Craniopharyngioma.AI: An Explainable Ensemble Learning Framework for Personalized Craniopharyngioma Outcome Prediction

Introduction Craniopharyngiomas are rare sellar/suprasellar tumors with long-term postoperative morbidity, including permanent diabetes insipidus (DI) and tumor recurrence. Existing clinical tools offer limited individualized prognostication. We introduce a pilot artificial intelligence (AI) platform to hypothesize postoperative outcomes in craniopharyngioma surgery. Methods Patients with histologically confirmed craniopharyngioma who underwent surgical resection between 2008 and 2023 at a single institution were retrospectively analyzed. Clinical, radiographic, and intraoperative features (e.g., tumor dimensions, volume, intraventricular extension, surgical approach, and extent of resection) were modeled using machine learning algorithms with a unified ensemble pipeline to predict permanent DI and tumor recurrence. Explainable AI interpretation using Shapley Additive Explanations (SHAP) quantified individual feature-level contributions, which were integrated into a FastAPI-powered web application for interactive prediction. Results 59 patients were used in a stratified cross-validation model training, of whom permanent DI occurred in 65.5% and tumor recurrence was recorded in 28.8%. The blended Ridge-Catboost ensemble achieved the highest performance for permanent DI prediction (AUC: 0.77±0.05, F1: 0.81±0.04), and a blended Naïve Bayes-Catboost ensemble was chosen for recurrence prediction (AUC: 0.78±0.12, F1: 0.67±0.15). SHAP-based feature ranking revealed that pre-operative BMI, age, and endocrine or visual dysfunction were the strongest predictors of DI and recurrence. Surgical approach and postoperative BMI changes further modulated DI risk. The app can be accessed at tinyurl.com/craniopharyngiomas. Conclusion In this small, single-institution cohort, ensemble learning and model tuning helped stabilize model accuracy and discriminative power. This exploratory AI prototype for prognostication might eventually support data-driven risk assessment in skull base oncology.