Phenotyping of atrial fibrillation with cluster analysis and external validation
Yuki Saito, Yuto Omae, Koichi Nagashima, Katsumi Miyauchi, Yuji Nishizaki, Sakiko Miyazaki, Hidemori Hayashi, Shuko Nojiri, Hiroyuki Daida, Tohru Minamino, Yasuo Okumura
Abstract
OBJECTIVES: Atrial fibrillation (AF) is a heterogeneous condition. We performed a cluster analysis in a cohort of patients with AF and assessed the prognostic implication of the identified cluster phenotypes. METHODS: We used two multicentre, prospective, observational registries of AF: the SAKURA AF registry (Real World Survey of Atrial Fibrillation Patients Treated with Warfarin and Non-vitamin K Antagonist Oral Anticoagulants) (n=3055, derivation cohort) and the RAFFINE registry (Registry of Japanese Patients with Atrial Fibrillation Focused on anticoagulant therapy in New Era) (n=3852, validation cohort). Cluster analysis was performed by the K-prototype method with 14 clinical variables. The endpoints were all-cause mortality and composite cardiovascular events. RESULTS: The analysis subclassified derivation cohort patients into five clusters. Cluster 1 (n=414, 13.6%) was characterised by younger men with a low prevalence of comorbidities; cluster 2 (n=1003, 32.8%) by a high prevalence of hypertension; cluster 3 (n=517, 16.9%) by older patients without hypertension; cluster 4 (n=652, 21.3%) by the oldest patients, who were mainly female and with a high prevalence of heart failure history; and cluster 5 (n=469, 15.3%) by older patients with high prevalence of diabetes and ischaemic heart disease. During follow-up, the risk of all-cause mortality and composite cardiovascular events increased across clusters (log-rank p<0.001, p<0.001). Similar results were found in the external validation cohort. CONCLUSIONS: Machine learning-based cluster analysis identified five different phenotypes of AF with unique clinical characteristics and different clinical outcomes. The use of these phenotypes may help identify high-risk patients with AF.