LB-3
Artificial Intelligence for Extracting Non-COronary Data from Angiography: The AI-ENCODE Study
Presenter
Mohamad Alkhouli, M.D., FSCAI, Mayo Clinic College of Medicine and Science, Rochester, MN
Mohamad Alkhouli, M.D., FSCAI1, Behrouz Rostami, PhD1, Zachi Attia, PhD1, Paul A Friedman, MD2 and Rajiv Gulati, M.D. Ph.D., FSCAI1, (1)Mayo Clinic College of Medicine and Science, Rochester, MN, (2)Mayo Clinic, Rochester, MN
Keywords: Coronary and Imaging & Physiology
Background:
Artificial intelligence (AI) presents a unique opportunity to transform cardiovascular imaging. In interventional cardiology, AI has shown exceptional performance in quantifying coronary disease and predicting fractional flow reserve. However, current AI applications in the Cath lab overlooked the potential to extract important non-coronary data from conventional angiograms. Methods:
Using a carefully labeled library of >20,000 angiograms performed at Mayo Clinic (2016-2021), we developed and validated multiple AI algorithms to extract left ventricular ejection fraction (LVEF), diastolic dysfunction (LVDD), right ventricular (RV) dysfunction, and cardiac index (CI) from 1-2 angiographic videos. The gold standard was echocardiograms performed within 30 days for the first 3 models, and simultaneous right heart Cath for the CI models.
Results:
The AI models yielded
excellent ability in predicting LVEF, LVDD, RV function and CI with an area under a receiver operating curve (AUC) of 0.87, 0.87, 0.78, and 0.74, respectively. In proof of concept experiments, adding routinely collected electrocardiographic and invasive hemodynamic data further enhanced the AI models' performance. Conclusions:
AI-ENCODE illustrated the ability of AI to broaden the diagnostic scope of coronary angiography by automatically extracting key functional and physiological data from routine angiograms. Real time access to AI-ENCODE data in the Cath lab
may enhance clinical decision-making and positively impact patient outcomes. 