Transforming Electrocardiogram to Participant-Level Diagnosis in Atrial Fibrillation Screening

Area of Science:

• Cardiology
• Biomedical Engineering
• Machine Learning in Healthcare

Background:

• Single-lead electrocardiograms (ECGs) are increasingly captured outside clinical settings, enabling widespread atrial fibrillation (AF) screening.
• Challenges include low AF prevalence and numerous ECGs per individual, complicating accurate participant diagnosis.
• Effective AF detection models are crucial for early intervention and stroke prevention.

Purpose of the Study:

• To introduce and evaluate a dynamic training approach for AF detection models using single-lead ECGs.
• To compare the performance of the dynamic approach against a fixed training strategy.
• To assess the efficacy of both approaches for both ECG-level classification and participant-level diagnosis.

Main Methods:

• An RR-interval-based AF detection model was trained using the Screening for Atrial Fibrillation with ECG to Reduce Stroke (SAFER) dataset.
• A dynamic training approach was implemented, incorporating more non-AF ECGs during model training compared to a fixed approach.
• Participant-level testing was performed by applying the trained ECG classification models.

Main Results:

• The dynamic training strategy consistently outperformed the fixed strategy across all metrics.
• Dynamic approach achieved higher Area Under the Receiver Operating Characteristic Curve (AUROC) scores (0.92-0.94 vs. 0.90-0.92).
• Dynamic approach yielded superior Area Under the Precision-Recall Curve (AUPRC) scores (0.38-0.75 vs. 0.12-0.66) for both ECG and participant levels.

Conclusions:

• The dynamic training approach significantly enhances the accuracy of AF detection from single-lead ECGs.
• This method demonstrates robust performance at the participant-level diagnosis, vital for clinical utility.
• Findings support the dynamic approach as a key advancement for real-world AF screening and management.