Abstract 11482: A Neural Network-Based Model Detects Low Ejection Fraction from Electrocardiograms Without Overt Rhythm or Conduction-System Abnormalities

Abstract

Introduction: While detection of asymptomatic left ventricular dysfunction (LVD) is important given the effectiveness of preventive medical treatments, the gold standard method using echocardiograms requires trained experts and thus is less suitable for widespread screening. Previous studies have found that convolutional neural network (CNN) models could detect patients with LVD from 12-lead electrocardiogram (ECG). However, they were developed in an unselected population which included an unknown number of patients with overt rhythm and conduction system abnormalities, which frequently coexist with LVD. It is unclear to what extent these models relied on such obvious features to detect LVD from ECGs. Herein, we aimed to assess the ability of a CNN-based model to detect LVD using ECGs having no apparent rhythm or conduction abnormalities. Methods: 56,985 ECGs with an echocardiogram within a 2-week interval were identified from the BWH ECG database. The dataset was split into 28,465, 11,306 and 17,214 ECGs for derivation, validation and test, respectively, with no overlap of patients across groups. A CNN model was trained as a binary classifier to detect LVD, defined as LV ejection fraction < 35 %. The final model was selected as that with the best area under the receiver operating characteristics curve (AUROC) on the validation set and performance assessed on the test set. The model was further tested after excluding ECGs with overt rhythm and conduction system abnormalities defined as the presence of atrial fibrillation, premature atrial complexes, premature ventricular complexes, electronically paced rhythms, left or right bundle branch block. Results: The mean patient age in the test set was 69.8 years, and 57.3% of patients were male. prevalence of low LVEF was 9.7%. The test dataset contained 7,633 ECGs with overt rhythm and conduction system abnormalities. The model yielded good accuracy with an AUROC of 0.91 (95% CI:0.90-0.91) to identify patients with LVD in the test data set. The model performed similarly in the presence or absence of rhythm or conduction-system abnormalities (AUROC 0.90 95% CI:0.89-0.92 and 0.89 95% CI:0.88-0.90 respectively). Conclusion: The model detected low LVEF from ECG without overt rhythm and conduction system abnormalities.