Background: Transcatheter aortic valve implantation (TAVI) is established as standard of care for severe symptomatic aortic stenosis across all surgical risk strata. The risk scores used in routine practice (STS, EuroSCORE II) were derived from surgical cohorts and discriminate one-year mortality only modestly (C-statistic 0.58 for STS and 0.66 for EuroSCORE II). Reliable, individualized prediction of one-year mortality therefore remains an unmet need, and heart-team decisions still rely on clinical judgment.
Purpose: To develop and internally validate an interpretable machine-learning model for predicting one-year all-cause mortality after TAVI, and to quantify the incremental predictive value of a small set of procedural variables added to baseline data.
Methods and Results: We analyzed 1,200 complete datasets of patients undergoing TAVI for severe aortic stenosis at our center; one-year all-cause mortality was 18%. Sixty-five baseline features (including valve type and size) were screened; highly collinear variables were removed using the variance inflation factor, and missing values were imputed according to feature class. Final feature selection was based on impurity- and permutation-based feature importance. A gradient-boosted decision-tree classifier (XGBoost) was trained, and performance was assessed by five-fold nested stratified cross-validation using ROC analysis and the mean area under the curve (AUC). The baseline model reached a mean AUC of 0,71. Adding procedural variables, recorded at the end of the procedure / at ICU admission, increased the mean AUC to 0,82.
Conclusions: Current heart-team decision-making relies largely on pre-procedural data, yet its discriminative ceiling is low. Our findings show that a decisive share of prognostic information arises during and after the procedure. This is biologically coherent: periprocedural complications independently predict one-year mortality (major stroke adjusted HR 5.4; stage-3 acute kidney injury HR 4.9). Adding procedural variables raised the mean AUC. However thoroughly a case is discussed beforehand, outcome is ultimately shaped by procedural factors, device haemodynamics, and the early course—risk assessment should therefore remain continuous, extending into the intensive-care phase.
