Deep Learning–Based Automated Echocardiographic Quantification of Left Ventricular Ejection Fraction: A Point-of-Care Solution. (16th June 2021)

• Record Type: Journal Article
• Authors: Asch, Federico M.; Mor-Avi, Victor; Rubenson, David; Goldstein, Steven; Saric, Muhamed; Mikati, Issam; Surette, Samuel; Chaudhry, Ali; Poilvert, Nicolas; Hong, Ha; Horowitz, Russ; Park, Daniel; Diaz-Gomez, Jose L.; Boesch, Brandon; Nikravan, Sara; Liu, Rachel B.; Philips, Carolyn; Thomas, James D.; Martin, Randolph P.; Lang, Roberto M.
• Is Part Of: Circulation. Volume 14:Number 6(2021)
• Page Start: e012293
• Publication Date: 2021-06-16
• Subjects: algorithm -- artificial intelligence -- echocardiography -- machine learning -- ventricular function, left; Cardiovascular system -- Imaging -- Periodicals; Heart -- Imaging -- Periodicals; 616.1075405
• DOI: 10.1161/CIRCIMAGING.120.012293
• Languages: English
• ISSNs: 1941-9651

Abstract

Abstract : Background: We have recently tested an automated machine-learning algorithm that quantifies left ventricular (LV) ejection fraction (EF) from guidelines-recommended apical views. However, in the point-of-care (POC) setting, apical 2-chamber views are often difficult to obtain, limiting the usefulness of this approach. Since most POC physicians often rely on visual assessment of apical 4-chamber and parasternal long-axis views, our algorithm was adapted to use either one of these 3 views or any combination. This study aimed to (1) test the accuracy of these automated estimates; (2) determine whether they could be used to accurately classify LV function. Methods: Reference EF was obtained using conventional biplane measurements by experienced echocardiographers. In protocol 1, we used echocardiographic images from 166 clinical examinations. Both automated and reference EF values were used to categorize LV function as hyperdynamic (EF>73%), normal (53%–73%), mildly-to-moderately (30%–52%), or severely reduced (<30%). Additionally, LV function was visually estimated for each view by 10 experienced physicians. Accuracy of the detection of reduced LV function (EF<53%) by the automated classification and physicians' interpretation was assessed against the reference classification. In protocol 2, we tested the new machine-learning algorithm in the POC setting on images acquired by nurses using a portable imaging system. Results: Protocol 1: the agreement with the reference EF values was good (intraclass correlation, 0.86–0.95), with biases <2%. Machine-learning classification of LV function showed similar accuracy to that by physicians in most views, with only 10% to 15% cases where it was less accurate. Protocol 2: the agreement with the reference values was excellent (intraclass correlation=0.84) with a minimal bias of 2.5±6.4%. Conclusions: The new machine-learning algorithm allows accurate automated evaluation of LV function from echocardiographic views commonly used in the POC setting. This approach will enable more POC personnel to accurately assess LV function.