A proof-of-concept machine-learning model improves pre-test probability estimation for obstructive coronary artery disease: benchmarking against the ESC guidelines

Background Pre-test probability (PTP) tables remain central to referral pathways for suspected chronic coronary syndromes. We developed an interpretable, baseline-only machine-learning (ML) PTP model for obstructive coronary artery disease (CAD). Purpose To internally validate the machine-learning model and benchmark it against ESC 2013 and ESC 2024 PTP models in order to improve discrimination, calibration and clinical utility of PTP estimation. Methods Single-centre study using prospectively collected data of 187 patients referred for diagnostic coronary angiography for suspected CAD. Predictors were restricted to pre-angiography variables: age, sex, symptom points and cardiovascular risk factors (count and individual factors). Outcomes were ≥50% stenosis endpoint (y50) and ≥70% stenosis endpoint (y70) at invasive angiography. Penalized logistic regression strategies (ridge/elastic-net) were internally validated using 2000 stratified bootstrap optimism correction. ESC 2013 and ESC 2024 probabilities were analyzed as fixed comparators with 2000-bootstrap AUC confidence intervals. Decision-curve analysis (DCA) quantified net benefit and interventions avoided per 100 patients versus ESC 2024 and treat-all at 10-20% thresholds. The pipeline was fully reproducible (fixed seed) and produced explicit probability equations for bedside use. Results The primary analytic cohort comprised 187 patients (66 y50 and 54 y70 events). For y50, ESC 2013 AUC was 0.750 (95% CI 0.668-0.822) and ESC 2024 AUC 0.744 (0.657-0.818). The internally validated ML model (minimal predictors with combined symptom points) achieved optimism-corrected AUC 0.771 (0.711-0.832), Brier 0.186 (0.164-0.219), calibration intercept 0.030 (-0.182 to 0.244) and slope 1.045 (0.749-1.409), improving discrimination versus ESC 2024 (delta AUC 0.061; p=0.043). DCA favoured ML, avoiding 21.4/18.4/21.9 additional interventions per 100 versus ESC 2024 at 10%/15%/20% thresholds (and 11.8/16.9/19.3 per 100 versus treat-all). For y70, ESC 2013 AUC was 0.748 (0.660-0.825) and ESC 2024 AUC 0.755 (0.671-0.834). The ML model achieved optimism-corrected AUC 0.754 (0.687-0.825), Brier 0.176 (0.156-0.206), intercept 0.015 (-0.241 to 0.340) and slope 1.006 (0.703-1.422) with no significant discrimination difference versus ESC 2024 (delta AUC 0.033; p=0.370). DCA was non-inferior, avoiding 13.4/0.2/5.3 interventions per 100 versus ESC 2024 at 10%/15%/20% thresholds (and 9.1/16.6/21.4 per 100 versus treat-all). Conclusion A true pre-test ML model using only age, sex, symptoms, and risk-factor burden improved prediction of ≥50% obstructive CAD over ESC tables while preserving calibration and clinical utility, and was non-inferior for ≥70% stenosis. These findings support data-driven refinement of pre-test risk estimation and justify external validation in larger multicentre registries and prospective impact evaluation before clinical deployment. Reporting followed TRIPOD-AI principles.