Explainable machine learning and unsupervised anatomical phenotyping of mitral valve prolapse in patients with severe mitral regurgitation.

backgroundMitral annular disjunction (MAD) is increasingly recognized in mitral valve prolapse (MVP), but its anatomical embedding within degenerative MVP remains incompletely defined. We assessed whether explainable supervised learning and unsupervised clustering could identify MAD-associated MVP phenotypes.

methodsWe analyzed a public surgical cohort of 979 patients with MVP and clinically significant mitral regurgitation. Etiology was classified as Barlow disease (BD) or fibroelastic deficiency (FED). Logistic regression, Random Forest and XGBoost were evaluated for MAD prediction using nested cross-validation with threshold optimization, and assessed for discrimination, calibration and net benefit (decision-curve analysis). XGBoost explainability was assessed using SHAP. UMAP-based clustering using anatomical and clinical variables (excluding MAD and etiology from the input) was performed, with prespecified cluster-number and stability sensitivity analyses.

resultsUnsupervised phenotyping identified three anatomical clusters. The smallest cluster (n = 91, 9.3%) showed the highest MAD prevalence (36.3% vs. 12.0% and 15.8%), marked BD enrichment (95.6%), universal bileaflet prolapse and ≥ 4-scallop involvement, larger MV diameters and less chordal rupture; this phenotype was reproducible across UMAP settings. Supervised models showed only moderate MAD discrimination (XGBoost AUROC 0.695 ± 0.057; PR-AUC 0.314 ± 0.066), with modest calibration (Brier 0.13-0.20; slopes < 1) and limited net benefit, and were therefore interpreted as explanatory rather than diagnostic. SHAP identified BD status as the strongest contributor, followed by number of prolapsing scallops, mitral valve diameters, LV volumes and bileaflet prolapse.

conclusionExplainable and unsupervised analyses concordantly identified a MAD-enriched, BD-dominant bileaflet/multiscallop MVP phenotype, supporting the concept that MAD is embedded within broader annular-leaflet remodeling rather than being an isolated binary abnormality. These findings require prospective, multimodality, outcome-linked external validation before clinical application.