A prospective cohort study develops and validates a machine learning model for predicting ecchymosis after total knee arthroplasty

Total knee arthroplasty (TKA) is a standard procedure for end-stage knee diseases, yet perioperative bleeding and its complications, such as ecchymosis, remain significant challenges. Predicting ecchymosis can guide personalized perioperative blood management, enhancing patient recovery and reducing the socioeconomic burden. In this prospective study (June 2023-March 2024), 416 TKA patients were split into ecchymosis (n = 135) and non-ecchymosis (n = 281) groups. The first seven months' data (312 patients) were used for training, and the next three months' (104 patients) for validation. Feature selection used Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest-Recursive Feature Elimination (RF-RFE), and BORUTA to identify key risk factors. The model was tested with advanced machine learning: Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Support Vector Machines (SVM), and Light Gradient Boosting Machine (LGBM). Major risk factors included low prealbumin, reduced coagulation index (CI) and its change (XCI), high fibrinogen degradation products (FDP), and postoperative day 1 total blood loss (TBL). The model showed robust performance with area under the curve values (AUCs) of 0.927 in training and 0.954 in validation set. The model's superior high performance offers significant guidance for early and accurate prediction, enabling personalized anticoagulation therapy and optimizing perioperative blood management in TKA patients.Trial registration: Chinese Clinical Trial Registry (registration No. ChiCTR2400080173).