Machine learning for the prediction of blood transfusion risk during or after mitral valve surgery: a multicenter retrospective cohort study

Yuhan Wang ¹ Leping Liu ² Kexin Fan ³ yongjun Wang ⁴ Jiyan Zhang ⁵ Xianjun Ma ⁶ Yuanshuai Huang ⁷ Xinhua Wang ⁸ Bingyu Chen ⁹ Rong Gui ^1*

• ¹ Department of Transfusion, Third Xiangya Hospital, Central South University, Changsha, China
• ² Department of Pediatrics, Third Xiangya Hospital, Central South University, Changsha, China
• ³ Department of Laboratory Medicine, Second Xiangya Hospital, Central South University, Changsha, Hunan Province, China
• ⁴ Department of Transfusion, Second Xiangya Hospital, Central South University, Changsha, Hunan Province, China
• ⁵ Anesthesia Center, Fuwai Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, Beijing, China
• ⁶ Department of Transfusion, Qilu Hospital of Shandong University (Qingdao), Qingdao, Shandong, China
• ⁷ Department of Transfusion, The Affiliated Hospital of Southwest Medical University, Luzhou, Sichuan, China
• ⁸ Department of Transfusion, Beijing Aerospace General Hospital, Beijing, Beijing Municipality, China
• ⁹ Department of Transfusion, Zhejiang Provincial People's Hospital, Hangzhou, Zhejiang Province, China

The study aimed to determine the best prediction method for red blood cell (RBC) transfusion during or after mitral valve surgery by utilizing the feature selection method to determine preoperative variables that can forecast the risk of RBC transfusion. This information can aid in managing RBC transfusions during or after mitral valve surgery.: 1477 patients underwent mitral valve surgery at eight large tertiary hospitals in China, all receiving cardiopulmonary bypass treatment. Thirty preoperative factors were collected, including demographic characteristics, medical history, physical signs, laboratory indicators, and surgery type. Key variables were screened via The Max-Relevance and Min-Redundancy (mRMR) method. Various machine learning models were developed to sort and analyze patient data concerning mitral valve surgery. Every model was evaluated using five-fold crossvalidation. The SHapley Additive exPlanations (SHAP) was utilized to interpret the models with the best predictions.Results: Out of the 1,477 patients included in the study, 862 patients who underwent mitral valve surgery received RBC transfusions during or after mitral valve surgery.The mRMR method was used to screen ten preoperative variables from a pool of thirty for subsequent modeling analysis. Among all models, LightGBM exhibited superior performance (AUC of 0.935 in the training set and 0.734 in the validation set) and performed well in the prospective dataset with an accuracy of 74.2%. SHAP analysis visualized the ten most influential variables for outcome prediction, ranked as follows: hematocrit, red blood cells (RBCs), weight, body mass index (BMI), fibrinogen, hemoglobin, height, age, left ventricular dilation, and sex.This study evaluated preoperative variables related to RBC transfusion and determined that LightGBM was the best model for RBC transfusion during or after mitral valve surgery. The model demonstrated high accuracy in predicting RBC transfusion likelihood from clinical datasets. This facilitates clinicians in anticipating the need for RBC transfusion before the patient's surgery and assists them in making decisions about blood management.