Unsupervised learning identifies patient phenotypes and recurrence risk after cryoballoon PVI

S. Ganesh (Baltimore)1, I. Vatsaraj (Baltimore)1, L. Steffens (Köln)2, H. Horlitz (Köln)2, J. Fernholz (Köln)2, J. Hein (Köln)2, S. Ak (Köln)2, I. Novikov (Köln)2, L. Grüne (Köln)2, M. Knitter (Köln)2, M. Horlitz (Köln)2, F. Stöckigt (Köln)2, N. Trayanova (Baltimore)3, Y. Mohsen (Köln)2
1Johns Hopkins University Biomedical Engineering Department Baltimore, USA; 2Krankenhaus Porz am Rhein gGmbH Klinik für Kardiologie, Elektrophysiologie u. Rhythmologie Köln, Deutschland; 3Johns Hopkins University Department of Biomedical Engineering Baltimore, USA

Background

Recurrence after cryoballoon pulmonary vein isolation (PVI) remains common. How multiparametric procedural patterns interact with comorbidities and demographics to shape recurrence risk is unclear. We tested whether model-derived patient embeddings can uncover latent phenotypes linked to recurrence.

Objective

To identify clinically interpretable patient/procedural clusters associated with atrial fibrillation (AF) recurrence using unsupervised machine learning.

Methods

We analyzed 621 patients and 152 variables (demographics, comorbidities, medications, cryo parameters, outcomes). A supervised XGBoost model was trained to predict recurrence with the recurrence label used only as the target (not as a feature). Per-patient, per-feature Shapley values (SHAP) quantified feature contributions and served as embeddings for clustering. K-means on SHAP profiles identified clusters; the number of clusters was chosen using silhouette, Calinski–Harabasz, and Davies–Bouldin indices and corroborated with UMAP visualization. Cluster-level recurrence proportions and feature importance patterns were then summarized.

Results

AF recurrence occurred in 267/621 patients (43%),59.6% male; mean age 66.9 ± 11.0 years.
Four clusters were identified (Figure 1): C1 (n=197, 4% recurrence), C2 (n=190, 89%), C3 (n=89, 85%), and C4 (n=145, 10%).

  • C1 (low risk): longer total freeze time in RSPV/LIPV, digitalis use, and lower RSPV minimum freeze temperature; EHRA class and EP-lab duration were secondary.

  • C2 (high risk): lower symptom class (EHRA) and fewer RIPV freezes; mixed signals from ACE use, PV temperature metrics (LSPV/RSPV/RIPV), CRP, and height contributed to the inclusion in ths cluster.

  • C3 (high risk): prior myocardial infarction and higher BMI/weight; higher RIPV freeze temperature at 30 sec, fluoroscopy dose/time, longer procedure duration, and LIPV isolation temperature.

  • C4 (low–intermediate risk): prior cardioversion enriched; absence of cardioversion favored inlcusion in this cluster. Hemoglobin, weight, valve surgery, MCV, prior ablation, height, RSPV temperature, and creatinine were associated with the inclusiono in this cluster without a consistent direction.

Conclusion

Unsupervised learning on model-derived embeddings revealed four reproducible patient–procedure phenotypes with markedly different recurrence risks (4–89%). This approach captures multivariate, non-linear patterns—spanning cryo dosing, temperatures, and clinical comorbidity profiles—that are not readily detected with conventional univariate or standard multivariable analyses. If validated prospectively, such phenotyping could enable pre-procedural risk stratification, guide intraprocedural dosing strategies, and inform enrichment/stratification in future trials.