Abstract 13853: A Fully Automated Deep Phenotyping of Late Gadolinium Enhancement Cardiac MRI for Predicting Heart Failure Progression in Hypertrophic Cardiomyopathy

Abstract

Introduction: Myocardium scar volume in late gadolinium enhancement (LGE) MRI is associated with cardiac adverse outcomes in hypertrophic cardiomyopathy (HCM). However, scar quantification is time-consuming and challenging. Also, LGE contains information about scar heterogeneity, location, and shape not readily quantified by current LGE analyses. Hypothesis: Automated deep phenotyping of LGE images can extract novel imaging markers of heart failure (HF) progression in HCM. Methods: We developed a deep convolutional neural network (CNN) (Fig 1) to extract LGE image features associated with progressive HF in HCM. We developed the CNN using data for 572 HCM patients (183 F; 51±16 years) evaluated at Tufts medical center. The CNN input was 5 LGE images and the output was the probability of developing a HF outcome (defined as a change of New York Heart Association class I/II at baseline to class III/IV during follow-up). The dataset was split into training, 50%, validation, 25%, and independent testing, 25%. We trained the CNN for 50 epochs (learn-rate = 0.001, batch size = 12, class-weighted cross-entropy loss). To avoid overfitting, the model that best predicted HF outcomes in the validation cohort was used as the final model. We compared the CNN model to a single-variable logistic regression (LR) model (L2 regularization, class-weighted loss) that predicts HF outcomes using scar volume (manually quantified by an expert reader). Results: Progressive HF occurred in 96 (17%) patients during follow-up (4±2 years). In the testing cohort (n = 138), CNN correctly predicted HF in 60% of patients with positive HF outcomes (n = 22; specificity 77%, accuracy 73%, AUC = 0.73). Using the same dataset, the LR classifier showed much lower HF prediction power with AUC = 0.54 (accuracy 38%, specificity 30%). Conclusions: Deep phenotyping of LGE allows accurate prediction of progressive HF in HCM patients and outperforms classical LR models that use scar volume as a sole predictor of HF.