An interpretable machine learning model for predicting prognosis of medulloblastoma integrating genetic and clinical features.

backgroundMedulloblastoma (MB), the most common malignant pediatric brain tumor, lacks prognostic tools integrating clinical, molecular, and treatment-related characteristics for individualized management.

methodsWe developed machine learning models using multicenter data from 729 Chinese patients (2001-2023), of whom 509 were assigned to the training set and 220 to the testing set, and further validated the models on 201 patients from international MB consortia. To accommodate patients and researchers with varying datatypes, four application scenarios were established, including clinical-molecular-radiotherapy (CMR), clinical-molecular (CM), clinical-radiotherapy (CR), and clinical-only (CO).

resultsWe construct four model scenarios and assess their predictive performance in the testing set: an XGBoost-based CMR model (incorporating 11 features, including molecular subgroup, radiotherapy dose, and key gene expression) with a C-index of 0.612; an XGBoost-based CM (C-index = 0.609); a GBM-based CR (C-index = 0.637); and a GBM-based CO (C-index = 0.635). External validation demonstrates robust performance, with radiotherapy and molecular data contributing significantly to enhanced efficacy. In addition, interactive web-based Shiny applications have been launched to facilitate dynamic risk assessment and treatment optimization.

conclusionsBy integrating multidimensional data, our framework enables the tailored prognostication and clinical decision to meet the multidimensional requirements of research and medicine.