Atrial fibrillation (AF) is the most common sustained cardiac arrhythmia, affecting 1% to 2% of the global population, and is expected to increase due to aging populations and the rising burden of cardiovascular risk factors [1]. AF poses a significant public health challenge due to its strong associations with thromboembolic events, anticoagulation-related bleeding, cardiovascular complications, and increased mortality [2,3]. Despite its high prevalence, AF remains clinically heterogeneous, with patients presenting diverse profiles, comorbidities, and risk factors [4].

Traditional risk scores and classifications based on arrhythmia patterns or subtypes often fail to capture this complexity, limiting their effectiveness for accurate risk stratification and personalized treatment [5]. Given this heterogeneity, classifying patients into clinically relevant phenotypes could enhance risk assessment and support more individualized therapeutic strategies [6].

Previous methods for identifying AF phenotypes, such as hierarchical clustering [7,8] and k-prototype [5] algorithms, have limitations in fully capturing patient heterogeneity. These approaches are often constrained by predefined assumptions and may overlook subtle relationships between clinical variables [9]. In contrast, a novel machine learning methodology [10] that uses generative topographic mapping (GTM) offers a more flexible, data-driven approach. This methodology [10] facilitates the identification and visualization of complex patient distributions in a probabilistic and interpretable manner [11], enabling the discovery of clinically meaningful phenotypes [10,12]. To date, GTM has not been applied to identify phenotypes of AF in the context of a prospective registry.

In this study, we applied a GTM-based clustering approach to a large, real-world, prospective AF cohort to identify clinically relevant phenotypes and assess their associations with clinical outcomes.

In this study, we applied a novel machine learning approach based on GTM to identify clinical phenotypes within the MAFP-III cohort. Our findings are twofold: (1) GTM proved highly effective in detecting clinically meaningful phenotypic subgroups, thereby enhancing patient stratification; and (2) these phenotypes were strongly associated with distinct prognostic outcomes. Its probabilistic framework enabled refined patient classification by managing uncertainty, capturing latent structures, and visualizing complex, high-dimensional data in a low-dimensional space, ultimately improving interpretability and supporting clinical decision-making.

This probabilistic framework also enables the identification of patients with lower maximum membership probabilities, typically located near cluster boundaries in the latent space. These individuals may represent transitional or overlapping phenotypes, reflecting more heterogeneous or evolving clinical profiles [11,12]. This may be particularly relevant in AF, where multimorbidity and dynamic risk trajectories are common. Such probabilistic information may therefore provide additional clinical value by capturing uncertainty and supporting more nuanced, personalized risk assessment, although this warrants further investigation.

From a clinical perspective, GTM complements existing risk scores such as CHA₂DS₂-VASc, CHA₂DS₂-VA, and HAS-BLED by integrating the broader cardiometabolic burden. This enables identification of patient subgroups simultaneously at high thromboembolic, bleeding, and cardiovascular risk profiles that may be overlooked when risks are considered separately. Such information could support more precise risk stratification, prioritize closer follow-up, optimize preventive strategies, and tailor therapeutic intensity in clinical practice [17,20].

In this context, GTM-derived phenotypes may provide complementary prognostic information beyond conventional risk scores by capturing complex, nonlinear interactions between cardiometabolic factors that are not fully reflected in additive scoring systems. This approach may be particularly valuable in refining risk stratification among patients with intermediate or overlapping risk profiles [10].

Beyond direct clinical application, GTM-based stratification may also enhance clinical research by enriching trial populations and identifying subgroups with differential treatment responses [21]. As electronic health records (EHRs) and decision-support tools expand, GTM could be embedded as an analytical layer, automatically classifying patients and translating complex data into actionable insights without increasing clinician workload.

From a digital health and implementation perspective, this framework could be operationalized within EHR systems to enable automated, real-time phenotypic assignment at the point of AF diagnosis. Once trained, the model allows projection of new patients into the latent space using routinely collected structured clinical data, facilitating scalable deployment and seamless integration into clinical workflows. Importantly, the probabilistic and visual nature of GTM may enhance interpretability and support clinician trust in AI-driven outputs. Future work should focus on prospective validation, interoperability with existing EHR infrastructures, and the development of user-centered clinical decision support tools [22].

