B-258 Laboratory Data–Driven Explainable Ensemble Machine Learning for Predicting Postoperative Mortality and Acute Kidney Injury in Hip Fracture Patients

Abstract

Postoperative mortality and acute kidney injury (AKI) remain major complications following hip fracture surgery, particularly in elderly patients. Although numerous clinical risk factors have been identified, the prognostic potential of routinely available laboratory parameters is often underutilized in perioperative risk stratification. Unlike prior studies focusing on single outcomes or regression-based approaches, our framework integrates diverse laboratory biomarkers within an explainable ensemble architecture. We aimed to develop an explainable, laboratory data–driven ensemble machine learning (ML) framework to simultaneously predict 6-month mortality and postoperative AKI in hip fracture patients. We retrospectively analyzed electronic health record data from 1,276 patients who underwent hip or femur fracture surgery between 2019 and 2023. A total of 106 laboratory and clinical variables were included for mortality prediction. For AKI prediction, 103 variables were used after excluding renal function–derived features to avoid data leakage. Routine laboratory parameters included complete blood count indices, renal function markers, inflammatory biomarkers, cardiac biomarkers, electrolytes, and urinalysis results. Five ML models—CatBoost, XGBoost, LightGBM, Random Forest, and a multilayer perceptron (MLP)—were trained using stratified five-fold cross-validation. A weighted soft-voting ensemble model was constructed by assigning optimized weights based on model performance. Model interpretability was assessed using Shapley Additive Explanations (SHAP) to quantify the contribution of individual laboratory biomarkers. The ensemble approach consistently improved discrimination compared with individual boosting and deep learning models, highlighting the complementary predictive contributions of heterogeneous algorithms applied to structured laboratory datasets. The weighted ensemble model demonstrated superior performance compared with individual algorithms. For 6-month mortality prediction, the ensemble achieved an area under the receiver operating characteristic curve (AUROC) of 0.889, outperforming all standalone models. For AKI prediction, the ensemble achieved an AUROC of 0.779 with improved F1-score compared to individual models. SHAP analysis consistently identified laboratory biomarkers as dominant predictors. Mortality risk was strongly associated with renal function markers (creatinine, estimated glomerular filtration rate, blood urea nitrogen), inflammatory biomarkers (procalcitonin), nutritional markers (serum albumin), and cardiac stress indicators (proBNP). For AKI prediction, body weight, blood urea nitrogen, urine white blood cell count, and inflammatory markers were prominent across tree-based models. Notably, the MLP model uniquely emphasized LDH, fibrinogen, and proBNP, suggesting nonlinear associations captured by deep learning architectures that were less apparent in tree-based approaches. Routine laboratory parameters contain substantial prognostic information for predicting postoperative mortality and AKI after hip fracture surgery. An explainable ensemble ML approach enhances predictive performance while preserving interpretability of clinically meaningful laboratory biomarkers. These findings support the integration of laboratory data–driven, explainable AI models into clinical laboratory medicine workflows for early risk stratification and improved perioperative management.