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

Area of Science:

• Biomedical Engineering
• Cardiology
• Machine Learning

Background:

• Electrocardiogram (ECG) data are crucial for detecting cardiac arrhythmias.
• Wearable devices enable real-time ECG monitoring, but labeled datasets are often scarce and costly.
• Developing effective ECG analysis requires methods that work with limited labeled data.

Purpose of the Study:

• To develop a pretraining framework for ECG abnormality detection using minimal labeled data.
• To create a morphology-aware embedding space for ECG signals.
• To improve the detection of Atrial Fibrillation, Sinus Bradycardia, and Sinus Tachycardia.

Main Methods:

• Utilized a supervised contrastive pretraining framework.
• Focused on generating a morphology-aware embedding space.
• Applied the framework to detect three specific ECG abnormalities.

Main Results:

• Achieved a macro Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.96.
• Obtained a balanced accuracy of 0.91.
• Outperformed a fully supervised alternative (macro AUROC 0.89, balanced accuracy 0.86).

Conclusions:

• Supervised contrastive pretraining is effective for ECG abnormality detection with limited labeled data.
• The proposed method enhances the performance of real-time cardiac arrhythmia monitoring.
• This approach addresses the challenges of creating augmented biomedical waveforms while preserving physiological features.