Interpretable and reproducible machine learning model for coronary calcification and segment-level stenoses stratification on computed tomography angiography.

Abstract

BACKGROUND: Coronary computed tomography angiography (CCTA) is widely used as a first-line tool for diagnosing and managing coronary artery disease (CAD), and machine learning (ML)-based analysis shows promise for quantitative CAD assessment. METHODS: In this post hoc analysis of 909 participants from the SCOT-HEART trial (median follow-up, 5.8 years), we first evaluated the distribution of CCTA-derived imaging features in a cohort (n = 221) with a zero calcium score, stenoses < 10%, and no evidence of CAD on CCTA, across 21 image processing settings. Interpretable ML models were then developed and validated to quantify coronary calcification and stenoses in major coronary segments (LMA, LCX, LAD, pRCA, mRCA). Calcified plaques, stenoses, and myocardial infarction outcomes were comprehensively assessed. RESULTS: A total of 549 stable imaging features was identified across processing settings. Six ML algorithms (SVM, KNN, MLP, Naïve Bayes, gradient boosting, LightGBM) were evaluated for predicting coronary calcification and stenoses. The best model achieved an accuracy of 84.2% and an AUC of 0.973. Stenosis stratification accuracy exceeded 84.8% across all segments, with minimal (< 0.05) differences between models using all versus stable features. SHAP analysis indicated heterogeneous contributions of imaging phenotypes and clinical risk factors. CONCLUSIONS: Stable imaging features provide a reference for future ML-based coronary quantitatively assessments. Interpretable ML models demonstrated promising performance in quantifying coronary calcification and segment-level stenoses.