Identification of atrial fibrillation differential clinical risk phenotypes in the GLORIA-AF Registry: An analysis using Generative Topographic Mapping, a novel machine learning approach to clustering

Background

Atrial fibrillation (AF) is a clinically heterogeneous condition associated with substantial variation in cardiovascular outcomes. Conventional risk stratification approaches rely on predefined combinations of clinical variables and may not fully capture the multidimensional structure of AF populations. Generative topographic mapping (GTM) is a novel probabilistic machine learning (ML) method that preserves topological relationships in high-dimensional data, enabling data-driven identification of clinically meaningful phenotypes. We evaluated the clinical and prognostic relevance of such GTM-derived phenotypes in a large, prospective, real world global AF registry.

Purpose

To derive AF phenotypes using GTM, evaluate prognostic differences across phenotypes, and secondarily assess the associations between adherence to holistic or integrated care based on the Atrial fibrillation Better Care (ABC) pathway and clinical outcomes.

Methods

From the prospective GLORIA-AF Registry, GTM was applied to baseline continuous clinical measures. Missing baseline values were imputed using an iterative multivariable approach prior to GTM phenotyping. GTM node reference vectors were aggregated into macro-phenotypes using Ward’s minimum-variance hierarchical clustering (Euclidean distance). Associations between phenotypes and the risk of major outcomes were evaluated using Kaplan–Meier analysis and Cox proportional hazards models. Primary outcome was the composite of all-cause death, stroke, and major bleeding.

Results

A total of 36,617 patients were studied (mean age 70.14 ± 10.47 years; 45% female). Three phenotypes were identified: Cluster 1 (n=16,736) a hypertensive–metabolic phenotype, Cluster 2 (n=12,943) an older, high-comorbidity phenotype, and Cluster 3 (n=6,938) a younger, lower-comorbidity phenotype.

Survival analyses included 26,046 patients (71.1%) with available follow-up for the primary outcome. Kaplan–Meier curves demonstrated clear separation for the primary composite outcome across phenotypes (log-rank p<0.001), with the poorest event-free survival in Cluster 2 and the most favourable in Cluster 3. In Cox models with Cluster 1 as the reference group, Cluster 2 had higher risk of the primary composite outcome (HR 1.63, 95% CI 1.52–1.76), whereas Cluster 3 had lower risk (HR 0.54, 95% CI 0.48–0.62). Similar risk gradients were observed across all-cause death, cardiovascular death, MACE, and major bleeding. Results were consistent after adjustment for region and phase.

ABC pathway adherence was associated with lower risk overall (HR 0.73, 95% CI 0.66–0.79), with no evidence of effect modification by phenotype (global pint=0.262).

Conclusions

GTM-derived AF phenotypes showed clear and clinically meaningful prognostic separation for the primary composite outcome and major secondary outcomes, supporting the utility of data-driven ML phenotyping for risk stratification in a global real-world AF population.Derived phenotypes of AF in GLORIA-AFTime-to-Event Outcomes Stratified by Cluster