SimAECLR: A Contrastive Learning Approach with Autoencoder for ECG Signals Classification Under Limited Labeled Data

Yantao Li, Zhenyu Yang, Zhongyin Wang, Mengying Xie, Shaojiang Deng

Abstract: Electrocardiogram (ECG) signals provide a direct reflection of cardiac electrical activity and are critical for the early detection and diagnosis of cardiovascular diseases. However, the scarcity of labeled data and the challenge of extracting discriminative features significantly hinder the performance of automated ECG analysis. To address these issues, this paper proposes SimAECLR, a self-supervised ECG classification framework that integrates a memory-augmented autoencoder with contrastive learning. SimAECLR automatically generates diverse views of unlabeled ECG signals through a memory-enhanced autoencoder and employs the InfoNCE loss within a dilated convolutional feature extractor to improve intra-class compactness and inter-class separability in the representation space. The model is first pretrained on unlabeled data to learn robust feature representations, followed by supervised fine-tuning using only a small amount of labeled data. This enables accurate multi-class ECG classification under limited supervision, supporting the development of collaborative intelligence systems in real-world clinical settings. Experimental results demonstrate the effectiveness of SimAECLR, achieving 91.26% and 85.26% accuracy on the MIT-BIH dataset, and 94.80% and 91.70% accuracy on the ECG5000 dataset using only 10% and 1% labeled data, respectively. These results highlight the model’s strong classification performance and its ability to significantly reduce reliance on annotated data.

External IDs:doi:10.1007/978-3-032-21168-2_18