Table 2_A SHAP-interpretable machine learning framework for predicting delayed discharge in ambulatory total knee arthroplasty: comparative validation of 14 models.docx

BackgroundThe rising global demand for total knee arthroplasty (TKA) has accelerated the shift toward ambulatory surgery, aimed at same-day or next-day discharge. However, significant variability in discharge protocols and high rates of delayed discharge in unselected patients challenge its widespread implementation. This study develops an interpretable machine learning framework to preemptively identify risk factors for delayed discharge in ambulatory TKA.
MethodsThis retrospective study analyzed data from 449 patients who underwent ambulatory total knee arthroplasty between September 2021 and June 2024. Fourteen machine learning models were developed and validated using preoperative variables selected via LASSO and multivariate regression. The dataset was split into training (70%) and validation (30%) sets, with hyperparameter tuning performed through grid search and 5-fold cross-validation. SHAP analysis was applied to interpret feature importance in the optimal model.
ResultsAnalysis of 449 patients identified five key predictors—ejection fraction, preoperative eGFR, preoperative ESR, diabetes mellitus, and Barthel Index—via LASSO and multivariate regression. Among 14 machine learning models, CATBoost exhibited optimal performance, with an AUC of 0.959 in training and 0.832 in validation, supported by high net benefit in decision curve analysis. SHAP analysis identified EF and preoperative ESR as the most influential features, confirmed risk directionality for low EF and low Barthel Index, and revealed nuanced interactions, such as the inverse relationship of EF with risk, enhancing model interpretability.
ConclusionThis study establishes that machine learning, particularly the CATBoost model, effectively predicts delayed discharge in ambulatory total knee arthroplasty using five key preoperative variables. SHAP analysis further enhanced model interpretability by revealing feature interactions, such as the modulating role of ejection fraction. These predictors enable improved risk stratification and personalized discharge planning, supporting optimized resource use and patient management. While limitations like single-center data require cautious interpretation, the findings highlight the potential of predictive analytics for clinical deployment. Further validation in diverse settings is warranted to translate these findings into clinical practice.