Unsupervised machine learning of multimodal data to identify atrial fibrillation subgroups likely to benefit from pulmonary vein isolation

Background The heterogeneity in response to pulmonary vein isolation emphasises the requirement to develop new methods to phenotype patients with atrial fibrillation (AF). Multimodal data characterising patients in both hospital and community settings could facilitate a personalised approach towards management, improving risk stratification and informing shared decision making prior to catheter ablation. Purpose To apply unsupervised machine learning to clinical, ambulatory monitoring and pre-procedural magnetic resonance imaging (MRI) data to define AF phenotypes and compare rates of arrhythmia recurrence following pulmonary vein isolation. Methods We conducted a multicentre, prospective observational cohort study in 61 patients (34% persistent AF) undergoing first-time pulmonary vein isolation (±cavotricuspid isthmus ablation). Pre-procedural late gadolinium enhancement MRI was performed to measure bi-atrial size and quantify the extent of atrial fibrosis. A two-week ambulatory monitor and AF6 questionnaire assessed AF burden and patient-reported outcomes. These data were combined with clinical data, and K-means clustering was employed to define AF phenotypes using principal component analysis. Missing data were addressed using mean imputation. Patients were followed up for 12 months post-ablation, and time-to-event analyses were performed. Results Three AF phenotypes were identified (Figure 1). Cluster 1 (n = 7; 57% persistent AF) had the lowest recurrence rate (29%) despite having the highest median CHA2DS2-VA score (2; IQR 1-4) and heart failure rate (71%). Cluster 2 (n = 40, 13% persistent AF) had a similar recurrence rate (35%) but lower median CHA2DS2-VA score (1, IQR 0-2). Cluster 3 (n = 14, 93% persistent AF) had the highest recurrence rate (57%) and a median CHA2DS2-VA score of 1.5 (IQR 1-2.75). Compared with Cluster 3, recurrence risk was numerically lower in Cluster 1 (HR 0.34, 95% CI 0.07–1.62, p = 0.176) and significantly lower in Cluster 2 (HR 0.41, 95% CI 0.17–0.98, p = 0.044) (Figure 2). Of the 21 persistent AF patients, 43% (9/21) were in Clusters 1 or 2. Conclusions Unsupervised machine learning identified distinct AF subgroups who varied in response to ablation. This approach may help identify patients with persistent AF who are most likely to benefit from pulmonary vein isolation.