Multisite, External Validation of an AI-Enabled ECG Algorithm for Detection of Low Ejection Fraction

TL;DR

This study validates an AI-based ECG algorithm for detecting low ejection fraction across multiple U.S. sites, showing strong accuracy and potential for clinical use.

Contribution

The study provides a large-scale, multisite external validation of an AI-enabled ECG algorithm for low ejection fraction detection.

Findings

The AI algorithm demonstrated an AUROC of 0.92 for detecting low ejection fraction.

The algorithm had 84.5% sensitivity and 83.6% specificity for LEF detection.

The algorithm produced a negative result in 78% of cases, suggesting potential as a rule-out strategy.

Abstract

Low left ventricular ejection fraction (LEF) can progress undiagnosed. Artificial intelligence–based electrocardiogram (ECG-AI) screening may provide a scalable means to detect LEF.

The purpose of this study was to validate a complete ECG-AI software as a medical device for LEF detection.

Four geographically diverse sites in the United States identified patients with both ECGs and transthoracic echocardiograms performed within 30 days of each other in clinical practice. Data were electronically extracted to specific guidelines and transmitted to the coordinating center for analysis.

Records of 16,000 subjects were extracted, resulting in an evaluable set of 13,960 subjects (mean age 66 years; 52% male). The device demonstrated excellent discrimination (AUROC: 0.92 [95% CI: 0.91-0.93]) and was 84.5% (95% CI: 82.2%-86.6%) sensitive and 83.6% (95% CI: 82.9%-84.2%) specific for LEF. The overall prevalence of LEF in the study data set was 7.9%, with LEF among 1.6% of the ECG-AI negative and 30.5% of ECG-AI positive subjects, contributing to positive and negative predictive values of 30.5% (95% CI: 28.8%-32.1%) and 98.4% (95% CI: 98.2%-98.7%), respectively.

External validation studies such as this one provide a rigorous framework to validate an algorithm’s performance. This study demonstrated the algorithm’s strong diagnostic accuracy over a geographically diverse, independent set of patients. In this generally unselected population, the algorithm produced a test negative result in 78% of the cases, suggesting potential utility as a rule-out strategy to defer echocardiography when other clinical findings are absent.

Figures