Aims

Models predicting the likelihood of obstructive coronary artery disease (CAD) on invasive coronary angiography exist. However, as stable patients with new-onset chest pain frequently have lower clinical likelihood and preferably undergo index testing by non-invasive tests such as coronary computed tomography angiography (CCTA), clinical likelihood models calibrated against observed obstructive CAD at CCTA are warranted. The aim was to develop CCTA-calibrated risk-factor- and coronary artery calcium score-weighted clinical likelihood models (i.e. RF-CL_CCTA and CACS-CL_CCTA models, respectively).

Methods and results

Based on age, sex, symptoms, and cardiovascular risk factors, an advanced machine learning algorithm utilized a training cohort (n = 38 269) of symptomatic outpatients with suspected obstructive CAD to develop both a RF-CL_CCTA model and a CACS-CL_CCTA model to predict observed obstructive CAD on CCTA. The models were validated in several cohorts (n = 28 340) and compared with a currently endorsed basic pre-test probability (Basic PTP) model. For both the training and pooled validation cohorts, observed obstructive CAD at CCTA was defined as >50% diameter stenosis. Observed obstructive CAD at CCTA was present in 6443 (22.7%) patients in the pooled validation cohort. While the Basic PTP underestimated the prevalence of observed obstructive CAD at CCTA, the RF-CL_CCTA and CACS-CL_CCTA models showed superior calibration. Compared with the Basic PTP model, the RF-CL_CCTA and CACS-CL_CCTA models showed superior discrimination (area under the receiver operating curves 0.71 [95% confidence interval (CI) 0.70-0.72] vs. 0.74 (95% CI 0.73-0.75) and 0.87 (95% CI 0.86-0.87), P < 0.001 for both comparisons).

Conclusion

CCTA-calibrated clinical likelihood models improve calibration and discrimination of observed obstructive CAD at CCTA.