Building upon this implementation framework, although phenotypes were derived from baseline characteristics, the GTM model also enables projection of longitudinal patient data onto the same latent space, allowing dynamic remapping as clinical profiles evolve over time [23]. This longitudinal extension may facilitate tracking of phenotypic trajectories and provide insights into disease progression and time-updated risk stratification. Future studies should incorporate repeated measurements and temporal validation to determine the clinical relevance of these dynamic phenotypic transitions.

Compared to traditional clustering methods such as hierarchical clustering and k-prototype, GTM showed superior performance in complex clinical datasets with mixed variable types [10,24]. Unlike approaches that impose rigid cluster boundaries or assume linear separability, GTM uncovers nonlinear relationships and accommodates the natural heterogeneity of AF populations [7,16,25,26]. Its probabilistic nature provides flexible clustering and a topologically coherent 2D representation, facilitating intuitive understanding of patient distributions and clinical trajectories [27].

In our cohort, GTM identified 4 distinct clinical phenotypes: (1) older patients with the highest cardiometabolic burden, (2) older patients with lower cardiometabolic risk, (3) comparatively younger patients with intermediate-high cardiometabolic risk, and (4) comparatively younger patients with intermediate cardiometabolic risk.

Phenotype 1 represented the least favorable reference profile, with the highest overall burden of adverse outcomes. After accounting for competing mortality risk, phenotype 4 demonstrated significantly lower risks of thromboembolic events, major bleeding, and MACE compared with phenotype 1, in addition to reduced cardiovascular and all-cause death. Phenotype 2 showed lower risks of thromboembolic events and MACE; however, its associations with cardiovascular and all-cause death were attenuated and no longer statistically significant after adjustment for age, sex, and comorbidities. Phenotype 3 was associated with lower all-cause death, although its associations with cardiovascular death and nonfatal outcomes were attenuated and less consistent after full adjustment. Age and cardiometabolic comorbidities, such as hypertension, diabetes, obesity, and vascular disease, play a central role in atrial remodeling, systemic inflammation, and endothelial dysfunction, thereby creating a prothrombotic milieu that increases the risk of thromboembolic events in AF [28,29]. This relationship was reflected in our cohort, where the phenotype characterized by the highest cardiometabolic burden (phenotype 1) consistently showed the least favorable prognosis, whereas the comparatively younger phenotype with intermediate cardiometabolic burden (phenotype 4) exhibited the most favorable overall outcome profile, particularly for thromboembolic events, major bleeding, MACE, and mortality. By integrating multidimensional clinical information, GTM-derived phenotypes capture the complex interplay between cardiometabolic status and prognosis, complementing conventional risk scores and refining risk stratification.

To the best of our knowledge, the only other study deriving AF phenotypes with a robust AI-based probabilistic approach was conducted by Bellfield et al [10]. Our findings build on and extend their work, showing that GTM can generate clinically meaningful phenotypes across diverse populations, with notable parallels to those reported in the UK Biobank and MIMIC-IV (Medical Information Mart for Intensive Care IV) cohorts. Although comparisons with the general AF cohort from the UK Biobank may be complex, we observed several similarities. For instance, our phenotype 1 appears analogous to their phenotype 3, characterized by older age and extensive cardiovascular and renal comorbidities. Similarly, our phenotype 2, comprising predominantly women with low cardiometabolic risk, mirrors their phenotype 5.

Moreover, comparisons with studies using traditional clustering techniques revealed partial concordance. For instance, Chen et al [7] applied hierarchical clustering to a cohort of 3348 patients with AF from the KERALA-AF Registry, identifying 5 distinct clinical phenotypes. Among these, Cluster 2, comprising patients aged older than 65 years with a high burden of multimorbidity, including cardiovascular-renal-metabolic syndrome, and Cluster 4, characterized by patients with heart failure and multiple comorbidities, appear comparable to our phenotypes 1 and 3, respectively. In their analysis, both clusters were associated with a significantly increased risk of MACE when compared with cluster 3, which included patients aged 65 years or younger with few comorbidities (cluster 2: odds ratio [OR] 1.79, 95% CI 1.42-2.25; cluster 4: OR 1.76, 95% CI 1.31-2.36).

