%0 Journal Article
%T Explainable and Reproducible Machine Learning Framework for Assessment of Coronary Calcification and Segmental Stenosis on Computed Tomography Angiography
%A Santiago Morales
%A Laura Rojas
%A Camila Vega
%A Diego Molina
%J Interdisciplinary Research in Medical Sciences Specialty
%@ 3062-4401
%D 2026
%V 6
%N 1
%R 10.51847/vlyNZ5PaMM
%P 194-206
%X CCTA has become a standard noninvasive imaging technique for the initial diagnosis and clinical management of coronary artery disease (CAD). Advanced machine learning techniques applied to CCTA data demonstrate considerable potential for precise quantitative evaluation of CAD. In a post hoc study involving 909 subjects from the SCOT-HEART trial (with a median follow-up period of 5.8 years), we initially examined the variability of CCTA-based imaging parameters in a reference group of 221 individuals characterized by zero coronary artery calcium scores, stenoses under 10%, and absence of any CAD findings on imaging. This evaluation spanned 21 distinct image processing configurations. We subsequently constructed and validated interpretable ML models designed to measure the extent of coronary calcification and degree of narrowing in the primary coronary artery branches (LMA, LCX, LAD, pRCA, mRCA). Comprehensive evaluation was performed on calcified plaque burden, stenosis severity, and myocardial infarction outcomes. Across the different processing approaches, a set of 549 robust and stable imaging features was determined. Performance of six machine learning classifiers (SVM, KNN, MLP, Naïve Bayes, gradient boosting, and LightGBM) was assessed for the prediction of coronary calcification and stenoses. The leading model delivered 84.2% accuracy along with an AUC of 0.973. Segment-level stenosis classification accuracy remained above 84.8% for all major vessels. Model performance showed only minor variations (less than 0.05) between those utilizing the complete feature set and those limited to stable features alone. SHAP interpretability analysis highlighted differing levels of importance for imaging characteristics and traditional clinical risk variables. The identified stable imaging features establish a reliable foundation for future development of ML-driven quantitative coronary assessments. The proposed interpretable machine learning models exhibited encouraging results in the quantification of coronary calcification and location-specific stenoses.
%U https://galaxypub.co/article/explainable-and-reproducible-machine-learning-framework-for-assessment-of-coronary-calcification-and-co1jsfeugkc2jdj