Machine learning–based decision support for coronary angiography referral in vascular surgery patients

Background Coronary artery disease (CAD) is highly prevalent among patients undergoing vascular surgery, and identifying those who may benefit from myocardial revascularisation before surgery is crucial for perioperative risk reduction. However, appropriate referral for coronary angiography (CA) remains challenging in routine clinical practice. Machine learning (ML)-based decision support systems may improve patient selection using routinely available clinical data. Methods A retrospective database of 7464 consecutive patients referred for CA at a tertiary care centre was analysed. After data cleaning and exclusion of incomplete records, 15 intrinsic patient variables (2 numerical, 13 binary) were retained following rarity filtering, correlation analysis, and exclusion of medication-related confounders. Outcomes of CA served as the reference standard (positive vs. negative findings). Feature scaling and exhaustive feature-combination analysis were performed, generating 32,767 variable sets. Seven ML algorithms (Random Forest, CatBoost, LightGBM, SVM, Logistic Regression, Naïve Bayes, KNN) were trained using 5-fold cross-validation. Model performance was assessed using accuracy, precision, sensitivity, and F1-score with a 5% hold out test dataset. An expert baseline was established by three physicians with different experience levels, who independently evaluated a representative randomised subset of 125 patients using the same clinical variables. Results In the full cohort, physician-driven referral achieved a precision of 62.9%. In the randomised subset, physician precision remained comparable (61.6%, p=0.881), confirming representativeness. All physicians had high false-negative rates, with the most experienced achieving the best balance (F1-score 0.45; accuracy 0.50). ML models consistently outperformed physicians. The best-performing models (LightGBM, Random Forest, CatBoost) achieved F1-scores 0.726, 0.69 and 0.692 respectively, with sensitivities up to 0.86 evaluated on the hold out test dataset. Aneurismatic disease and left ventricular ejection fraction emerged as the most influential predictors across models. Conclusion ML-based decision support significantly outperforms physician judgment in identifying patients with positive coronary angiography findings, primarily by reducing false-negative referrals. Integration of ML tools using simple clinical variables may improve patient safety and optimise invasive diagnostic strategies in vascular surgery patients.