Predictive Factors of Inpatient Rehabilitation Outcomes and Stay: A Machine Learning Study with Temporal Validation

Background/Objectives: Despite the increasing rates of disability associated with aging, obesity, and osteoarthritis requiring surgery, optimizing rehabilitation after hospital discharge remains a major challenge and a key determinant of care safety, quality, and sustainability. The aim of this exploratory study is to evaluate the predictive performance of machine learning (ML) models for predicting Inpatient Rehabilitation Length Of Stay (IRLOS), Function in the Activities of Daily Living (FADL) and discharge destination (DD) in patients who underwent total joint replacement for hip and knee osteoarthritis, using Real-World Data routinely collected in a tertiary orthopedic hospital. Methods: 2103 patients were included and temporally split into a development cohort (2019; n = 1711) and a temporal validation cohort (2018; n = 392). A total of 73 routinely collected perioperative variables were used to train multiple ML models, including both black-box and transparent methods. IRLOS and FADL were modeled as regression tasks, while DD was treated as a binary classification task. Model development followed a rigorous pipeline with feature selection, cross-validation, and hyperparameter tuning. Performance was assessed using appropriate metrics and evaluated across joint type (hip/knee) and via temporal validation. Model interpretability was examined using SHAP and model-specific analyses, further supported by clinical analysis. Results: In temporal validation, models achieved modest performance for IRLOS (R² = 0.17; MAE = 2.52 days) and FADL (R² = 0.25; MAE = 2.25 days), with no significant performance degradation over time. DD prediction showed good discrimination (AUC = 0.85; balanced accuracy = 0.80) despite outcome imbalance, with high sensitivity (0.92) and negative predictive value (≈1.00), but low positive predictive value (0.09). Performance was stable across hip and knee subgroups. The interpretability analysis further highlighted several key predictors related to perioperative complexity (e.g., surgical duration and transfusion), baseline functional status, and social factors (e.g., living situation and employment), confirming that rehabilitation outcomes are multidimensional and influenced by both medical and non-medical determinants. Conclusions: The analysis identified several relevant predictors related to perioperative complexity, baseline functional status, and social context. The models showed stable behavior across intervention type and across time through temporal validation, suggesting that they capture relevant patterns in rehabilitation pathways, although predictive performance was moderate.