Machine Learning-Driven Identification of Distinct Persistent Atrial Fibrillation Phenotypes: A Cluster Analysis of DECAAF II

Introduction: Catheter ablation of persistent atrial fibrillation yields sub-optimal success rates partly due to the considerable heterogeneity within the patient population. Identifying distinct patient phenotypes based on post-ablation prognosis could improve patient selection for additional therapies and optimize treatment strategies.

Methods: We studied all patients who underwent catheter ablation of persistent atrial fibrillation in the DECAAF II trial. Out of 44 participating centers, 25% were randomly chosen as a validation set. A Gradient Boosting Method determined essential features for arrhythmia recurrence prediction and the number of clusters was determined according to the average silhouette width. K-medoids cluster analysis identified subgroups based on these features, and Kaplan-Meier curves were further compared among different clusters.

Results: Among 815 patients, 570 served as a training set and 245 as a validation set. Using the training set, the GBM model achieved an AUC of 0.874. K-medoids cluster analysis used LA volume, BMI, baseline fibrosis, and age, resulting in two clusters. Cluster 1 patients were older, had higher baseline fibrosis, higher BMI, and greater LA volume compared to Cluster 2. Atrial arrhythmia recurrence rates were significantly higher in Cluster 1 (51.7% vs. 35.0%, p = 0.0002), and survival analysis showed a significant difference in primary recurrence outcomes (HR = 1.71, p < 0.0001). The validation set confirmed these findings.

Conclusion: Utilizing machine learning, we identified a high-risk cluster for procedural failure in catheter ablation of persistent atrial fibrillation within the DECAAF II trial population. The primary differentiating factors of this high-risk cluster include older age, high left atrial fibrosis, elevated BMI, and increased left atrial volume.