Optimal model for predicting intraoperative blood transfusion in elective surgery patients: a comparative study of eight machine learning methods

Abstract

Background - To identify the optimal machine learning model for predicting intraoperative blood transfusion requirements in elective surgery patients by systematically evaluating eight algorithms.

Materials and methods - A retrospective cohort of 1,500 elective surgery patients was screened, with 1,017 meeting inclusion criteria. Demographic (gender, age, height, weight), preoperative (liver, cardiac, pulmonary, coagulation functions), and intraoperative indicators (surgery grade, anesthesia score, estimated blood loss, vital signs) were collected. After univariate analysis, 20 significant variables were selected for modeling. The dataset was split 7:3 into training and testing sets. Eight models −Random Forest (RF), Generalized Linear Model (GLM), Support Vector Machine (SVM), Gradient Boosting Machine (GBM), k-Nearest Neighbors (KNN), Neural Network (NNet), λ determined by cross-validation (LASSO), and Decision Tree (DT)− were trained using five-fold cross-validation. Performance was evaluated based on the area under the receiver operating characteristic curve
(ROC-AUC), precision-recall curves (PR-AUC), residuals, and variable importance.

Results - In the training set, RF achieved the highest AUC (1.000), followed by GBM (0.992) and SVM (0.987). In the testing set, RF maintained superior performance (AUC=0.992), with high precision-recall (AUC-PR=0.988) and minimal residuals. Key predictors included preoperative hemoglobin (preHGB), EBL, and coagulation markers (preAPTT, preDD).

Discussion - RF is the most reliable model for intraoperative transfusion prediction, offering high accuracy and clinical interpretability. This study provides a data-driven tool to optimize transfusion strategies and reduce adverse outcomes.