A predictive model based on non-contrast CMR and texture analysis for the prognostic evaluation of cardiac amyloidosis

Abstract

Abstract We aimed to develop a predictive model based on texture features with non-contrast cardiac magnetic resonance (CMR) imaging for risk stratification toward adverse events in cardiac amyloidosis (CA) patients. A cohort of 78 CA patients was classified into a training set (n = 54) and a validation set (n = 24) at a ratio of 7:3. A total of 275 texture features were extracted from CMR images. MaZda and the support vector machine (SVM) were utilized for feature selection and model construction. A SVM model incorporating radiological and texture features was built for prediction of endpoint events by evaluating area under curve (AUC). In the whole cohort, 52 patients were major adverse cardiovascular events (MACE) occurred and 26 patients were No MACE occurred. By combining 2 radiological features and 8 texture features extracted from cine and T2-weighted imaging (T2WI) images, the SVM model achieved AUCs of receiver operating characteristic (ROC) and precision-recall (PR) of 0.930 and 0.962 in the training cohort and that of 0.867 and 0.941 in the validated cohort. The Kaplan–Meier curve of this SVM model criteria excellently stratified CA outcomes (Log rank test, P < 0.0001). The SVM model based on radiological and texture features derived from non-contrast CMR images can be a reliable biomarker for adverse events prognostication in CA patients.