Lead-fusion Barlow twins: A fused self-supervised learning method for multi-lead electrocardiograms

Abstract

Nowadays, deep learning depends on large-scale labeled datasets, which limits its broader application in electrocardiogram (ECG) analysis, as manual labeling of ECGs is consistently costly. To overcome this issue, this paper proposes a fused self-supervised learning (SSL) method for multi-lead ECGs: lead-fusion Barlow twins (LFBT). It utilizes unlabeled ECG datasets to pretrain an encoder group using a fused loss. This loss fuses intra-lead and inter-lead BT loss. By employing BT, LFBT avoids the need for additional techniques to prevent trivial solutions (collapse) in pretraining. Moreover, multi-branch concatenation (MBC) fuses information from all leads when transferring pretrained encoders to downstream tasks. According to the experiments, LFBT can extract prior knowledge from unlabeled ECG datasets, making a deep learning model yield comparable performances with its supervised counterpart (trained from scratch) using 3$\sim 5\times$ fewer labels. Furthermore, LFBT is robust when applied to uncurated ECGs from real-world hospitals, with no significant performance decline observed after pretraining. Model interpretation based on gradient-weighted class activation mapping (GradCAM) indicates that LFBT helps models focus on critical waveform changes when training data and labels are insufficient. Compared with previous methods, LFBT demonstrates advantages in performance and implementation. To summarize, LFBT shows considerable potential in reducing the need for manual labeling of ECGs, thereby advancing deep learning applications in real-world ECG-based diagnoses. Code is available at https://github.com/Aiwiscal/ECG_SSL_LFBT.