Prediction Model of Perioperative Blood Transfusion for Cardiovascular Surgery Patients Based on Machine Learning: Retrospective Study Using Electronic Medical Records (Preprint)

BACKGROUND Blood transfusion was related to postoperative adverse events and increased medical costs in patients underwent cardiovascular surgery. Predicting transfusion risk or major bleeding risk will help reduce transfusion. Machine learning (ML) methods show good performance at predicting risk, but transfusion risk prediction based on ML models among Chinese population were unavailable. OBJECTIVE To establish and validate prediction models using ML methods for perioperative transfusion risk of patients undergoing cardiovascular surgery in the Chinese population. METHODS Analysis was performed using electronic medical records from patients underwent cardiovascular surgery in Fuwai hospital between January 1, 2016 and June 30, 2019. Based on the 66402 unique patients, a retrospective cohort (N=61892) and a prospective cohort (N=4510) were formed for model derivation and validation. Four ML algorithms including eXtreme Gradient Boosting (XGBoost), random forest, naive Bayes, logistic regression with least absolute shrinkage and selection operator were adopted using 10-folds cross-validation to build prediction models of perioperative blood, red blood cell, plasma and platelet transfusion. According to the model evaluation in the validation cohort, the optimal perioperative blood transfusion prediction model was selected to compare with the Association of Cardiothoracic Anaesthetists perioperative risk of blood transfusion score (the ACTA-PORT score) established in previous research. RESULTS Among ML models, the XGBoost(area under the receiver-operating characteristic curve[AUC]:0.823; 95% confidence interval[CI]: 0.810 to 0.836) outperformed other models for perioperative blood transfusion and showed better prediction ability than ACTA-PORT score (AUC:0.690; 95% CI: 0.673 to 0.707; P<.001) in the validation cohort. While ML prediction models for perioperative red blood cell transfusion, plasma transfusion and platelet transfusion, achieving good model performance as AUC levels were 0.836(95% CI: 0.823 to 0.849), 0.766(95% CI: 0.745 to 0.787) and 0.948(95% CI: 0.937 to 0.959) respectively. CONCLUSIONS The study retrospectively developed and prospectively validated discriminative perioperative transfusion prediction models, which may promote the early warning and intervention against perioperative transfusion, and benefit patient blood management.