Abstract 13795: Deep Learning of Electrocardiogram Waveforms for Diagnosis of Heart Failure With Reduced Left Ventricular Ejection Fraction

Abstract

Introduction: A deep learning-aided electrocardiogram (Deep-ECG) was trained to identify left ventricular systolic dysfunction (LVSD) of ejection fraction (EF) less than 40%. We evaluated the performance of Deep-ECG among patients with symptomatic heart failure (HF) in Korean acute HF registry. Hypothesis: Deep-ECG may be useful for identifying EF in HF with reduced EF (HFrEF). Methods: We included 1,292 symptomatic HF patients from 2011 to 2014 who underwent echocardiography and ECG within 30 days. The performance of Deep-ECG for LVSD was determined using receiver-operator characteristic curve (AUC) values. Results: Mean age was 67.8±14.4 years and 56% of patients were men. HFrEF, HF with mid-range EF and HF with preserved EF(HFpEF) was present in 42.3%, 17.9% and 39.8%, respectively. Using the optimal cut-off by Youden’s index, AUC value was 0.844 for identifying HFrEF among HF patients. AUC values were significantly higher for patients with normal AV conduction (0.872 vs. 0.759, p =0.002) and normal interventricular conduction (0.848 vs. 0.812, p=0.037) when compared to each conduction delay. Interestingly, corrective (true-positive) results in Deep-ECG for LVSD were acquired in patients bigger LV end systolic dimension (LVESD) than in false negative group (54.4±10.1mm vs. 49.7±8.7mm, p<0.001). Conclusions: The Deep-ECG algorithm had very good discrimination for HFrEF in real world HF cohort. This algorithm may be useful for identification of LVSD in resource-limited settings. Figure 1. Scatterplot demonstrating LVESD and the performance of Deep-ECG for identifying LVSD (LVEF < 40%)