Interpretable machine learning analysis of the impact of tranexamic acid combined with antibiotics on renal function and its risk factors in hip replacement surgeries.

BACKGROUND: Early prediction of renal function changes in patients undergoing hip replacement surgery enables timely intervention for acute kidney injury and acute renal failure, which in turn improves treatment outcomes. This study aimed to develop and validate an interpretable machine learning model that uses real-world clinical parameters to predict postoperative renal function changes. The model incorporates tranexamic acid (TXA) and antibiotics as key predictive factors, examines their combined effect on renal function, and identifies associated risk factors. METHODS: This observational study was conducted from January 2022 to June 2024 at Quzhou Affiliated Hospital of Wenzhou Medical University. After data preprocessing, the full dataset was randomly split into a training cohort and an internal validation cohort at a 7:3 ratio. We performed feature selection via support vector machine-recursive feature elimination (SVM-RFE) to identify the most clinically relevant predictive variables. Changes in serum creatinine (Scr) levels were defined as the primary outcome variable. We evaluated the predictive performance of four machine learning models using multiple standard assessment metrics. Feature importance was calculated for each candidate model, and the top-performing model was further interpreted using SHAP and LIME algorithms. RESULTS: In both the training and validation cohorts, the LightGBM model outperformed RF, GBDT, and XGBoost in predicting Scr changes. SHAP analysis showed that the top five features contributing to Scr prediction in the LightGBM model were D-dimer difference, C-reactive protein (CRP) level, sodium difference, hemoglobin difference, and antibiotic use. DCA analysis indicated that all these models delivered superior predictive performance compared to traditional methods. CONCLUSIONS: The LightGBM model demonstrated higher AUC values in predicting Scr changes. TXA dosage was not identified as a top predictor of postoperative acute kidney injury in our models; however, antibiotic dosage was correlated with renal function outcomes. Our study identified and validated critical risk factors for postoperative renal impairment, including elevated CRP levels, sodium fluctuations, intraoperative blood loss, and transfusion volume. These findings have significant clinical implications for perioperative management, monitoring protocols, and risk stratification.