Artificial Intelligence Analysis of ECG to Determine Fractional Flow Reserve (FFR)

BackgroundThe current gold standard of coronary artery disease (CAD) diagnosis is invasive angiography, during which fractional flow reserve (FFR) measurement may be performed to confirm the clinical significance of a stenosis. The yield of routine and indiscriminate FFR in identifying hemodynamically significant stenoses is low. To combat this, we have developed an artificial intelligence model - ECGio - designed to be deployed at the point of care to determine FFR through the analysis of a resting digital 12-lead electrocardiogram (ECG), a fast, real-time, cost-effective, widely accessible, and safe diagnostic method. This study assessed the ability of ECGio to train, tune, and test itself through a cross-validation paradigm to predict the presence of a reduced FFR in the left anterior descending artery in a patient population presenting for invasive FFR. MethodsIn a single-center study the ECGs of 209 consecutive patients (61.3{+/-}9.5 years, 35.4% female) from 2014 to 2021 were recorded within 7 days prior to angiography during which FFR was measured in the left anterior descending artery. Collected ECGs were used to train and test the AI model using a five-fold cross-validation methodology. ResultsThe ability of ECGio to predict the presence of a reduced FFR (<0.80) in this cohort was a sensitivity, specificity, PPV, NPV, Accuracy, and F-1 Score of 43.2%, 86.7%, 64.0%, 73.6%, 71.3%, and 51.6%, respectively. ConclusionsThis study demonstrated the feasibility of using a deep learning AI algorithm to analyze a digital 12-lead ECG to provide a similar level of information as the invasive FFR. HighlightsO_LICoronary angiography is invasive and expensive and exposes the patient to radioactive dyes and risk of complications. Clinicians tend to fail-safe, overperforming testing and struggling to identify patients who would benefit from invasive testing resulting in procedures having low yield. C_LIO_LIOur AI model determines FFR by analyzing the patients resting digital 12 lead ECG which is fast, cheap, safe, and real-time. C_LIO_LIAI ECG analysis has the potential to play a crucial role in CAD diagnostics. C_LI PerspectivesA major obstacle in CAD screening is that there is no quick, accurate, non-invasive test to differentiate patients that require additional testing and treatment from those that can be safely dismissed. The 12-lead digital ECG is the most easily acquired diagnostic test; it does not involve stress, radioactive dyes, or risk. Invasive FFR is the most accurate technique to identify an ischemia-producing stenosis. The current study demonstrated the feasibility of training AI to analyze ECG signals to recognize reduced FFR and even estimate the actual FFR value. Future studies will analyze patients who enter the diagnostic process through other clinical pathways in order to better understand the performance of ECGio in a more general patient population. EthicsThis study was conducted in a deidentified, retrospective fashion from a pre-existing registry. The ethics committee waved the need for informed consent.