Unsupervised Phenotyping of Atrial Fibrillation Using Variational Autoencoders in a Critical Care Cohort: The...

Atrial fibrillation (AF) is a heterogeneous disorder associated with substantial morbidity and mortality, yet management remains largely guided by traditional classifications and risk scores. We aimed to identify distinct AF phenotypes using variational autoencoder (VAE)-based unsupervised learning and to evaluate differences in mortality, healthcare utilization, and treatment patterns across them.

Retrospective observational cohort study in MIMIC-IV v3.1, a single-center critical care database, restricted to the 2014 to 2016, 2017 to 2019, and 2020 to 2022 anchor year groups. In 13,967 patients with AF, a patient-level matrix of 35 demographic, comorbidity, medication, laboratory, AF characteristic, and utilization features was used to train a VAE for nonlinear dimensionality reduction, followed by K-means clustering at k = 6. Phenotypes were compared by Kaplan-Meier and Cox analysis of all-cause mortality, with Phenotype 4 (High AF Chronicity) as the reference because it had the longest reverse Kaplan-Meier follow-up, and by incremental discrimination beyond routinely available variables. Reproducibility was assessed by two-level bootstrap resampling against a permuted-data null.

Six phenotypes were identified (n = 1,906 to 2,925): Phenotype 1 (Cerebrovascular-Predominant), Phenotype 2 (Hypoalbuminemic-Anemic), Phenotype 3 (Low Comorbidity Burden), Phenotype 4 (High AF Chronicity), Phenotype 5 (Warfarin-Predominant Anticoagulation), and Phenotype 6 (Renal Dysfunction). All-cause mortality differed across phenotypes (log-rank p<0.001), with 14 of 15 pairwise comparisons significant after Bonferroni correction. Observed-death proportions ranged from 11.1% to 48.1%, but accrued over markedly unequal observation windows (reverse Kaplan-Meier median follow-up 5.3 to 854.9 days) and are not comparable risks. Phenotype membership remained associated with mortality after adjustment for age and sex, the adjustment set that minimizes overlap with the phenotype-defining features. Adding phenotype to a model of age, sex, CHA₂DS₂-VASc, and three routine laboratory values raised the concordance index from 0.683 to 0.733, although a conventional model containing utilization variables and no phenotype term reached 0.745. Thirty-day readmission ranged from 13.0% to 42.1% and intensive care admission from 33.6% to 77.8%; because utilization variables also helped define the clusters, these are descriptive characteristics of the phenotypes rather than independent outcomes, and rebuilding the phenotypes without those variables halved the intensive care spread to 25.1 percentage points.

VAE-based unsupervised learning identified six AF phenotypes with substantial differences in mortality, healthcare utilization, and treatment patterns, highlighting clinically meaningful heterogeneity within AF. Cluster separation metrics were modest and individual phenotype assignment was only moderately reproducible when the model was refit end to end (Adjusted Rand Index 0.406), indicating that phenotypes lie along overlapping continua rather than representing discrete entities, although a between-phenotype mortality gradient was present in all 100 end-to-end resamples (minimum range 18.8 percentage points). Temporal validation did not meet the pre-specified ±5% threshold for phenotype distributions (maximum difference 9.0%). These findings are hypothesis-generating and support future prospective evaluation of phenotype-guided risk stratification, with the longer-term goal of informing precision cardiovascular medicine.