Development and Statistical Validation of a Machine Learning Model for Predicting Functional Outcomes at Discharge...

Hip fractures in aging populations pose a major healthcare burden, and predicting discharge motor function may enable timely interventions. We aimed to develop a machine learning model with statistically validated reliability to predict motor Functional Independence Measure (mFIM) scores at discharge in patients who underwent bipolar hip arthroplasty (BHA).

This retrospective study was conducted at a regional hospital providing integrated care. A total of 201 hips treated with BHA for femoral neck fractures were analyzed. The primary outcome was the discharge mFIM score. Ten predictors were assessed: age; sex; presence of hypertension, diabetes, or heart failure; body mass index; Hasegawa Dementia Scale-Revised (HDS-R) score; time from admission to surgery; time from surgery to transfer to the convalescent ward; pre-fracture mobility status; pre-fracture independence level; and mFIM score at transfer. Six machine learning models were developed with hyperparameter tuning. Feature importance was evaluated using SHapley Additive exPlanations (SHAP), and results were compared with multiple regression for consistency.

Light gradient boosting machine achieved the highest predictive performance (R²= 0.84). SHAP analysis revealed that the most influential predictors were mFIM score at transfer to the rehabilitation ward, HDS-R score, and pre-fracture independence level. These factors were identified as significant variables in multiple linear regression.

We developed a reliable and interpretable machine learning model to predict discharge mFIM scores in patients who underwent BHA. Key predictors of the model were supported by SHAP and multiple regression analysis. This model can assist clinicians in setting appropriate rehabilitation goals and discharge plans early during recovery.