Deriving novel atrial fibrillation phenotypes using a tree-based artificial intelligence-enhanced electrocardiography...

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Traditional atrial fibrillation (AF) classification by duration has limitations.
• Mechanistic and prognostic diversity within AF subtypes remains poorly understood.

Purpose of the Study:

• To develop an AI-driven framework for mapping AF heterogeneity.
• To identify distinct AF phenogroups with varying disease risks and outcomes.
• To augment traditional AF classification with a risk-stratified dimension.

Main Methods:

• Utilized a variational autoencoder trained on over 1.1 million ECGs to extract features from 20,291 AF patients.
• Applied unsupervised tree-based clustering to these features to create a phenogroup structure.
• Analyzed phenogroup characteristics for disease risk stratification and clinical correlation.

Main Results:

• Identified five distinct AF phenogroups, stratified by future disease risk.
• Phenogroup 2 represented highest-risk AF with heart failure (HF), indicating advanced disease and mortality risk.
• Paroxysmal AF phenogroups (4 and 5) showed differences in risk and ventricular structure, with phenogroup 5 having more adverse features.
• The AI-ECG framework demonstrated explainability through tree trajectories mapping individual patient traits.

Conclusions:

• An AI-ECG framework can effectively map AF heterogeneity beyond traditional duration-based subtypes.
• The identified phenogroups provide a novel risk-based stratification for atrial fibrillation patients.
• This approach supports personalized medicine by offering deeper insights into AF prognosis and management.