AI-Based Phenotyping of Atrial Fibrillation Through Generative Topographic Mapping: Prospective Murcia Atrial Fibrillation Project III Cohort Study.

BACKGROUND

The clinical heterogeneity of atrial fibrillation (AF) challenges current classifications and risk scores, limiting their real-world applicability. AI-driven methods may enhance phenotyping and risk stratification.

OBJECTIVE

This study aimed to apply a generative topographic mapping (GTM)-based clustering approach to a large, real-world, prospective AF cohort to identify clinically relevant phenotypes and assess their associations with clinical outcomes.

METHODS

We conducted a prospective observational cohort study including consecutive adult outpatients with newly diagnosed AF who initiated oral anticoagulation between January 2016 and November 2021. Cardiometabolic risk was modeled as a multidimensional construct integrating cardiovascular and metabolic comorbidities. GTM was applied to project high-dimensional clinical data into a low-dimensional latent space, enabling probabilistic patient representation and visualization of phenotypic structure. Unsupervised hierarchical clustering using Ward's method was performed on the latent space to identify AF phenotypes, and patients were assigned to clusters based on their highest posterior probability. Clinical outcomes, including thromboembolic events, major bleeding, major adverse cardiovascular events (MACE), cardiovascular death, and all-cause death, were assessed over a maximum follow-up of 2 years. Nonfatal outcomes were analyzed using Fine-Gray competing-risk models and reported as subdistribution hazard ratios (sHRs), whereas cardiovascular and all-cause death were analyzed using adjusted Cox proportional hazards models and reported as adjusted hazard ratios (aHRs).

RESULTS

Among 3259 patients with AF (median age 77, IQR 70-83 years; 1721/3259, 52.8% female; mean follow-up 1.81 years), four phenotypes emerged: (1) older with highest cardiometabolic burden, (2) older with lower cardiometabolic risk, (3) comparatively younger with intermediate-high cardiometabolic risk, and (4) comparatively younger with intermediate cardiometabolic risk. Compared with phenotype 1, phenotype 4 demonstrated lower risks of thromboembolic events (sHR 0.70, 95% CI 0.50-0.98), major bleeding (sHR 0.61, 95% CI 0.42-0.89), and MACE (sHR 0.67, 95% CI 0.47-0.94). Regarding cardiovascular and all-cause death, all phenotypes demonstrated a lower risk compared with phenotype 1: phenotype 2 (aHR 0.52, 95% CI 0.33-0.85; and aHR 0.66, 95% CI 0.50-0.87, respectively), phenotype 3 (aHR 0.57, 95% CI 0.32-0.99; and aHR 0.44, 95% CI 0.30-0.65, respectively), and phenotype 4 (aHR 0.49, 95% CI 0.33-0.72; and aHR 0.56, 95% CI 0.44-0.71, respectively). However, these associations were attenuated after further adjustment for age, sex, and major comorbidities, with phenotype 4 retaining a significant association with lower cardiovascular death and phenotypes 3 and 4 retaining significant associations with lower all-cause death.

CONCLUSIONS

GTM-based clustering analysis identified prognostically distinct AF phenotypes, highlighting the potential of machine learning approaches to support personalized AF care. Further external validation is needed to establish the generalizability of these findings.