Identifying New Atrial Fibrillation Phenotypes via Tree-Based AI-Enhanced ECG
Overview
This study utilized a variational autoencoder and a tree-based clustering method on over 1.1 million ECGs to identify five distinct atrial fibrillation (AF) phenogroups with varying risks and clinical characteristics. The AI-ECG approach revealed phenogroups stratified by future disease risk, including subtypes with differing ventricular structure and heart failure burden, enhancing AF classification beyond traditional duration-based categories.
Background
Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia, associated with increased morbidity and mortality, including a five-fold increased risk of ischemic stroke and elevated heart failure risk. Traditional AF classification relies on arrhythmia duration but fails to capture the heterogeneity and prognostic diversity of AF. Machine learning approaches have been applied to identify AF subgroups, yet prior methods often omit electrophysiological data from ECGs. Incorporating ECG signals into unsupervised AI models may better characterize AF phenotypes and improve risk stratification.
Data Highlights
Parameter
Value
Total ECGs analyzed
1,163,401
Unique patients with AF
20,291
ECGs recorded during AF
5,621
ECGs recorded during sinus rhythm
14,670
Additional ECGs from AF patients
338,242
Number of phenogroups identified
5
Latent features from VAE
51
Key Findings
Five AF phenogroups were identified using a tree-based clustering of AI-extracted ECG features, stratified by future disease risk.
Phenogroups included higher-risk AF, highest-risk AF with heart failure, average paroxysmal AF, lower-risk paroxysmal AF, and higher-risk paroxysmal AF.
Paroxysmal phenogroups 4 and 5 differed in risk and ventricular structure, with phenogroup 5 showing more adverse cardiac features.
The mixed AF phenogroup 2 represented advanced AF with greater heart failure burden and mortality risk.
The tree-based method provided an explainable trajectory positioning individuals based on shared electrophysiological traits.
Clinical Implications
This AI-ECG framework enhances AF classification by integrating electrophysiological data to identify phenogroups with distinct risk profiles, potentially guiding personalized management strategies. Recognizing phenogroups with higher heart failure burden or adverse ventricular remodeling may inform targeted interventions and improve prognostication beyond traditional AF subtypes.
Conclusion
The study demonstrates that AI-enhanced ECG analysis combined with tree-based clustering can uncover novel AF phenotypes with prognostic significance, offering a refined, risk-based dimension to AF classification that supports personalized clinical care.
References
AHA 2023 Guidelines -- Refinement of AF Classification
Beth Israel Deaconess Medical Centre Dataset -- ECG Collection
DDRTree Method -- Dimensionality Reduction and Clustering
Variational Autoencoder in ECG Analysis -- Feature Extraction
by Mehak Gurnani, Konstantinos Patlatzoglou, Joseph Barker, Libor Pastika, Boroumand Zeidaabadi, Ibrahim Antoun, Riyaz Somani, G. Andre Ng, Paolo Inglese, Lara Curran, Declan O’Regan, Nicholas S. Peters, Daniel B. Kramer, Jonathan W. Waks, Arunashis Sau, Fu Siong Ng