PO-700-04 PREDICTING RISK OF SUDDEN CARDIAC DEATH BY ARRHYTHMIA IN HYPERTROPHIC CARDIOMYOPATHY PATIENTS BY APPLYING MACHINE LEARNING TO CARDIAC MAGNETIC RESONANCE IMAGES AND CLINICAL COVARIATES

Abstract

Hypertrophic cardiomyopathy (HCM) is a leading cause of sudden cardiac death (SCD) by ventricular arrhythmia (VA) in young adults. Current clinical criteria outlined in AHA/ACC and ESC guidelines fail to accurately identify HCM patients at risk for SCD. Mechanistically, myocardial fibrosis, a hallmark of HCM, creates substrates for VA. However, few studies have explored whether features of myocardial fibrosis visualized on late gadolinium enhanced cardiac magnetic resonance imaging (LGE-CMR) can be used, in combination with other clinical data, to forecast VA in a large cohort of HCM patients. This study aims to predict risk of VA in HCM patients by applying machine learning (ML) to LGE-CMR images and clinical covariates. In a single-center retrospective study, LGE-CMR images and clinical covariates were acquired from 819 patients with HCM to develop an algorithm for VA (tachycardia or fibrillation) risk prediction. The left ventricular (LV) myocardium was automatically segmented from LGE-CMR images by a pre-validated deep learning model which preserves the shape of the LV with anatomical fidelity. Multiple features quantifying fibrosis burden were calculated using various standard thresholding techniques. These were fed, together with most discriminative clinical covariates selected in a univariate analysis with outcome, into a tree-based ensemble learning model to identify patients who experienced VA. In the univariate analysis, 9 numerical covariates exhibited significant (p<0.05) discrimination of the outcome in Mann-Whitney U test. The top 3 covariates were maximum interventricular septal thickness, LV global longitudinal systolic strain and LV outflow tract stress. The model incorporating LGE-CMR-derived information and clinical covariates identified HCM patients with VA with area under the ROC curve (AUC) of 0.69 and 0.72, respectively, in a 5-fold cross validation (n=570) and a held-out test (n=143). As a comparison, a model we developed using only demographic covariates and major risk factors outlined in AHA/ACC guidelines had AUC of 0.63 and 0.60.View Large Image Figure ViewerDownload Hi-res image Download (PPT) Our ML algorithm offers improved and generalizable prediction of VA in HCM patients compared with current clinical guidelines. It could be used to identify HCM patients at high risk for SCD.