Contrastive self-supervised pre-training on 1 million ECGs achieved 3%-4% higher AUROC for low LVEF and 5%-7% higher AUROC for high potassium chloride compared with baseline in low-label settings.
Does contrastive pre-training improve ECG classification performance for low LVEF and high serum potassium chloride compared to randomly initialized models?
Contrastive pre-training on a large corpus of unlabeled ECGs substantially enhances downstream classification performance for low LVEF and hyperkalemia, particularly when labeled data is scarce.
Effect estimate: AUROC +3-4% (LVEF), +5-7% (potassium)
Background: Self-supervised contrastive learning has emerged as a powerful paradigm for learning generalizable representations from unlabeled data. In the context of electrocardiogram (ECG) analysis, such pre-training can significantly enhance classification performance, especially when labeled data is scarce. Objective: We aimed to investigate and improve contrastive self-supervised learning techniques for ECGs by systematically combining recent methodological advances in augmentation design, contrastive loss formulation, and encoder architectures. Methods: We implemented a contrastive pre-training framework combining vectorcardiography (VCG)-based physiologically-inspired augmentations, interlead, intersegment, contrastive loss, and patient-aware positive sampling. In addition, we developed a dual-stream architecture, extending the TemporalNet model by processing grouped ECG leads independently. Pretraining was conducted on a large corpus of approximately 1 million unlabeled ECGs. We evaluated performance on 2 downstream classification tasks-low left ventricular ejection fraction (LVEF) and high serum potassium chloride-using various levels of labeled supervision (1%, 5%, 10%, 50%, and 100%). The pre-trained models were compared with the randomly initialized models under both frozen and finetuned conditions. Results: Contrastive pre-training consistently improved performance across all supervision levels. In low-label settings (1%-10% supervision), the pre-trained model achieved 3%-4% higher area under the receiver operator curve on the LVEF task and 5%-7% higher area under the receiver operator curve on the potassium chloride task compared with the baseline. The performance gap narrowed with increased supervision but remained favorable toward pre-trained models. Conclusion: Our findings demonstrate that contrastive pre-training can substantially enhance ECG classification, especially when labeled data is limited. By unifying and extending ideas from recent literature into a scalable framework trained on 1 million ECGs, we provide practical guidance and architectural innovations for building strong ECG foundation models applicable to a broad range of clinical prediction tasks.
Dade et al. (Thu,) conducted a other in Low left ventricular ejection fraction and high serum potassium chloride (n=1,000,000). Contrastive self-supervised pre-training vs. Randomly initialized models was evaluated on Area under the receiver operator curve for low LVEF and high serum potassium chloride (AUROC +3-4% (LVEF), +5-7% (potassium)). Contrastive self-supervised pre-training on 1 million ECGs achieved 3%-4% higher AUROC for low LVEF and 5%-7% higher AUROC for high potassium chloride compared with baseline in low-label settings.