Project 467197

Ejection Fraction (EF) is a cardiovascular measurement indicating the amount of blood pumped by the heart. A low EF value can be an indicator of imminent heart failure, and therefore, monitoring changes in EF is crucial in determining the best treatments and preventing life-threatening cardiovascular abnormalities.With the emergence of accessible point-of-care diagnostic medical devices, the ability to quickly estimate EF by non-professional users gains particular importance. In this project, we use a novel Machine Learning technique called Graph Neural Networks to enable real-time, high-accuracy EF estimation while providing performance interpretability.Doctors estimate EF by examining ultrasound videos of the heart. They first find the video frames corresponding to the points in time when the heart is fully contracted (End-Systole (ES)) and fully expanded (End-Diastole (ED)). The ratio between the volume of blood at these points is then used to approximate EF. And to make an accurate estimate, doctors rely on multiple ultrasound videos obtained from different angles of the heart for the same patient.Following the clinical procedure above, for each patient, we create a fully connected graph with ultrasound video frames as its nodes. This graph contains videos obtained from different angles of the heart, similar to how doctors consider multiple angles when determining EF. Graph Neural Networks are then used to assign an importance measurement to each video frame before performing EF estimation. These importance scores can be investigated to ensure higher weights are given to ES and ED frames of the ultrasound videos, which act as a measure of reliability and interpretability. This approach enables explainable, real-time EF estimation in POC devices.