Abstract
OBJECTIVE: The primary objective of this study was to develop a length of stay (LOS) prediction model. BACKGROUND: Predicting the LOS is crucial for patient care, planning, managing expectations, and optimizing hospital resources. Prolonged LOS after colorectal surgery is largely influenced by complications, and an accurate prediction model could significantly benefit patient outcomes and healthcare efficiency. METHODS: This study included patients who underwent colorectal surgery in 14 different hospitals between January 2016 and December 2020. Two distinct random forest models were developed: one solely based on preoperative variables (preoperative prediction model [PP model]) and the other incorporating both preoperative and intraoperative variables (intraoperative prediction model [IP model]). Both models underwent validation using 10-fold cross-validation. The discriminative power of the model was assessed using the area under the curve (AUC), and calibration was evaluated using a calibration curve. The 2 developed models were compared using DeLong test. RESULTS: A total of 2140 patients were included in the analysis. After internal validation, the PP model achieved an AUC of 0.75 (95% confidence interval [CI]: 0.73-0.77), and the IP model achieved an AUC of 0.84 (95% CI: 0.82-0.85). The difference in discrimination between the 2 models was statistically significant (DeLong test, P < 0.001). Both models exhibited good calibration. CONCLUSIONS: Incorporating intraoperative parameters enhances the accuracy of the predictive model for LOS after colorectal surgery. Improving LOS prediction can assist in managing the increasing number of patients and optimizing the allocation of healthcare resources.