False-negative constrained machine learning for coronary angiography referral in vascular patients
J Petrovic, N Ruzic, B Vucurevic, S Graovac, I Burazor, M Bojic, S Tanaskovic, B BozicAbstract
Background
Coronary artery disease (CAD) is highly prevalent among patients undergoing vascular surgery. False-negative referral decisions for coronary angiography (CA) expose high-risk vascular patients to potentially preventable adverse cardiac events. Conventional clinical judgment is associated with substantial false-negative rates. Machine learning (ML) models optimised to constrain false negatives may improve patient safety while maintaining acceptable overall performance.
Methods
A retrospective cohort of 7464 vascular patients referred for CA at a tertiary care centre was analysed, with angiographic findings serving as the reference standard. After preprocessing, feature selection, and standardisation, ML models were trained for binary classification (positive vs. negative CA findings). To prioritise patient safety, models were optimised using decision thresholds explicitly constrained to maintain a false-negative rate (FNR) ≤10% (α_FNR=0.1). Random Forest and CatBoost classifiers were identified as best-performing architectures. Model performance was assessed using accuracy, precision, recall, F1-score, and confusion matrix analysis. Results were compared to physician-based decision-making reported in prior analyses.
Results
A Random Forest model using four routinely available clinical variables (hyperlipidemia, prior myocardial infarction, prior PCI, and ejection fraction) achieved an FNR of 6.9%, with accuracy 68.2%, precision 69.3%, recall 93%, and F1-score 0.64. A CatBoost model incorporating a broader clinical feature set (age, diabetes, smoking status, hypertension, hyperlipidemia, sex, atrial fibrillation, prior myocardial infarction, prior PCI, angina pectoris, ejection fraction, and aneurysmatic disease) achieved an FNR of 10.0%, accuracy 71.1%, precision 71.4%, recall 90%, and F1-score 0.69. Both ML approaches substantially reduced false-negative referrals compared with physician decision-making, while maintaining balanced overall classification performance. High recall values indicate effective identification of patients with true positive angiographic findings.
Conclusion
ML models optimised with false-negative constraints can reduce missed coronary disease to below 10% using simple clinical variables. Such safety-oriented ML decision support systems have the potential to substantially improve coronary angiography referral strategies in vascular patients, minimising underdiagnosis while preserving acceptable diagnostic efficiency.