Machine-learning prediction of patient-specific mitral valve anatomical parameters for reconstructive planning

Reconstruction of the mitral valve using autologous pericardium may benefit from individualized leaflet dimensions, whereas conventional proportional rules provide only simplified anatomical estimates. This study aimed to develop a machine-learning framework for patient-specific prediction of mitral leaflet dimensions. Clinical and morphometric data from 72 adult autopsy cases without structural mitral valve disease were analyzed. Eight regression approaches were compared in separate target-specific modeling workflows. A clinically oriented configuration used 17 patient- and valve-level inputs to estimate eight anatomical dimensions required for leaflet template construction. Predictive performance varied across anatomical targets. The final model set comprised four ElasticNet and four random forest regressors. Prediction was most accurate for P1 and P3 heights, for which the mean absolute error (MAE) was 1.68–1.70 mm and R 2 was 0.54–0.55. Among the anterior-leaflet parameters, A3 height showed the best performance (MAE 2.33 mm; R 2 0.21). Posterior leaflet free-edge length remained the least predictable target (MAE 13.32 mm; R 2 0.02). Model outputs were incorporated into an interactive tool for generating individualized computer-aided leaflet templates. The proposed proof-of-concept workflow connects anatomical measurements, machine-learning estimation, and digital template construction for personalized mitral valve reconstruction. Prospective comparison of imaging-derived and anatomical measurements, followed by validation in independent cohorts, is required before clinical implementation.