Over 85 million units of RBCs are transfused annually worldwide, and the decision to transfuse is usually based on a few key laboratory values, such as hemoglobin. Other laboratory and vital sign data, however, may also help predict when a patient will need a transfusion. Researchers used machine learning to develop a meta-model based on over 40 variables routinely collected at Emory University Hospital over five years on all non-trauma patients admitted to the intensive care unit (ICU) to predict the likelihood of a transfusion. Four years of data from 2016 to 2020 was used to train five distinct machine learning algorithms (logistic regression, random forest, feedforward neural networks, support vector machines, and XG Boost), and a collective meta-model was tested on the final year of data. Based on 72,072 patient encounters (median age, 63 years; 47% female) and 24 hours of data either after ICU admission for non-transfused patients (n=53,758) or the 24 hours before transfusion (n=18,314), researchers found that their meta-model had an accuracy rate of 0.93 in predicting the likelihood of a transfusion. Furthermore, the area under the curve receiver operating characteristic curve was 0.97, and the F1 score was 0.89 supporting the robustness of the meta-model. Not surprisingly, platelet counts and hemoglobin levels were important variables in the model, but researchers found many other significant variables including lipase levels, creatine levels, and SpO2/FiO2 ratios. Further prospective studies and fine-tuning of using machine learning to predict the likelihood and type of transfusions are needed.
Reference:
Leave a Reply