Unsupervised Clustering of Patients Undergoing Thoracoscopic Ablation Identifies Relevant Phenotypes for Advanced Atrial Fibrillation

Background/Objectives: The rate of recurrence after ablation for atrial fibrillation (AF) is considerable. Risk stratification for AF recurrence after ablation remains incompletely developed. Unsupervised clustering is a machine learning technique which might provide valuable insights in AF recurrence by identifying patient clusters using numerous clinical characteristics. We hypothesize that unsupervised clustering identifies patient clusters with different clinical phenotypes, including AF type and cardiovascular morbidities, and ablation outcomes. Methods: Baseline and procedural characteristics of 658 patients undergoing thoracoscopic ablation for advanced AF (persistent, with enlarged left atria, or with previous failed catheter ablation) between 2008 and 2021 were collected. Principal component analysis (PCA) was used as an unsupervised dimensionality reduction technique, followed by K-Means clustering for unsupervised data clustering. The silhouette score was used to determine the optimal number of clusters, resulting in the formation of three clusters. CHA2DS2-VASc score and AF recurrence were not included in the clustering, but were compared between clusters. Moreover, we compared the patients with and without previously established risk factors for AF recurrence for each cluster. Results: Unsupervised clustering resulted in three distinct clusters. Cluster I had a significantly lower rate of AF recurrence than Cluster II, which contained significantly more persistent AF patients than the other clusters. The CHA2DS2-VASc score in Cluster III was significantly higher than in the other clusters. In all clusters, but particularly in Cluster III, the recurrence risk was higher for persistent AF patients and female patients. In Cluster II, the recurrence risk was not influenced by an increased left atrial volume index, unlike other clusters. Conclusions: Using unsupervised clustering of clinical and procedural data, we identified three distinct advanced AF patient clusters with differences in AF type, CHA2DS2-VASc score, and AF recurrence. We found that established risk factors like BMI, AF type, and LAVI vary in importance across clusters.