Short-term atrial fibrillation onset prediction using machine learning.

Area of Science:

• Cardiology
• Biomedical Engineering
• Data Science

Background:

• Atrial fibrillation (AF) poses significant health risks, and early detection is crucial for preventive strategies.
• Machine learning (ML) models integrated into wearable devices offer potential for real-time AF prediction.
• Current methods often lack the short-term predictive capability needed for timely intervention.

Purpose of the Study:

• To develop and evaluate ML models for predicting imminent paroxysmal AF episodes using Holter ECG recordings.
• To identify patients in sinus rhythm who will experience AF within hours.

Main Methods:

• A large database of 95,871 Holter ECG recordings was analyzed, identifying 1319 AF episodes.
• Deep learning (DL) models were trained on raw ECG data.
• Traditional ML models, including random forest and XGBoost, were trained using heart rate variability (HRV) parameters for comparison.

Main Results:

• Decision tree models utilizing HRV parameters demonstrated superior predictive performance.
• The XGBoost model achieved an area under the ROC curve of 0.919 for predicting AF episodes lasting over 5 minutes.
• High accuracy (84.5%), sensitivity (83.0%), and specificity (86.6%) were reported for the best-performing model.

Conclusions:

• HRV parameters are critical for short-term AF onset prediction, supporting preventive healthcare.
• Integrating these predictive models into mHealth technologies could enable a 'pill-in-the-pocket' approach for AF management.
• Further prospective studies are needed to validate these findings and their clinical utility.