Self-Supervised Learning with Electrocardiogram Delineation for Arrhythmia Detection.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Electrocardiogram (ECG) signals are crucial for diagnosing heart conditions like arrhythmia and heart failure.
• Deep learning models show promise in ECG analysis but require large, high-quality annotated datasets, which are costly and time-consuming to acquire.
• Self-supervised learning (SSL) offers a solution by enabling model pre-training on unlabeled data, reducing reliance on manual annotations.

Purpose of the Study:

• To propose and evaluate a novel SSL algorithm for ECG classification, specifically targeting arrhythmia detection.
• To demonstrate the effectiveness of SSL in enhancing deep neural network (DNN) performance for ECG analysis.
• To assess the algorithm's ability to improve classification with limited labeled data and facilitate feature transfer between different ECG datasets.

Main Methods:

• Developed a self-supervised learning (SSL) algorithm leveraging ECG delineation for pre-training deep neural networks (DNNs).
• Evaluated the proposed SSL algorithm on arrhythmia classification tasks using multiple ECG datasets.
• Compared the performance of DNNs pre-trained with the SSL method against traditional supervised learning approaches, varying the amount of labeled data.

Main Results:

• The proposed SSL algorithm significantly enhanced DNN performance in arrhythmia classification across various datasets.
• The method demonstrated improved accuracy even with limited fractions of labeled data, highlighting its efficiency.
• Features learned through pre-training on one dataset were successfully transferred and fine-tuned on a different dataset, showing generalization capabilities.

Conclusions:

• Self-supervised learning, particularly the proposed ECG delineation-based approach, is a viable and effective strategy for ECG classification.
• This method reduces the dependency on extensive manual annotations, making advanced ECG analysis more accessible.
• The demonstrated feature transferability suggests potential for developing robust, generalizable models for diverse cardiac monitoring applications.