An artificial intelligence prediction model for optimizing patient selection for cardiac imaging for the investigation of suspected coronary artery disease

Aims — Nearly, 40% of patients undergoing elective invasive coronary angiography (ICA) are diagnosed with non-obstructive coronary artery disease (CAD) or normal coronary anatomy, resulting in unnecessary risk exposure and increased costs to the healthcare system. In this study, we externally validate an artificial intelligence model for optimizing patient selection for ICA vs. coronary computed tomography angiography (CCTA) to reduce unnecessary ICAs.

Methods and results — The model was trained on data from outpatients undergoing elective ICA at two cardiac centres in Ontario, Canada between 2008 and 2019. It uses 42 predictors including demographic characteristics, risk factors, and medical history (including ECG stress testing and/or functional imaging) to predict the probability of obstructive CAD. Geographical validation assessed the discrimination performance on patients seen at the other 20 cardiac centres in Ontario, Canada during the same period. Temporal validation evaluated the model’s performance on outpatients receiving ICA at the original centres between 2020 and 2023. Reclassification analysis was employed to estimate health system impact. Subgroup analysis was used to assess model fairness. Following external validation, the model was updated on data from the entire outpatient population (N = 319 012) receiving ICA from 2008 to 2023. As expected, performance of the local model was lower on the geographical validation cohort (AUROC = 0.714 [95% CI: 0.714, 0.714], Sensitivity = 0.73, Specificity = 0.57) than on the local test set. However, following updating, performance recovered (AUROC = 0.794 [95% CI: 0.794, 0.795], Sensitivity = 0.708, Specificity = 0.699). Analysis of health system impact indicates that 27 of every 39 unnecessary ICAs could be avoided through use of the model.

Conclusion — Use of the model could result in an absolute reduction of 27% in the proportion of ICAs that result in a diagnosis of normal/non-obstructive disease. This could contribute to a reduction in complications from ICA and more efficient utilization of cardiac catheterization lab capacity for higher-value cardiac interventions such as revascularization and structural procedures. Additionally, use of the model would create significant efficiencies for payors, given the much lower cost of CCTA compared with ICA. If implemented within clinical practice, the model has the potential to improve the patient experience and reduce existing health inequities.