Parsimonious Machine Learning of Preprocedural Cardiac CT Predicts All-Cause Mortality After Transcatheter Aortic Valve Implantation
Davide Vignale, Andrea Corvaglia, Anna Palmisano, Patrick Scuri, Chiara Gnasso, Bruno Fabiani, Gabriella Amaro, Simone Barbieri, Davide Margonato, Alberto Traverso, Ludovica Bognoni, Takahiro Nishihara, Chloe Yeabin Jung, Marco Denti, Joao Cavalcante, Eustachio Agricola, Francesco Maisano, Matteo Montorfano, Carlo Tacchetti, Antonio EspositoAbstract
Purpose
To develop a machine learning (ML) model using preprocedural cardiac CT data for predicting 1-year mortality after transcatheter aortic valve implantation (TAVI) for severe aortic stenosis (AS), deployable as a web-based solution.
Materials and Methods
A retrospective single-center study was conducted on consecutive participants undergoing TAVI from October 2020 to March 2023 (training) and April to November 2023 (internal testing), with CT performed on a dual-source scanner SOMATOM Definition Flash or Drive. Two logistic regression models were developed: full-variable (clinical, echocardiographic, and cardiac CT features) and CT-only (left ventricular ejection fraction and indexed right atrial end-diastolic volume). ROC analysis and risk stratification by tertiles were used for performance evaluation. The CT-only model was further tested with features automatically extracted using TotalSegmentator.
Results
The training set included 495 participants (260[53%] women; median age 82 years [IQR,78–85]); the internal test set included 143 (72[50%] women; median age 82 years [IQR,77–85]). One-year mortality occurred in 56/495 (11%) and 14/143 (10%), respectively. The CT-only model achieved an AUC of 0.67 [95%CI,0.59-0.75], comparable to the full-variable model (AUC: 0.70 [95%CI,0.62-0.78]). Internal testing confirmed performance (AUC: 0.68 [95%CI, 0.54-0.81]) and risk stratification into groups of low- (1.7%), intermediate- (13.5%), and high-risk (17.0%) of 1-year mortality. Automatically extracting CT metrics with TotalSegmentator preserved discrimination (AUC: 0.67 [95%CI, 0.54-0.79]), with outcome probability of 4.3% (low-risk), 11.1% (intermediate-risk), and 14.3% (high-risk). The model was successfully deployed via a web-application (https://uc3-model.srace.ai-hub.it/).
Conclusion
An ML model using two preprocedural cardiac CT-derived metrics, automatically extractable with open-source tools, predicts 1-year mortality after TAVI. A web-based solution has been made publicly available to facilitate external validation, deployment for prognostication independent of commercial software or multi-source data integration.