Automatic Detection of Cardiac Remodeling Using Global and Local Clinical Measures and Random Forest Classification

Abstract

Myocardial infarction leads to a change in geometry and a modified motion characteristics of the heart, called remodeling. The detection of patients with subclinical remodeling is clinically relevant because effective therapies have to be initiated early to avoid a progressive dilatation, and deterioration in contractile function. In this paper, we propose a classification approach to detect patients with cardiac remodeling based on established global and local clinical parameters, like end-diastolic and end-systolic volume, ejection fraction or local myocardial thickness. The functional parameters are extracted based on segmented endo- and epicardial contours using an in-house developed software tool. A random decision forest is trained for recognition of patients with impaired shape or motion characteristics. The 17 segment model of the left ventricle proposed by the American Heart Association is compared to a higher resolution model using 97 left ventricle segments in terms of classification performance. The classification results are submitted to the left ventricle statistical shape modelling challenge with the aim to compare the classification performance of classical clinical parameters with other probabilistic or model-based approaches. A leave-one-out cross-validation shows an accuracy of 0.93 using global and local parameters compared to an accuracy of 0.86 using global parameters only.