Heartperfect: Data mining in a large database of myocardial perfusion scintigraphy

Abstract

We are presenting a method to obtain diagnosis and prognosis information by searching similar images into a large database of Myocardial Perfusion Scintigraphy (MPS) cases for which diagnosis is known. We are applying similarity measures to cardiac images pre-registered with a template. Our database is composed of 1430 patient cases with associated clinical information. For each new case, we sort all the patients of the database from most to less similar ones and compute a severity criterion, based on a statistical analysis of normal and diseased most similar patients. By varying a threshold on the severity criterion and testing the classification of controlled cases, we have measured the operational characteristic of this test (ROC curves), and shown increased performance in sensitivity and specificity for disease detection with respect to clinicians and to experts in consensus. Through the extension of database to patients’ outcome information, we expect to extend this method to prognosis. © Springer-Verlag Berlin Heidelberg 2000.