A hybrid learning framework for automated multiclass electrocardiogram classification with SimCardioNet.

Area of Science:

• Artificial Intelligence
• Cardiology
• Medical Imaging

Background:

• Electrocardiography (ECG) is crucial for diagnosing cardiovascular diseases.
• Accurate ECG interpretation requires expertise and is hindered by data scarcity and high annotation costs.
• Deep learning offers potential for automated ECG analysis but requires substantial labeled data.

Purpose of the Study:

• To develop and evaluate SimCardioNet, a hybrid deep learning framework for multi-class ECG image classification.
• To address challenges of data scarcity and annotation costs in ECG analysis.
• To improve the accuracy and interpretability of automated ECG classification.

Main Methods:

• A hybrid self-supervised and supervised deep learning framework (SimCardioNet) was proposed.
• The framework utilizes a custom multi-scale CNN with residual connections and self-attention.
• Pretraining involved a modified SimCLR strategy with a hybrid loss, followed by supervised fine-tuning with progressive layer unfreezing.

Main Results:

• SimCardioNet achieved high performance on three distinct ECG datasets, including 0.975 accuracy on a clinical dataset and perfect classification on an external dataset.
• The model attained 0.921 accuracy and F1-score on the PTB-XL benchmark, outperforming existing state-of-the-art methods.
• Ablation studies confirmed the effectiveness of self-supervised pretraining, attention mechanisms, and data augmentations.

Conclusions:

• SimCardioNet demonstrates robust and interpretable ECG classification capabilities, reducing reliance on labeled data.
• The framework shows strong generalization and clinical viability, particularly in resource-constrained settings.
• This approach has the potential to significantly aid cardiovascular disease diagnosis through automated ECG analysis.