Artificial Intelligence-Derived Computed Tomography Phenotyping for Cardiovascular Risk Stratification in Transcatheter Mitral Valve Replacement
Liliane Zillner, Tillmann Kerbel, Paul Sautner, Mirjam G. Wild, Michaela M. Hell, Elmar W. Kuhn, Tanja K. Rudolph, Thomas Walther, Lenard Conradi, Andreas Zierer, Francesco Maisano, Marco Russo, Fabrizio Rosati, Andrea Colli, Miguel Piñón, David Reineke, Gaby Aphram, Christophe Dubois, Jörg Hausleiter, Ralph Stephan von Bardeleben, Iuliana Coti, Christian Loewe, Daniel Zimpfer, Georg Heinze, Martin AndreasBackground: Despite high procedural success of transcatheter mitral valve replacement (TMVR), mortality remains substantial, and conventional risk scores inadequately reflect the specific risk profile of this high-risk population. Artificial intelligence (AI)-derived analysis may enable TMVR-specific risk assessment using routinely acquired computed tomography (CT). Methods: In a retrospective European multicenter registry, AI-based segmentation of harmonized real-world pre-procedural cardiac CT with variable thoracic and upper-abdominal coverage was applied in patients undergoing TMVR with the TendyneTM system (134 patients; 346 CT-derived variables retained for high-dimensional screening). One-year cardiovascular mortality was assessed using Cox proportional hazards models with Benjamini–Hochberg false discovery rate correction. STS- and EuroSCORE II-adjusted analyses and a targeted coronary-ventricular coupling analysis were performed. Results: In the STS-adjusted Cox analysis, nine AI-derived CT parameters met the 10% FDR discovery threshold for cardiovascular mortality, spanning skeletal (n = 3), muscular (n = 1), and cardiopulmonary phenotypes (n = 5), whereas three parameters met this threshold in CT-only analyses. Coronary–ventricular coupling, defined as coronary artery volume relative to left ventricular end-diastolic volume, was significantly associated with lower 1-year cardiovascular mortality in the CT-only analysis (HR per IQR 0.47, 95% CI 0.22–1.00; p = 0.049) and after EuroSCORE II adjustment (HR 0.47, 95% CI 0.22–0.99; p = 0.046), while the STS-adjusted estimate remained directionally consistent but did not reach statistical significance (HR 0.49, 95% CI 0.23–1.05; p = 0.067). Conclusions: AI-based CT phenotyping identifies imaging-derived markers associated with cardiovascular mortality beyond information captured by conventional surgical risk scores in TMVR patients. Coronary-ventricular coupling emerges as a promising automated CT-derived imaging marker with a biologically plausible pathophysiological basis, linking coronary arterial volume to ventricular size in mitral valve disease.