The role of anesthesia in TKA length of stay: a data-driven analysis using Boruta algorithm

Background Optimizing the length of hospital stay (LOHS) after total knee arthroplasty (TKA) is essential for improving healthcare efficiency. The effect of anesthesia method on LOHS remains controversial and challenging to accurately assess using conventional statistics. This study leverages robust machine learning with dedicated feature selection to clarify its impact within a limited but well-characterized patient cohort. Methods We analyzed data from 157 patients undergoing primary unilateral TKA. The Boruta algorithm—a robust feature selection method based on Random Forest—was employed to identify the most predictive variables for LOHS from perioperative parameters. Four machine learning models (Linear Model, Random Forest, XGBoost, and K-Nearest Neighbors) were built using Boruta-confirmed features and evaluated via 10-fold cross-validation. Results General anesthesia (GA) was associated with significantly longer LOHS (15.93 ± 4.04 days) compared to intrathecal anesthesia (IA) (14.31 ± 4.09 days, p = 0.012). The Boruta algorithm confirmed anesthesia method as an important predictor, along with age and preoperative hospitalization duration. While the Linear Model offered the best interpretability (mean R 2 = 0.475), XGBoost showed a strong ability to model nonlinear patterns. Subgroup analyses revealed that GA was linked to prolonged anesthesia time, higher blood loss, and increased transfusion requirements, factors that are collectively associated with extended hospitalization. Conclusion Despite the limited sample size, the use of Boruta feature selection ensured model focus on clinically meaningful predictors. We identify anesthesia method as a key modifiable factor associated with LOHS after TKA. These data-driven insights support the preferential consideration of IA for eligible patients in similar clinical contexts to potentially shorten hospital stays and optimize resource use, demonstrating the value of explainable machine learning in small-sample clinical studies.