Echocardiographic Video-Driven Multi-task Learning Model for Coronary Artery Disease Diagnosis and Severity Grading Running head: Echo-Video AI for CAD Diagnosis and Severity

Ying Guo¹Yu-Han Cai²TAO XU¹Xinyang Song¹Hong-Xia Guo¹Min Dong¹Dong Ni²Fang Wang^1*Wufeng Xue^2*

• ¹Beijing Hospital, Beijing, Beijing, China
• ²School of Medicine, Shenzhen University, Shenzhen, Guangdong Province, China

Echocardiography is a first-line noninvasive test for diagnosing coronary artery disease (CAD), but it depends on time-consuming visual assessments by experts. This study constructed an echocardiographic video-driven multi-task learning model, denoted Intelligent echo for CAD (IE-CAD), to facilitate CAD screening and stenosis grading. A 3DdeeplabV3+ backbone and multi-task learning were simultaneously incorporated into the core frame of the IE-CAD model to capture the dynamic myocardial contours. Multifarious features reflecting local semantic structures were extracted and integrated to yield echocardiographic metrics such as ejection fraction, strain, and myocardial work. For model training and testing, we used a total of 870 echocardiographic videos from 290 patients with clinically suspected CAD at Beijing Hospital (Beijing, China), split at an 8:2 ratio. To evaluate the model’s generalizability, we used an external dataset comprising 450 echocardiographic videos from 150 patients at Fuwai Hospital (Beijing, China). The IE-CAD model achieved an AUC of 0.78 and a sensitivity of 0.85 for detecting significant or severe CAD, with a Pearson correlation coefficient of 0.545 for predicting the Gensini score. When applied to the external dataset, the model achieved an AUC of 0.77 and a sensitivity of 0.78 for detecting significant or severe CAD. Thus, the IE-CAD model demonstrated effective CAD diagnosis and grading in patients with clinical suspicion.