Using a Machine Learning Algorithm to Detect Depressed Ejection Fraction From a Single-Lead ECG
Abstract
Background: Multiple studies demonstrate the benefit of intervention for left ventricular ejection fraction (LVEF) below 40%, so the development of a low ejection fraction algorithm to detect LVEF below 40% can aid in early screening of initial asymptomatic Heart Failure with reduced Ejection Fraction (HFrEF).
Objective: To demonstrate the performance of a low ejection fraction algorithm using single-lead ECG data to detect LVEF below 40%.
Methods: We collected 1325 single-lead ECG recordings (15s duration) at various chest positions using an electronic stethoscope from 197 patients. We analyzed these ECG recordings using a deep neural network model trained on individual leads extracted from a 12-lead ECG to discriminate left ventricular ejection fractions (EFs) above or below different thresholds. We compared the model output to ejection fraction measured using echocardiograms.
Results: Across all recordings from all patients, we obtained an AUROC of 0.89, with a sensitivity of 88% and specificity of 74% using a model output threshold of 0.35 (Figure 1). The AUROC of recordings taken at different orientations and stances ranged from 0.85 to 0.92 (Table 1), with a sensitivity of at least 78% and specificity of at least 66% at any orientation.
Conclusion: Using a single lead ECG measured by an electronic stethoscope and a deep neural network model, we were able to detect depressed ejection fraction (≤40%) with a sensitivity of 88% and specificity of 74%. This work demonstrates the utility of a low-cost electronic stethoscope and machine learning for early screening and detection of depressed left ventricular ejection fraction.
