Automated and Explainable Coronary Angiogram Interpretation for Selection of Coronary Artery Bypass Grafting Candidates using Artificial Intelligence.

Visual estimation of coronary artery stenosis on angiography is subject to human error. While machine learning may facilitate more accurate interpretation, clinical utility has been limited by lack of human interpretable models. We developed an automated computer vision model to identify candidates for coronary artery bypass (CABG) from coronary angiograms. Medical records for primary CABG between 2018 and 2023 were screened for coronary angiogram video with angiographic and operative reports. A clinically determined reference group of angiographically normal, single and double-vessel disease was compared with automated angiogram reports to identify patients with indications for CABG per AHA/ACC 2021 guidelines. A total of 4,472 angiographic video clips from 349 patients were analyzed identifying 682 lesions. Mean analysis was 51.1 seconds/study, or 4.6 seconds/video clip. Detection algorithm results were compared to original reports/images in cases where the model differed in recommendation. The model detected stenotic lesions in the left main, left anterior descending, circumflex, and right coronary arteries and calculated whether multivessel disease met criteria for CABG (Accuracy:74%; Positive Predicted Value:61%; Negative Predictive Value:87%) compared with lesions documented in angiogram reports. Incorrect angiographically normal prediction occurred in 20 cases (6%) due to the selection of an incorrect maximum contrast frame. Prediction of percutaneous intervention when bypass was recommended (n=31) was due to under-recognition of a left main or circumflex lesion. With future improvements, automated identification of CABG candidates could augment visual reading at the time of angiography to improve quality control and guideline directed revascularization strategies.