Cardiothoracic surgery patients have a high risk for hemorrhage and occasionally require massive transfusions, though most patients require 0 to 3 units of red blood cells (RBCs). Patient blood management models commonly rely on maximum surgical blood order schedules to avoid excessive blood waste, but a more personalized prediction on the risk of surgical hemorrhage and blood use would be useful to help conserve blood resources. To this end, researchers at the University of Utah developed a hybrid machine-learning based algorithm to predict blood usage based on data from 2410 cardiothoracic surgery patients at their institution from 2014 to 2019. A random forest non-linear algorithm was used to determine important patient variables (extracorporeal membrane oxygenation initiation, usage, and cannulation were the top three), and then a Gaussian process regression model was used to determine the relationship between the patient variables and RBC blood usage. This machine learning-based algorithm was validated on an additional 437 cardiothoracic surgery patients and was able to predict intraoperative RBC use with high accuracy, especially when 0 and 1-3 units of RBCs were transfused. Additional fine tuning for higher units of RBCs and validation at other institutions is needed, but a personalized algorithm to predict blood usage during surgery may help to avoid blood wastage.
Reference:
