Self-supervised learning for Electrocardiogram classification using Lead Correlation and Decorrelation

Abstract

In recent years, the development of deep learning has shown potential in the automatic analysis of electrocardiogram (ECG), aiding cardiologists in detecting cardiovascular diseases (CVDs). Generally, deep learning models depend on numerous labeled ECGs to train, but manual labeling of ECGs is costly as it requires considerable time and expertise. Self-supervised learning (SSL) can solve this problem by pretraining deep learning models with unlabeled ECGs, mitigating their reliance on labeled ECGs. This work proposes lead correlation and decorrelation (LCD) for effective and efficient SSL of ECGs. Concretely, LCD combines intra-lead correlation, inter-lead correlation, intra-lead and inter-lead decorrelation in pretraining. These mechanisms utilize multilead ECG characteristics: intra-lead invariance, inter-lead invariance, inter-lead variance, and intra-lead redundancy. After pretraining, LCD can provide a generic encoder for feature extraction of any ECG lead in a classification task. Benefitting from the effective pretraining mechanism, models with the encoders pretrained by LCD outperform most of the baselines. Compared with the best baseline, they achieve better/comparable classification performances in the same tasks with less pretraining time. Furthermore, LCD helps the models focus on critical features when training with insufficient labeled ECGs, reducing the reliance on labeled ECGs by 4$\sim$6$\times$. All the results demonstrate that LCD is an effective and efficient method, boosting a broader application of deep learning to automatic ECG analysis. The code is available at https://github.com/Aiwiscal/ECG_SSL_LCD.