Abstract 17893: Deep Learning for the Prediction of Myocardial Scarring Through Electrocardiograms

Abstract

Introduction: Cardiac Magnetic Resonance Imaging (MRI) with Late Gadolinium enhancement (LGE), the gold standard for myocardial scarring evaluation, is limited by cost, time, and expertise required. The Electrocardiogram (ECG), in contrast, is a widely available and affordable diagnostic tool. We applied deep learning (DL) models to ECGs to directly predict the presence of LGE on MRI. Methods: Patients who underwent a cardiac MRI and had at least one ECG collected from 2012 to 2022 were considered. Patients were classified into positive and negative for myocardial scarring via their MRI reports, labeled by experts. Patients with non-specific scarring were excluded from the dataset. Patient ECGs most recent to their MRI were utilized as input for the model. MRI reports and ECG waveforms were collected from NYU Langone Health. We implemented a 34-layer Neural Network architecture, adapted from a state-of-the-art arrhythmia detection model, to learn from a 10 second ECG sampled at 250hz. We evaluated the model using area under the receiver-operating curve (AUROC), specificity, sensitivity, and precision. Results: 8813 patients with MRI and ECG were included. 1559 were excluded due to non-specific scarring. 2739 patients were positive for scarring. Patients were split into a 70% train, 10% validation, and 20% test dataset. The deep learning model obtained a test set AUROC of 0.80 [95%CI, 0.77 - 0.81]. At 20% false positive rate, we obtained a sensitivity of 0.63, specificity of 0.80, and precision of 0.64. Conclusions: This study suggests that our DL model can predict for the presence of LGE from a surface ECG. A future direction of this work would be to evaluate the use of DL for the prediction of LGE location, quantity and specific phenotypes of myocardial disease.