Campuses:

Application of artificial intelligence to the standard 12 lead ECG to identify people with left ventricular dysfunction

Tuesday, May 1, 2018 - 1:25pm - 2:25pm
Lind 305
Background
Asymptomatic left ventricular dysfunction (ALVD) is present in 2-9% of the population, is associated with reduced longevity and is treatable when found. Inexpensive, reliable, in office screening is not available. The area under the curve (AUC) for a BNP screening blood test is 0.79 to 0.89. We hypothesized that use of artificial intelligence (AI) would enable the ECG, a ubiquitous, inexpensive test, to identify ALVD.

Methods
We trained a convolutional neural network using digitally stored 12 lead ECG and echocardiogram pairs from 44,969 patients from the Mayo Clinic data vault to identify patients with an ejection fraction (EF) less than or equal to 35%. The network was then tested on 51,979 patients reserved for external validation.

Results
Of the 51,979 patients tested, 4,064 (8%) had an EF less than 35%. The AUC of the ROC was 0.93 (Fig). The sensitivity, specificity and accuracy were 85%, 86% and 86%, respectively. In patients with an abnormal AI screen but normal EF (false positives, 1317), 153 had at least one abnormal EF in the future (5 year incidence 10.1%). This five-fold increased risk of developing a future low EF suggests that the network may be detecting early, subclinical, metabolic or structural abnormalities that manifest in the ECG.

Conclusion
Applying artificial intelligence to the ECG - a ubiquitous, low cost test - permits the ECG to serve as a powerful tool to screen for asymptomatic ventricular dysfunction and furthermore to identify individuals at increased risk for its development in the future.