Interethnic Validation of an ECG Image Analysis Software for Detecting Left Ventricular Dysfunction in Emergency Department Population

Abstract

Background: We previously developed and validated an AI-based ECG analysis tool (ECG Buddy) in a Korean population. This study aims to validate its performance in a U.S. population, specifically assessing its LV Dysfunction Score and LVEF-ECG feature for predicting LVEF <40%, using NT-ProBNP as a comparator. Methods: We identified emergency department (ED) visits from the MIMIC-IV dataset with information on LVEF <40% or ≥40%, along with matched 12-lead ECG data recorded within 48 hours of the ED visit. The performance of ECG Buddy's LV Dysfunction Score and LVEF-ECG feature was compared with NT-ProBNP using Receiver Operating Characteristic - Area Under the Curve (ROC-AUC) analysis. Results: A total of 22,599 ED visits were analyzed. The LV Dysfunction Score had an AUC of 0.905 (95% CI: 0.899 - 0.910), with a sensitivity of 85.4% and specificity of 80.8%. The LVEF-ECG feature had an AUC of 0.908 (95% CI: 0.902 - 0.913), sensitivity 83.5%, and specificity 83.0%. NT-ProBNP had an AUC of 0.740 (95% CI: 0.727 - 0.752), with a sensitivity of 74.8% and specificity of 62.0%. The ECG-based predictors demonstrated superior diagnostic performance compared to NT-ProBNP (all p<0.001). In the Sinus Rhythm subgroup, the LV Dysfunction Score achieved an AUC of 0.913, and LVEF-ECG had an AUC of 0.917, both outperforming NT-ProBNP (0.748, 95% CI: 0.732 - 0.763, all p<0.001). Conclusion: ECG Buddy demonstrated superior accuracy compared to NT-ProBNP in predicting LV systolic dysfunction, validating its utility in a U.S. ED population.