ECG Abnormality Detection Using MIMIC-IV-ECG Data Via Supervised Contrastive Learning.

Electrocardiogram data provide a tremendous opportunity for the detection of various types of cardiac arrhythmia. Recent advancement in ubiquitous wearable devices with incorporated ECG sensors offers an opportunity for a real-time monitoring system for detecting abnormalities in ECG data. Nevertheless, access to a comprehensive labeled dataset can be expensive and not feasible. As a result, there is a need for a pretraining framework with a minimal amount of labeled data that generates a morphology-aware embedding space. Contrastive learning is a commonly used approach for creating these regularized embedding spaces where the latent representations of similar data points are close. However, creating an augmented biomedical waveform can be challenging as it may alter the underlying physiological features. To address these shortcomings, we utilize a supervised contrastive pretraining framework for detecting three types of ECG abnormalities, Atrial Fibrillation, Sinus Bradycardia, and Sinus Tachycardia. This proposed method achieves an increased macro AUROC of 0.96 (versus 0.89 for the fully supervised alternative) and a balanced accuracy of 0.91 (versus 0.86 for the baseline).