Comparison of seven machine learning models for predicting prolonged PACU length of stay: performance evaluation and identification of key influencing factors

Abstract

Abstract Background Prolonged stays in the post-anesthesia care unit (PACU) significantly increase healthcare resource utilization and delay postoperative recovery. Current research on PACU prolonged stays is mostly limited to univariate analyses, lacking multidimensional, high-accuracy predictive models to support precise clinical risk stratification. Methods This single-center retrospective study included 1,996 PACU patients (May-December 2024) at a tertiary hospital in Zhejiang, China. Based on literature review and expert input, 37 candidate predictors were selected. To prevent data leakage, the dataset was randomly split into training (70%) and validation (30%) sets. All feature selection-univariate/multivariate analysis ( P < 0.05), Boruta, and LASSO-was performed strictly on the training set, and the intersecting predictors were used for model building. Hyperparameters were tuned via grid search with 5-fold cross-validation on the training set. Seven machine learning algorithms including logistic regression (LR), decision tree (DT), random forest (RF), support vector machine (SVM), LightGBM, XGBoost, and artificial neural network (ANN) were developed using R and Python, and evaluated on the independent validation set using AUC, accuracy, sensitivity, specificity, F1-score, and Brier score. SHAP analysis was used to interpret key predictors in the best-performing model. Results The XGBoost model demonstrated the best performance, achieving an AUC of 0.901 (95% CI: 0.863–0.932) on the validation set and a sensitivity of 65.83% for detecting prolonged PACU stays. SHAP analysis identified surgical duration (mean SHAP value 0.19), anesthesia duration (0.08), anesthesia method (0.06), total protein (0.06), and albumin-globulin ratio (0.03) as the top five predictive factors. Conclusions The XGBoost model developed in this study achieved a relatively high AUC of 0.901. However, its sensitivity for detecting prolonged PACU stay was 65.83%, indicating that the model’s ability to confirm prolonged PACU stay remains somewhat insufficient. Therefore, it can be used as an auxiliary screening tool in clinical rather than a definitive diagnostic tool.