Classification of left ventricular thrombi by their history of systemic embolization using pattern recognition of two-dimensional echocardiograms

Abstract Although one can diagnose left ventricular (LV) thrombi by two-dimensional echocardiography (2DE), the factors associated with peripheral embolization, given a 2DE with LV thrombi, have not been well delineated. Therefore we looked at 2DE and clinical variables that included texture features in the 2DE of 38 patients whose 2DE had LV thrombi and questioned these patients to see if clinical embolization had occurred in the 8.9 ± 6.1 month (± SD) average follow-up period. Eight patients, four with acute myocardial infarction (AMI) and four with dilated LV and decreased LV systolic wall motion, had clinically apparent leg or brain emboli, whereas the remaining patients did not. Emboli occurred within a week of obtaining the 2DE in question. The variables considered were the age of the patient, the type of heart disease present, warfarin administration, exercise tolerance, standard M-mode measurements, LV dyssynergy by 2DE, clot size and mobility, and gray scale statistics which include run length, Sobel edge points followed by 50% gradient thresholding, gray level second-order statistics, offset 1 and gray level difference statistics, offset 1. The values of the variables were then entered into an expert system (Expert Ease) in order to achieve classification of patients into emboli versus no emboli groups, while using a minimal number of variables. The only variables that were needed included run length, long runs emphasis, gray level difference statistics (entropy, contrast, mean, and angular second moment), gray level second-order statistics (contrast), and warfarin status. When probability statistics were applied to this schema, its accuracy was predicted to be at least 96%. These results suggest that a small number of 2DE texture and clinical features can be used to detect LV thrombi that are at high risk for systemic embolization.