Abstract
Abstract
Objectives
Subclinical leaflet thrombosis (SLT) is an early form of bioprosthetic valve dysfunction after transcatheter aortic valve implantation. Predicting SLT remains challenging. We aimed to apply machine learning not only to identify predictors but also to interpret their dynamic, non-linear effects on SLT risk using accumulated local effects (ALEs) curves, a robust method that accounts for variable collinearity.
Methods
A prospective cohort of 128 consecutive patients receiving a self-expanding transcatheter heart valve underwent multimodality imaging and haematological profiling. The primary outcome was SLT on 6-month CT. An extreme gradient boosting classifier was trained on 126 variables. Model performance was evaluated via nested cross-validation. Interpretability used SHapley Additive exPlanations and ALEs plots.
Results
Subclinical thrombosis was detected in 22 patients (17.1%). Extreme gradient boosting model demonstrated excellent discrimination (AUC: 0.91, 95% CI, 0.87-0.94) and calibration (Brier score: 0.09). SHapley Additive exPlanations analysis identified the top predictors: bicuspid valve anatomy (mean: 0.45), baseline haemoglobin (0.28), periprocedural Δhaematocrit (0.19), prosthesis eccentricity (0.14), and post-procedural platelet nadir (0.09). ALEs curves revealed a U-shaped association for baseline haemoglobin (lowest risk around 13 g/dL) a monotonic rise in SLT risk with increasing prosthesis eccentricity and greater periprocedural declines in haematocrit, and a progressively higher risk with post-procedural thrombocytopaenia.
Conclusions
This study introduces ALEs plots to cardiovascular medicine, translating predictions into interpretable risk curves that show how leaflet thrombosis risk evolves across key predictors, linking valve geometry and periprocedural haematological changes to guide patient surveillance after transcatheter aortic valve replacement.