Abstract 12481: Automated Assessment of Video Quality and Ejection Fraction in Emergency Department Point-of-Care Echocardiograms

Abstract

Introduction: Bedside point of care ultrasound (POCUS) has greatly improved bedside emergency medicine through improved cost-effectiveness, prompt availability to augment physical exams, and the ability to evaluate unstable patients at the bedside. While most emergency rooms have ultrasound machines, there is severe underutilization of this bedside imaging modality, in part due to the lack of formalized ultrasound training in most emergency room physicians. Accuracy of bedside assessment of cardiac function remains variable and operator dependent. Methods: To aid with interpreting and improving the quality of emergency room transthoracic echocardiograms (TTEs), we train a deep learning algorithm to take POCUS videos and identify patients with a reduced ejection fraction. Our algorithm also identifies which POCUS videos are low quality and need to be retaken. We curated all patients who received POCUS echocardiograms in the apical 4 chamber view from Stanford ED between July 2020 and November 2020, including 927 videos from 574 patients recorded from three different machines (Phillips, GE venue, and Mindray). Each video is annotated with their ejection fraction (775 normal, 152 reduced) and their quality (768 sufficient, 159 low). Results: Using these videos, we trained and evaluated our model using 10-fold cross-validation over patients, achieving an AUROC of 0.92 (0.89 - 0.94) for predicting ejection fraction on videos of sufficient quality and an AUROC of 0.81 (0.78 - 0.85) for predicting video quality. We use our model to identify videos with qualities ranked in the top quartile and find that on these videos, our model achieves an AUROC of 0.94 (0.91 - 0.97). Conclusions: This algorithm can be run on commodity hardware in real-time, aiding in ultrasound education and significantly decreasing the barrier to entry into ED POCUS for TTEs. the quality predicted by the model can be given as feedback to physicians to aid with improving video quality and image acquisition.