Machine learning models to predict red blood cell transfusion in patients undergoing mitral valve surgery

Abstract

Background Red blood cell (RBC) transfusion therapy has been widely used in surgery, and has yielded excellent treatment outcomes. However, in some instances, the demand for RBC transfusion is assessed by doctors based on their experience. In this study, we use machine learning models to predict the need for RBC transfusion during mitral valve surgery to guide the surgeon's assessment of the patient's need for intraoperative blood transfusion. Methods We retrospectively reviewed 698 cases of isolated mitral valve surgery with and without combined tricuspid valve operation. Seventy percent of the database was used as the training set and the remainder as the testing set for 13 machine learning algorithms to build a model to predict the need for intraoperative RBC transfusion. According to the characteristic value of model mining, we analyzed the risk-related factors to determine the main effects of variables influencing the outcome. Results A total of 166 patients of the cases considered had undergone intraoperative RBC transfusion (24.52%). Of the 13 machine learning algorithms, CatBoost delivered the best performance, with an AUC of 0.888 (95% CI: 0.845-0.909) in testing set. Further analysis using the CatBoost model revealed that hematocrit (<37.81%), age (>64 y), body weight (<59.92 kg), body mass index (BMI) (<22.56 kg/m2), hemoglobin (<122.6 g/L), type of surgery (median thoracotomy surgery), height (<160.61 cm), platelet (>194.12×109/L), RBC (<4.08×1012/L), and gender (female) were the main risk-related factors for RBC transfusion. A total of 204 patients were tested, 177 of whom were predicted accurately (86.8%). Conclusions Machine learning models can be used to accurately predict the outcomes of RBC transfusion, and should be used to guide surgeons in clinical practice.