Construction of a predictive model for cerebrospinal fluid leakage after endoscopic transsphenoidal pituitary tumor surgery based on machine learning algorithm

Objectives: To develop and compare machine learning (ML) models predicting postoperative cerebrospinal fluid (CSF) leakage after endoscopic transsphenoidal pituitary adenectomy (TSPA). Design: Retrospective cohort study constructing prediction models using decision tree, logistic regression, and SMOTE algorithms based on clinical data. Setting: Five Class III Grade A general hospitals in Sichuan Province, China. Participants: The training cohort included 420 patients undergoing endoscopic TSPA from January 2021 to December 2022; a validation cohort of 140 patients from another hospital during the same period. Interventions: Data on tumor characteristics, operative details, and intraoperative events were collected. Patients were grouped by presence or absence of postoperative CSF leakage. Main Outcome Measures: The primary outcome was postoperative CSF leakage. Predictive model performance was evaluated by AUC, sensitivity, specificity, and calibration curves. Results: Significant differences (P<0.05) were observed in tumor diameter, texture, operative time, intraoperative diaphragma sellae rupture, reoperation, and CSF leakage grade between groups. Logistic regression identified tumor texture, reoperation, intraoperative CSF leakage grade, and diaphragma sellae rupture as risk factors. The models achieved high AUCs: logistic regression (0.792 training, 0.775 validation), decision tree (0.835 training, 0.809 validation), and SMOTE (0.839 training, 0.831 validation). The SMOTE algorithm showed the highest sensitivity and specificity. Calibration curves indicated good agreement between predicted and actual probabilities. Conclusions: ML models, especially the SMOTE algorithm, effectively predict postoperative CSF leakage after endoscopic TSPA. Incorporating multiple algorithms reduces bias and enhances predictive performance.