Abstract 9620: External Validation of Deep Learning-Based Echocardiography Algorithms to Automate View Selection and Annotation of the Echocardiogram

Abstract

Background: Manual measurements of echocardiography are time-consuming, error-prone, and highly user-dependent - all of which can be overcome with deep learning (DL) algorithms. Methods: We validated the automated DL measurement of the left ventricle (LV) and left atrial (LA) global longitudinal strain (GLS) and 20 conventional echocardiographic measurements using the Us.ai DL workflow. The workflow uses convolutional neural networks (CNNs) to automatically classify 2D videos and Doppler images with accuracies between 0.91-0.99. Separate CNNs automatically annotate and assess individual videos and images. We validated automated measurement in (1) an internal test set and (2) externally against core-lab measurements in the multinational PROMIS-HFpEF study. We compared the association of automated and manual parameters with N-terminal pro-B-type natriuretic peptide (NT-proBNP) in PROMIS-HFpEF. Results: The DL workflow successfully analysed 198 studies with an average of 7.8 minutes per study, with 100% reproducibility in the external validation set. In the internal validation set (N=86), automated LA (r=0.77, mean absolute error [MAE]=8.9%) and LV (r=0.91, MAE=2.8%) GLS showed good agreement with manual measurements. In the external dataset, automated measurements showed good agreement with the manual measurements, with an MAE range of 6.8 mL for LV volumes, 7.1% for LV ejection fraction, and 1.7 for the ratio of the mitral inflow E wave to the tissue Doppler e' wave ( Table ). Automated strain measurements showed good agreement (LV GLS: r=0.74, MAE=2.7%, and LA GLS: r=0.80, MAE=4.8%) with core lab measurements in PROMIS-HFpEF. Lastly, automated and manual measurements were similarly associated with NT-proBNP. Conclusions: DL-based algorithms can successfully classify and annotate echocardiographic studies, including LA and LV GLS. DL algorithms might increase reproducibility and democratise the use of strain measurements in clinical practice.