DEEP FUSION OF ULTRASOUND VIDEOS FOR FUROSEMIDE CLASSIFICATION

Intensive Care Unit (ICU) physicians spend a significant amount of time on fluid management, which involves the critical decision as to whether to administer fluids or diuretics to a patient in order to maintain appropriate vascular volume. These decisions are typically made on the basis of blood tests and sometimes radiographic imaging studies, but such tests are time-consuming and inconvenient. There thus remains a need for a cost effective and reliable point-of-care decision-support tool that can be deployed easily and repeatedly throughout a patient's course as needed. Point- of-care ultrasound (POCUS) can be used to evaluate cardiac structures and the inferior vena cava (IVC), both of which yield key information about intravascular volume status. However, POCUS requires significant skill on the part of the practitioner, and erroneous interpretation can delay appropriate treatment or lead to harm such as renal injury from over diuresis. Recent advancements in artificial intelligence have shown that computer-assisted data analysis can help physicians interpret POCUS images and thus aid them to make quick and accurate decisions at the bedside. Here we present a deep fusion approach to the analysis of POCUS videos of the heart and IVC that leverages a twostream 3D network architecture. We show that this approach is capable of providing high-accuracy (86%) determination of whether to administer fluids or diuretics.