Interpretable ensemble learning for hospital prediction duration following total knee arthroplasty: a retrospective temporal validation study

Estimating postoperative length of stay (LOS) following total knee arthroplasty (TKA) may help characterize resource use and identify factors associated with prolonged hospitalization. However, the predictive performance and clinical applicability of machine learning models for continuous LOS estimation remain uncertain. This study aimed to develop and temporally validate interpretable machine learning models for predicting LOS after TKA. A retrospective cohort study was conducted using data from patients undergoing primary TKA at a tertiary centre in Singapore. After missing data imputation, preprocessing, and feature selection within the training cohort, multiple regression-based machine learning models were developed and evaluated using a temporal split. Model performance was primarily assessed using mean absolute error, root mean squared error, R², and calibration-related measures. Tolerance-based accuracy was reported as a supplementary descriptive metric. SHapley Additive exPlanations (SHAP) analysis was used to explore model-dependent feature contributions, and regression-based sensitivity analyses were conducted to assess the consistency of identified associations. A total of 2,414 patients were included. Across the evaluated models, predictive performance was modest, with low explained variance in the temporal test cohort. The Ensemble Voting Regression (EVR) model showed small numerical advantages across several metrics and acceptable average calibration, although differences from other models were small. Operation duration, age, haemoglobin, body mass index, type of anaesthesia, and procedure site contributed to LOS prediction. SHAP-based findings were generally consistent with regression-based analyses for several predictors. Interpretable machine learning models provided modest but clinically informative estimates of postoperative LOS after TKA under temporal validation. EVR showed small numerical advantages and acceptable average calibration, although low explained variance limited individualized prediction. These findings support population-level LOS estimation and preliminary risk stratification and provide a realistic benchmark for future model refinement, recalibration, and validation.