Comprehensive Summary
In this study, researchers utilized a variational autoencoder (VAE) to extract ECG features representative of various atrial fibrillation (AF) outcomes. Then, these features were clustered into five AF outcome subgroups with a tree-learning method (DDRTree). The VAE, trained on median beat ECGs from one of two cohorts used in this study, initially identified 256 latent features. Only latent features with information capacity above 0.1 were selected and input into the DDRTree (51 features). Individuals with diagnosed AF were evaluated by the DDRTree, and patients were clustered into five AF outcome phenogroups. Phenogroup 1 identified higher-risk AF; phenogroup 2 identified highest-risk AF with heart failure; phenogroups 3-5 identified paroxysmal AF of average, lower, or higher risk, respectively. Multivariate logistics regression models identified the top three features that contributed to each subgroup’s identification. Phenogroup characteristics included data such as age, left atrial diameter and structure, and whether the ECGs were in sinus rhythm or AF.
Outcomes and Implications
The traditional methods of AF classification follow a stage-based classification, yet current AHA subtypes fail to address the heterogeneity of AF. Additionally, past machine-learning clustering of AF has not included ECG data, an easy-to-obtain piece of clinical data which may better characterize AF phenogroups rather than simply identifying general cardiovascular risk factors. The clusters identified by DDRTree, as well as the specific features that characterize each cluster, can help clinicians identify higher-risk patients and make more appropriate and individualized treatment decisions. However, before any clinical applications of AI-ECG tree-based phenotyping can be employed, future investigations must evaluate areas where this study was limited. For example, future research must assess how successful medical therapies are in providing better outcomes for each phenogroup.