Explainable Machine Learning Models for Predicting Functional and Pain Recovery After Fragility Fracture Surgery

Background: Recovery after hip fragility fracture surgery is heterogeneous. Explainable machine learning (ML) may help identify factors associated with short-term recovery during inpatient rehabilitation. Methods: This retrospective exploratory cohort included 88 patients who underwent surgery at our institution or two outside hospitals and received inpatient rehabilitation at our institution between 2017 and 2019. Improvements at discharge in activities of daily living (ADL), Harris Hip Score (HHS), and numeric rating scale (NRS) pain were defined using clinically established or data-derived thresholds. Admission demographic, clinical, and laboratory variables were analyzed using logistic regression, random forest, and XGBoost with fixed model configurations and class weighting. Model performance was assessed as a secondary feasibility analysis using repeated stratified 5-fold cross-validation (50 splits), corrected resampled t-tests, DeLong's tests, and McNemar's tests with Bonferroni correction. TreeSHAP was applied to out-of-fold predictions, and feature stability was evaluated across folds. Results: Discrimination was modest (AUC, 0.52-0.67), without significant between-model differences after correction. Prominent features were blood urea nitrogen, age, hemoglobin, and sodium for ADL; ALT (GPT), cardiovascular comorbidity, blood urea nitrogen, and age for HHS; and DXA, platelet count, sodium, and age for NRS. Stability analyses supported recurring feature associations across cross-validation folds. Conclusions: Explainable ML may be feasible as a hypothesis-generating approach for exploring factors associated with short-term recovery after fragility fracture surgery.