Validation of a Deep Learning-enabled Electrocardiogram Algorithm to Detect and Predict Cardiac Contractile Dysfunction in the Community

Background: Left ventricular systolic dysfunction (LVSD) is associated with increased morbidity, mortality and clinical heart failure but is underdiagnosed. We sought to validate a deep learning algorithm designed to detect low left ventricular ejection fraction (LVEF), using a 12-lead electrocardiogram (ECG). 
Methods: In a prospective, randomly-selected cohort, residents aged 45 years or older underwent clinical evaluation, medical record abstraction, ECG, baseline echocardiography and passive follow-up via record linkage of routine echocardiograms, those with history of heart failure were excluded. The primary outcome was the ECG algorithm’s ability to identify subjects with a low LVEF at baseline (i.e. ≤50%, ≤40% and ≤35%). The secondary outcome was to assess the association between a positive ECG screen and a clinically meaningful decrease in LVEF (LVEF of ≤35% or absolute change of ≤ 20%) over time, among those with normal LVEF (≥50%) at baseline. Kaplan-Meier curves and Cox proportional hazard models adjusted for age, sex and traditional risk factors. 
Results: We included 1737 subjects (52% women), mean age±SD 61.70±10.4 years. The prevalence of low LVEF at baseline was 5% for ≤50%, 1.7% for ≤40% and 1.0% for ≤35%, respectively. The network yielded excellent receiver operator characteristics’, area under the curve, sensitivity, and specificity values for the detection of low LVEF at baseline. Among those with LVEF≥50% (n=1639), after a median follow-up of 18.5±4.4 years, 108(6.6%) had a clinically meaningful drop in LVEF. Those with a positive ECG screen but normal LVEF were at higher risk of suffering a clinically meaningful drop in EF HR=2.49 (95%CI 1.57-3.96, p<.0001) when compared to those with a negative ECG screen. The risk increased when assessing shorter follow-up intervals, HR=3.24 (95%CI 1.91-5.48, p<0.0001) at 15 years, HR=3.78 (95%CI1.97-7.28, p<0.0001) at 10 years and HR=7.93 (95%CI 2.20-27.42, p=0.001) at 5 years. These results were only slightly attenuated after adjusting for age, sex, body mass index, history of hypertension, diabetes and coronary disease.
Conclusions: A deep learning-enabled ECG algorithm detected low LVEF with excellent accuracy in an unselected community population and predicted future LVSD. Application of deep learning to the ECG can serve as a powerful screening tool to identify LVSD.

02 - Validation of a Deep Learning-enabled Electrocardiogram Screening Algorithm to Detect and Predict Cardiac