the morphology and rhythm characteristics of multilead ECGs. Unlike existing studies only concentrating on ECG-specific data augmentations, MRC provides a systematic solution for ECG-based contrastive learning. It proposes two new ECG-oriented data augmentation methods termed random beat selection and 0–1 pulse generation for view creation, representing the morphology and rhythm characteristics of an ECG. Then, a triple-branch network maps the three views (raw ECG, morphology, and rhythm view) to a latent space for dual contrastive learning (raw ECG versus morphology view and raw ECG versus rhythm view). This dual contrastive learning can be adjusted to prefer invariance derived from ECG morphology and rhythm, making pretrained encoders suitable for different downstream tasks. Thus, MRC reduces the gap between pretraining and downstream tasks, which is a significant challenge in contrastive learning. More importantly, with only 10% of the training data, MRC-based classification models can yield better performances than the supervised models. Such a finding demonstrates that MRC can reduce the cardiologists’ labeling burden in real-world applications. Additionally, MRC achieves high performances in downstream tasks, outperforming existing studies under the same settings. To summarize, MRC is an effective contrastive learning framework for multilead ECGs. It has the potential to alleviate cardiologists’ workload by aiding diagnosis and reducing manual labels in real-world applications.
Liu et al. (Mon,) conducted a other in Multilead electrocardiogram (ECG) classification. Morphology-rhythm contrast (MRC) framework vs. Supervised models and existing studies was evaluated on Classification model performance. The morphology-rhythm contrast (MRC) framework yielded better ECG classification performance than supervised models using only 10% of the training data.