Machine learning to predict hemodynamically significant CAD based on traditional risk factors, coronary artery calcium and epicardial fat volume

We sought to establish an explainable machine learning (ML) model to screen for hemodynamically significant coronary artery disease (CAD) based on traditional risk factors, coronary artery calcium (CAC) and epicardial fat volume (EFV) measured from non-contrast CT scans. 184 symptomatic inpatients who underwent Single Photon Emission Computed Tomography/Myocardial Perfusion Imaging (SPECT/MPI) and Invasive Coronary Angiography (ICA) were enrolled. Clinical and imaging features (CAC and EFV) were collected. Hemodynamically significant CAD was defined when coronary stenosis severity ≥ 50% with a matched reversible perfusion defect in SPECT/MPI. Data was randomly split into a training cohort (70%) on which five-fold cross-validation was done and a test cohort (30%). The normalized training phase was preceded by the selection of features using recursive feature elimination (RFE). Three ML classifiers (LR, SVM, and XGBoost) were used to construct and choose the best predictive model for hemodynamically significant CAD. An explainable approach based on ML and the SHapley Additive exPlanations (SHAP) method was deployed to generate individual explanation of the model’s decision. In the training cohort, hemodynamically significant CAD patients had significantly higher age, BMI and EFV, higher proportions of hypertension and CAC comparing with controls ( P all < .05). In the test cohorts, hemodynamically significant CAD had significantly higher EFV and higher proportion of CAC. EFV, CAC, diabetes mellitus (DM), hypertension, and hyperlipidemia were the highest ranking features by RFE. XGBoost produced better performance (AUC of 0.88) compared with traditional LR model (AUC of 0.82) and SVM (AUC of 0.82) in the training cohort. Decision Curve Analysis (DCA) demonstrated that XGBoost model had the highest Net Benefit index. Validation of the model also yielded a favorable discriminatory ability with the AUC, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and accuracy of 0.89, 68.0%, 96.8%, 94.4%, 79.0% and 83.9% in the XGBoost model. A XGBoost model based on EFV, CAC, hypertension, DM and hyperlipidemia to assess hemodynamically significant CAD was constructed and validated, which showed favorable predictive value. ML combined with SHAP can offer a transparent explanation of personalized risk prediction, enabling physicians to gain an intuitive understanding of the impact of key features in the model.