Work

Machine-Learning Algorithms and Benchmarking Missing-Data Strategies for Predicting Angiographically Severe Coronary Artery Stenosis: A Retrospective Study

Year: 2026
Type: preprint
Study design: Observational Study
Abstract: Background Missing data is common in clinical datasets used to predict coronary artery disease (CAD) and can influence the relative performance of machine learning algorithms. This study evaluated and compared missing data handling strategies and machine learning models for predicting clinically significant coronary artery stenosis (70% or greater) using clinical variables. Methods In this retrospective single-center study, 623 patients undergoing coronary angiography were analyzed. We evaluated three missing-data strategies: indicator-based imputation, MICE-based iterative imputation, and k-nearest neighbor (KNN) imputation in combination with eight machine-learning algorithms, yielding 24 workflows. Model development used an 80/20 stratified train-test split, nested 5-fold cross-validation, randomized hyperparameter optimization, SMOTENC applied within training folds to handle class imbalance, and probability calibration. We selected the final model based primarily on mean outer cross-validated ROC-AUC, using mean outer Brier score as a tie-breaker. We derived the decision threshold from development-set out-of-fold predictions and applied it unchanged to the held-out test set. Model interpretation included SHAP analysis, permutation importance, and a surrogate decision tree. Results We selected the indicator-based imputation plus linear support vector machine (SVM) workflow, with a mean outer cross-validated ROC-AUC of 0.730 ± 0.069. On the test set, the model achieved a ROC-AUC of 0.759 (95% CI: 0.667–0.843) and a Brier score of 0.182. At the locked threshold of 0.6346, selected from development-set out-of-fold calibrated probabilities by maximizing balanced accuracy, balanced accuracy was 0.726, with a sensitivity of 75.3%, specificity of 70.0%, precision of 84.2%, and negative predictive value of 57.1%. SHAP analysis identified age, LVEF, WBC, WMA, and troponin among the variables with the greatest model-attributed importance. The surrogate decision tree reproduced the final model with 80.8% fidelity. Conclusions In this single-center cohort with substantial and heterogeneous missingness, predictive performance varied across missing-data and modeling strategies, while more complex learners did not demonstrate a discrimination advantage over a regularized linear SVM. The results support the feasibility of a calibrated, missingness-aware machine-learning framework for exploratory CAD risk stratification, but external validation in larger, multicenter populations with different case mix, disease prevalence, and missingness patterns is required before clinical deployment.
Language: English
Open Access status: green
OpenAlex ID: W7236197493
Cites: 72
Cited by:
Related to: 10