Objective: Medullary gliomas pose significant surgical risks, particularly the risk of postoperative lower cranial nerve (LCN) dysfunction, which profoundly affects quality of life. The lack of standardized risk assessment hinders optimal surgical planning. The aim of this study was to develop and validate an individualized predictive model for short-term postoperative LCN impairment integrating clinical and imaging data and to estimate individual risk across a range of resection extent to optimize surgical planning.
Methods: A retrospective cohort (n = 111, January 2020-February 2023) was used for model development, with prospective validation (n = 45, February 2023-December 2024). The primary outcome was postoperative LCN dysfunction (inability to be extubated within 14 days or requiring tracheotomy with persistent ventilation). Predictive modeling was performed using logistic regression, incorporating multistage feature selection, hyperparameter optimization, and bootstrapped validation. Model performance was evaluated using metrics such as area under the curve (AUC), Brier score, calibration, and decision curve analysis (DCA) in the prospective validation set. Shapley Additive Explanations (SHAP) analysis was used to interpret feature contributions, and a nomogram was constructed for clinical implementation. Optimal extent of resection (EOR) thresholds were explored to balance functional preservation and tumor clearance.
Results: Four independent predictors of LCN dysfunction were identified: EOR (OR 1.84, 95% CI 1.07-3.16), infiltrative growth (OR 10.46 [95% CI 2.70-40.52]), preoperative LCN impairment (OR 6.79 [95% CI 2.54-18.16]), and cervical cord involvement (OR 4.64 [95% CI 1.55-13.91]). The model demonstrated strong discrimination (training AUC 0.85 [95% CI 0.76-0.92], testing AUC 0.89 [95% CI 0.79-0.97]) and good calibration (Brier score 0.12). High-risk patients, defined as those with a model predicted risk probability > 0.471 based on Youden's index, had significantly higher rates of pneumonia, tracheostomy, and prolonged mechanical ventilation. Stratified resection plans showed that low-risk patients benefited from gross-total resection (EOR 88.4%, [95% CI 75.7%-100.0%]), while high-risk patients achieved optimal functional outcomes with limited resection (EOR 40.8%, [95% CI 38.2%-50.5%]). SHAP and nomogram analyses provided transparent, patient-specific information for risk consultations.
Conclusions: This study presents the first predictive model tailored to short-term postoperative LCN outcomes following medullary glioma surgery, proposing a dynamic resection paradigm based on individualized risk stratification. By guiding surgical planning and intraoperative decision-making, this model facilitates a balance between maximal tumor control and functional preservation.
Keywords: lower cranial nerve dysfunction; machine learning; medullary glioma; oncology; predictive model; risk stratification; skull base; surgical decision-making; tumor.