Explainable paroxysmal atrial fibrillation diagnosis using an artificial intelligence-enabled electrocardiogram.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Diagnostics

Background:

• Atrial fibrillation (AF) is a major cause of morbidity and mortality.
• Paroxysmal atrial fibrillation (PAF) often presents silently, particularly in cryptogenic stroke patients.
• Early detection of PAF is crucial for stroke prevention.

Purpose of the Study:

• Develop reliable AI algorithms for early AF detection in normal sinus rhythm (NSR) patients.
• Utilize 12-lead electrocardiograms (ECGs) for AF prediction.
• Identify key ECG features indicative of future AF onset.

Main Methods:

• Trained deep neural networks on a large dataset of 552,372 ECG traces.
• Evaluated model performance using the area under the receiver operating characteristic curve (AUROC).
• Employed explainable AI to understand model decision-making processes.

Main Results:

• Achieved an AUROC of 0.905 ± 0.007 for early PAF diagnosis.
• Identified T-wave vicinity (ST segment, S-peak) as critical for PAF prediction.
• Found nonspecific ST-T abnormalities and inverted T waves associated with PAF in NSR.

Conclusions:

• Deep learning models can predict AF onset from NSR ECGs.
• AI models identify crucial ECG features for AF diagnosis.
• This approach can serve as a predictive tool for PAF screening and stroke risk assessment.