Prediction of Atrial Fibrillation From the ECG in the Community Using Deep Learning: A Multinational Study.

Area of Science:

• Cardiology
• Artificial Intelligence
• Public Health

Background:

• Atrial fibrillation (AF) poses a significant health risk.
• Existing risk scores for AF have limitations.
• Diverse datasets are crucial for validating predictive models.

Purpose of the Study:

• To refine and validate a deep neural network model (ECG-AF) for predicting AF risk using ECG data.
• To compare the ECG-AF model's performance against the established CHARGE-AF risk score.
• To evaluate the association of the ECG-AF model with other cardiovascular outcomes.

Main Methods:

• The ECG-AF model was developed using 60% of Framingham Heart Study (FHS) samples.
• Model performance was assessed using Area Under the Curve (AUC) in FHS, UK Biobank, and ELSA-Brasil cohorts.
• Cox proportional hazards models were used to evaluate associations with cardiovascular outcomes.

Main Results:

• The ECG-AF model demonstrated moderate discrimination for incident AF (AUC, 0.82) in FHS, comparable to CHARGE-AF (AUC, 0.83).
• Combining ECG-AF and CHARGE-AF improved AF prediction (AUC, 0.85) in FHS and other cohorts.
• Higher ECG-AF scores correlated with increased risks of heart failure, myocardial infarction, stroke, and mortality.

Conclusions:

• The single-input ECG-AF deep neural network model shows strong performance in predicting AF and cardiovascular outcomes.
• The model is comparable to multivariable clinical risk scores and offers improved prediction when combined.
• This AI-driven approach holds promise for enhancing cardiovascular risk assessment in diverse populations.