Forecasting imminent atrial fibrillation in long-term electrocardiogram recordings

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Acute atrial fibrillation (AF) poses significant morbidity, yet lacks predictive models for imminent onset.
• Current inpatient care lacks tools to forecast the immediate occurrence of AF.
• Forecasting AF onset is crucial for timely intervention and improved patient outcomes.

Purpose of the Study:

• To evaluate deep learning's capability in forecasting imminent atrial fibrillation (AF) onset.
• To achieve a clinically relevant lead time for AF prediction using ECG data.
• To explore the potential of neural networks in managing inpatient AF.

Main Methods:

• Utilized the Physiobank Long-Term AF Database with 24-h labeled ECG recordings.
• Developed three deep learning models (convolutional and transformer layers) for AF forecasting.
• Evaluated models using AUC(t), precision-recall curves, and risk trajectories at various lead times (7.5-60 min).

Main Results:

• All models demonstrated risk trajectory divergence approximately 15 minutes before AF onset.
• The sinus rhythm model achieved the highest AUC (0.74) at a 7.5-min lead time.
• A model using all preceding waveform data showed comparable performance with higher AUCs at longer lead times.

Conclusions:

• Demonstrated the potential of neural networks to forecast AF onset in long-term ECG recordings with clinically relevant lead times.
• This proof-of-concept study highlights the utility of deep learning for AF prediction.
• External validation in larger cohorts is necessary for clinical deployment.