Evaluating Deep Learning Performance in Diagnosing Atrioventricular Valve Regurgitations Using Echocardiography:An External Validation Study - The 72nd Annual Conference of the Israel Heart Society in association with the Israel Society of Cardiothoracic Surgery

Ido Cohen ^1,2,3 Jeffrey G Malins ⁴ Michal Cohen-Shelly ^2,5,6 Yossi Asaf ^1,2,3 Michael Fiman ⁵ Kobi Faierstein ^1,2,3 Karin Sudri ⁶ Robert Klempfner ^2,3,5 Elad Maor ^2,3,5

¹Internal Medicine E, Sheba Medical Center, Ramat-Gan, Israel
²Leviev Cardiovascular Institute, Sheba Medical Center, Ramat Gan, Israel
³Faculty of Medical and Health Sciences, Tel Aviv University, Tel Aviv, Israel
⁴Department of Cardiovascular Medicine, Mayo Clinic, 200 First Street SW, Rochester, MN 55905, USA
⁵Aisap.ai, Ramat Gan, Israel
⁶ARC Innovation Center, Sagol AI Hub, Sheba Medical Center, Ramat Gan, Israel

Background: Mitral regurgitation (MR) and tricuspid regurgitation (TR) are prevalent valvular heart diseases associated with significant morbidity and mortality. Timely diagnosis and referral can enable effective treatment interventions that markedly improve patient outcomes. However, traditional echocardiography faces limitations in availability, cost, consistency, and reliability, leading to misdiagnosis and undertreatment. Artificial intelligence (AI) applied to echocardiographic scans has the potential to address these challenges.

Methods: This study evaluates the AI algorithm developed by AISAP on an external population. It utilizes deep learning networks to analyze echocardiographic exams for diagnosing atrioventricular valve disorders. We tested the algorithm on transthoracic echocardiography data from the Mayo Clinic Health System (MCHS) between 2013 and 2023. The model`s performance was compared to ground truth values using two classification schemes: distinguishing between normal-mild and moderate-severe regurgitation, and categorizing results into four groups: normal, mild, moderate, and severe.

Results: The MR cohort included 280 patients, whereas the TR cohort comprised 298 patients. The model demonstrated a robust ability to identify clinically significant (moderate and above) atrioventricular valve regurgitation. The MR model achieved an area under the curve (AUC) of 0.98 (95% CI: 0.97–0.99), with 91% accuracy, 95% sensitivity, and 89% specificity. In comparison, the TR model exhibited an AUC of 0.96 (95% CI: 0.94–0.98), with 84% accuracy, 91% sensitivity, and 80% specificity.

Conclusion: The model demonstrated high diagnostic accuracy and reliability in assessing atrioventricular valve regurgitation severity across diverse populations, highlighting its potential as a valuable clinical tool. The findings underscore the role of AI in complementing expert evaluations and improving access to effective diagnostics, with future applications potentially including point-of-care diagnosis and monitoring of disease progression.