Similarly, in the study by Bisson et al [30], conducted within the Loire Valley AF cohort, cluster 2—which included older patients with permanent AF, structural cardiac abnormalities, and a substantial comorbidity burden—resembles our phenotype 1. This cluster also showed a significantly elevated risk of all-cause death (hazard ratio 3.54, 95% CI 1.49-8.43) when compared with a reference cluster composed of patients with fewer comorbidities.

Altogether, our findings confirm that GTM is a powerful and reliable tool for uncovering actionable phenotypic patterns in AF, offering a pathway toward more personalized, prognosis-guided care. The identification of clinically meaningful and prognostically distinct phenotypes supports its potential integration into clinical workflows, where it could enable earlier identification of patients at high-risk and more precise allocation of preventive and therapeutic strategies. Future studies should validate these findings in independent cohorts and assess its use in real-time practice, as well as its added value when combined with longitudinal data and biomarkers. As health care moves toward precision medicine, GTM exemplifies how AI can translate complex data into actionable insights, bringing us closer to truly personalized AF management.

This study has several limitations. First, the nonrandomized nature of the study limits our ability to infer causality between phenotypes and outcomes. Although we adjusted for major confounders, residual confounding cannot be entirely excluded. Second, phenotypes were derived solely from baseline characteristics, without accounting for temporal changes. As risk factors evolve over time, some patients may shift phenotypic categories during follow-up. Although the GTM framework allows projection of new data onto the latent space, enabling potential dynamic remapping, this was not explored in the present study. Third, although we used rigorous internal validation to determine the optimal number of clusters, alternative methods or thresholds may yield different results. Furthermore, although internal cross-validation was used to optimize model performance, the identified phenotypes were not externally validated in an independent cohort. Therefore, the generalizability and reproducibility of these findings across independent AF populations remain to be established. Fourth, although key variables such as age and cardiometabolic comorbidities are central to AF pathophysiology and were intentionally included to enhance clinical relevance, we did not perform sensitivity analyses excluding these factors. Therefore, the robustness of the identified phenotypes under alternative modeling configurations remains uncertain and should be explored in future studies. Additionally, we did not include certain clinically relevant variables, such as cancer, in the phenotyping model. Although malignancy may have important prognostic implications [31], its heterogeneous nature and complex interactions with clinical management may limit its integration within a cardiometabolic-focused modeling framework. Future studies should explore the impact of incorporating such variables into GTM-based phenotyping. Fifth, all participants were already receiving OAC therapy. This may limit the applicability of findings to untreated AF populations or those on different therapeutic regimens. Sixth, the cohort was predominantly Caucasian, potentially restricting generalizability to more ethnically diverse populations. Given known disparities in AF outcomes by ethnicity, further validation in broader populations is warranted [32,33]. Finally, we did not perform direct head-to-head comparisons between GTM-derived phenotypes and established risk scores such as CHA₂DS₂-VASc and CHA₂DS₂-VA (for thromboembolism) or HAS-BLED (for bleeding). In addition, the applicability of GTM-based phenotyping in routine clinical practice remains exploratory and requires prospective validation before it can be integrated into decision-making.

In conclusion, this clustering analysis using GTM identified 4 distinct clinical phenotypes among patients with AF receiving oral anticoagulation, each associated with different prognostic profiles. GTM enabled the modeling of complex, nonlinear relationships within high-dimensional clinical data, facilitating stratification into clinically meaningful phenotypic subgroups. These findings support the potential utility of GTM-based phenotyping for improving risk characterization in AF. However, external validation studies are required to determine the generalizability and clinical applicability of these phenotypes.