The incremental value of PET perfusion imaging over coronary CT angiography in predicting outcome. A study using machine learning

Abstract

Abstract Funding Acknowledgements Type of funding sources: Public Institution(s). Main funding source(s): Academy of Finland, Turku University Hospital Background Coronary computed tomography angiography (CTA) and myocardial perfusion imaging (MPI) are powerful non-invasive tools to evaluate the patients with suspected coronary artery disease (CAD). Purpose The goal of this study was to evaluate the incremental value of these imaging methods in predicting short- and long-term cardiac events using machine learning (ML) approaches. Methods 2411 patients with clinically suspected CAD underwent coronary CTA and subsequent positron emission tomography (PET) MPI. Following our local routine, if obstructive CAD cannot be ruled-out by coronary CTA, PET MPI using 15O-water is performed. The incremental prognostic value of PET imaging was tested using several ML models of which XGBoost models consistently outperformed others and were chosen for further analyses. XGBoost models were trained by using clinical data from medical records and coronary imaging findings: one set of models used all the available variables, while a second set of models used only clinical and CTA-based variables. Differences in the performances of the models were then used to assess the incremental value of PET perfusion variables in outcome prediction. Results After the removal of incomplete data entries, data from 2284 patients was retained for further analysis. During 8-year follow-up period 210 patients had a major cardiac event (these endpoints correspond to 59 myocardial infarctions, 35 unstable angina pectoris, and 116 deaths). The PET perfusion imaging data improved the predictive power of CTA during the first 4 years of observation time. After that, no significant difference in the predictive power was observed between the considered sets of XGBoost models, which either included or did not include PET perfusion data in the input variables. The highest area under the receiver operating characteristic curve (AUC) was at the observation time of 2 years (0.81, 95% CI 0.805–0.823) when PET data were included. The corresponding AUC when PET data were not included was at the observation time of 2 years (0.79, 95% CI 0.785–0.802). Conclusions Based on ML approach PET perfusion imaging improves the power in predicting cardiac events over anatomical CTA imaging for the first 4 years. The results illustrate the differences and complementary nature of anatomic and perfusion information in predicting outcome of patients with suspected CAD.