A Clinically Interpretable Machine Learning Framework for Mortality Prediction in Critically Ill Orthopedic Patients

Background The global burden of geriatric orthopedic conditions (e.g., hip fractures, complex trauma) is rising, posing significant challenges to critical care medicine. Existing prognostic tools rely on single indicators or traditional scoring systems, lacking sufficient accuracy and interpretability in ICUs. This study aims to develop an interpretable machine learning (ML) framework with high predictive performance for individualized risk stratification. Methods Critically ill orthopedic patients ≥ 50 from MIMIC-IV/III were enrolled (excluding those with malignant tumors or incomplete data). MIMIC-IV was split 7:3 into training/testing sets. LASSO Cox regression with 20-fold cross-validation enabled dimension reduction and feature selection. Nine ML models were compared; the optimal model was selected by accuracy and AUROC. SHAP quantified feature impacts and individual decision processes. MIMIC-III served for external validation. Results 6,488 patients were included, with 11.8% in-hospital mortality in MIMIC-IV. Eight core features (age, APSIII, SOFA, blood glucose, WBC, lactate, body temperature, CRRT) were identified. Logistic regression performed best (AUROC = 0.82) and achieved AUC = 0.81 (95% CI: 0.78–0.83) in external validation. Conclusion This interpretable mortality prediction model for critically ill orthopedic patients aids preoperative risk assessment and postoperative ICU monitoring, supporting targeted early interventions to improve outcomes.