This abstract is published as part of the proceedings of the American Society Echocardiography Conference 2021.
Title: AI-Enabled LV Functional Assessment Reduced Operator Variability and Increased Statistical Power for Outcome Detection: World Alliance Societies of Echocardiography (WASE)-COVID Study
Authors: Tine Descamps, Gary M Woodward, Louise A Tetlow, Angela Mumith, Katie Ions, Ilya Karagodin, Cristiane Singulane, Mingxing Xie MD PhD FASE, Edwin S Tucay D FASE, Ana C Tude Rodrigues MD, Zuilma Y Vazquez-Ortiz MD PhD, Azin Alizadehasl MD FASE, Mark J. Monaghan PhD, Bayardo A Ordonez Salazar MD, Laurie Dufour MD, Atoosa Mostafavi MD, Antonella Moreo MD, Rodolfo Citro MD, Akhil Narang MD, Chun Wu MD PhD, Karima Addetia MD , Roberto M Lang, Federico M Asch.
BACKGROUND: The variability present with manual chamber quantitation of echocardiograms reduces accuracy and statistical power to predict patient outcomes. The ability to precisely assess cardiac function is crucial for both clinical diagnostics and clinical trial validity. Utilization of artificial intelligence (AI) to automate quantification in echocardiograms is thought to overcome these barriers. To demonstrate this, data from the WASE-COVID study was used to determine the effect of variability on clinical outcomes prediction and statistical powering.
METHODS: We studied 870 adult patients with acute COVID-19 infection from 13 medical centers in four world regions who had undergone transthoracic echocardiograms (TTEs). Biplane left ventricular (LV) ejection fraction (EF), volumes and global longitudinal strain (LVGLS) were calculated. All datasets were manually measured twice by the same operator and once by a second operator, as well as processed using an automated AI platform (EchoGo Core, Ultromics, UK). Operator influence upon variability in LVEF and LVGLS in manual and AI measurements was calculated using principal component analysis (PCA). Inter- and intra-operator variability, and inter-method variability was assessed by Bland-Altman plots. Trial outcomes detection (in-hospital mortality) was assessed by univariate and multivariate statistics. Receiver operator characteristic (ROC) curves were used to assess test performance.
RESULTS: PCA demonstrated that manual LV quantification was confounded by the operator, while AI-derived results were operator independent. In-hospital mortality was 22.2%. Assessment of clinical impact showed LVEF and LVGLS from manual measurements was associated with COVID-19 death on the 10% level (OR [95%CI] = 0.988 [0.974, 1.002], p= 0.082; OR [95% CI] = 1.027 [0.995, 1.060] p= 0.10 respectively). EF and GLS from AI measurements were associated with COVID-19 death on the 1% significance level (OR [95% CI = 0.974 [0.955, 0.993], p= 0.006; OR [95% CI] = 1.090 [1.043, 1.141], p= 0.0001, respectively) ROC area under the curve (AUC) was significantly greater for LVGLS when using automated AI quantification vs manual (AUC 0.624 and 0.535, p-value = 0.039), for LVEF this increase did not reach significance (AUC 0.598 and 0.522, p-value = 0.093). Both had greater correlation to clinical measures (such as BNP).
CONCLUSION: Echocardiographic parameters derived using automated technologies could reduce variability in data, increase clinical precision, and improve clinical trials power calculations.