Previous Abstract | Next Abstract
Printable Version
October 03, 2020
10/3/2020 12:00:00 PM - 10/3/2020 1:00:00 PM
Room Virtual
Point Of Care Ultrasound (pocus) Auto Ef Computer Learning
Michael Hall, M.D., Sam Surett, B.S., Frederico Asch, M.D., Jose L. Diaz-Gomez, M.D., Ali Chaudhry, M.B.A., Ha Hong, Pharm.D, Nicolas Poilvert, Ph.D., Rachel Liu, M.D., Sara Nikravan, M.D.
University Of Washington, Seattle, Washington, United States
Disclosures: M. Hall: None. S. Surett: Salary; Self; Caption Health. F. Asch: None. J.L. Diaz-Gomez: None. A. Chaudhry: Salary; Self; Caption Health. H. Hong: Salary; Self; Caption Health. N. Poilvert: Salary; Self; Caption Health. R. Liu: None. S. Nikravan: None.
Background: Point of Care Ultrasound (POCUS) is a useful tool for quickly and safely assessing cardiac function in the setting of shock. However, barriers to adoption of POCUS exist. The standard bi-plane Simpson’s method of calculating left ventricular ejection fraction (LVEF) requires acquisition of diagnostic quality Apical 2 (AP2) and 4 chamber (AP4) images, which are difficult to obtain. In addition, physicians, particularly those without echocardiography training, have high variability and less reliability in visually identifying abnormalities in LVEF. Machine learning has been shown to accurately assess LVEF comparable to an expert in echocardiography. A deep learning (DL) algorithm (Caption Interpretation, Caption Health) was designed to calculate LVEF from available AP2, AP4, and parasternal long-axis (PLAX) views. Given the difficulty of acquiring apical windows in the acute care setting, the investigators sought to understand if accurate automated LV assessments could be made from the more readily obtained PLAX view. Methods: Echocardiographic studies from 166 patients were traced by 3 sonographers and assessed by a panel of 3 cardiologists who independently performed volumetric calculations of AP2/AP4 tracings per standard clinical practice; this panel read formed the “reference standard” LVEF for each study. The DL algorithm processed all of the studies to produce an LVEF calculation for each. Ten (10) physicians, including 7 who use POCUS clinically and 3 cardiologists, provided a qualitative visual assessment of LV function (i.e., hyperdynamic (EF > 73%), normal (EF 53-73%), reduced (30-52%), and severely reduced (EF < 30%)), and LVEF integer assessment based on a single clip from each view (PLAX, AP4, AP2). The DL algorithm was assigned to the same LV function categories based on its integer estimate. For each patient, the DL LVEF and associated qualitative assessment was compared to the physician visual assessment. Results: The DL algorithm showed comparable or higher accuracy than physicians when predicting the correct qualitative LV function from a single view (Figure 1). For EF estimation, it showed a lower incidence of outliers (defined as LVEF estimates >15% from reference standard) than physician visual assessments (Figure 2). Conclusion: This study highlights the utility of a DL algorithm for POCUS users, automatically producing LVEF estimates similar to expert assessments from even single diagnostically acceptable images. Ultimately, such technology may allow bedside providers to more quickly and accurately diagnose cardiac dysfunction for the purpose of clinical decision making in the care of acutely ill patients.
Figure 1
Figure 2

Copyright © 2020 American Society of Anesthesiologists