Diagnostic Performance of Artificial Neural Network for Detecting Ischemia in Myocardial Perfusion Imaging

Background: The purpose of this study was to apply an artificial neural network (ANN) in patients with coronary artery disease (CAD) and to characterize its diagnostic ability compared with conventional visual and quantitative methods in myocardial perfusion imaging (MPI). Methods and Results: A total of 106 patients with CAD were studied with MPI, including multiple vessel disease (49%), history of myocardial infarction (27%) and coronary intervention (30%). The ANN detected abnormal areas with a probability of stress defect and ischemia. The consensus diagnosis based on expert interpretation and coronary stenosis was used as the gold standard. The left ventricular ANN value was higher in the stress-defect group than in the no-defect group (0.92 +/- 0.11 vs. 0.25 +/- 0.32, P<0.0001) and higher in the ischemia group than in the noischemia group (0.70 +/- 0.40 vs. 0.004 +/- 0.032, P<0.0001). Receiver-operating characteristics curve analysis showed comparable diagnostic accuracy between ANN and the scoring methods (0.971 vs. 0.980 for stress defect, and 0.882 vs. 0.937 for ischemia, both P=NS). The relationship between the ANN and defect scores was non-linear, with the ANN rapidly increased in ranges of summed stress score of 2-7 and summed defect score of 2-4. Conclusions: Although the diagnostic ability of ANN was similar to that of conventional scoring methods, the ANN could provide a different viewpoint for judging abnormality, and thus is a promising method for evaluating abnormality in MPI.