Machine learning improves interpretation of coronary artery disease using Rb-82 PET quantification of myocardial blood flow

110 Objectives: Clinical interpretation of stress myocardial perfusion imaging (MPI) is based primarily on regional reductions in tracer uptake to identify epicardial coronary artery disease (CAD), but also includes ECG-gated image assessment of left ventricular volumes, ejection fraction, regional wall motion, transient ischemic dilatation and the patient’s clinical history. Quantification of regional myocardial blood flow (MBF) is becoming a standard component of the stress MPI clinical report using PET, demonstrating improved defect contrast and sensitivity for multi-vessel disease, and has been proposed as the primary measurement for interpretation in some centers. This study compared the abilities of machine learning (ML) and conventional statistical methods to reproduce the clinical interpretation of CAD using only rest & stress MBF data in a large cohort of scans acquired over a 10-year period. Our models predict the presence of CAD on a per patient basis. Methods: Our dataset consisted of 7,573 rest and stress PET MBF studies, and excluded patients with an abnormal scan due to calcification. The clinical standard diagnoses for detection (normal/abnormal) and localization (LAD/LCX/RCA vessel) of CAD (scar/ischemia) were extracted by regular expression parsing of the structured perfusion reports, and were manually reviewed for accuracy. PET MBF values were tabulated at two levels of detail: 3 vascular territories and 17 polar-map segments. Logistic regression (LR) and the following three ML models were trained to predict the presence of CAD: support vector machine (SVM), multilayer perceptron (MLP), and random forest classifier (RF). MBF studies were divided into training (N=6,058) and testing (N=1,515) sets, stratified by class. Pairwise comparisons of areas under the receiver operator characteristic curves (AUC) were conducted between all model pairs using DeLong’s test, and confidence intervals for AUC values were computed with bootstrapping. Accuracy was calculated with the prediction threshold at 0.5. Results: All ML models outperformed logistic regression (P < 0.0001) for detection of CAD. This was consistent both for models trained on 3 vessel data and for those trained on 17 segment data. On the 3-vessel data, SVM outperformed RF and MLP (P < 0.05 and P < 0.05), with no significant differences observed between the AUCs of RF and MLP. On the 17-segment data, MLP outperformed both SVM and RF (P < 0.0001 and P < 0.05, respectively), with no significant differences observed between the AUCs of RF and SVM. Performance was consistently higher for models trained on the 17-segment data than for those trained on the 3-vessel territory data. Conclusions: Random forest, support vector machine, and multilayer perceptron models outperformed standard logistic regression for overall detection of CAD in this large retrospective cohort. These methods have potential to be used as decision support tools to improve the diagnosis of CAD using Rb-82 PET MBF imaging. Future studies will determine these ML methods’ accuracy for localization of disease in the 3 vessel territories, discernment of type of abnormality (scar/ischemia), and identification of features associated with coronary microvascular disease.