BIOS-05. USING DATA-DRIVEN ANALYTICS TO PREDICT SURVIVAL IN PEDIATRIC PATIENTS WITH MEDULLOBLASTOMA: THE UTILITY OF MACHINE LEARNING IN NEXT-GENERATION NEURO-ONCOLOGY PROGNOSTICATION

Abstract

Abstract Medulloblastoma is the most common malignant intracranial tumor affecting the pediatric population. Despite advancements in multimodal treatment over the past 2 decades yielding a >75% 5-year survival rate, children who survive often have substantial neurological and cognitive sequelae. We aimed to identify risk factors and develop a clinically friendly online calculator for prognostic estimation in pediatric medulloblastoma patients. Pediatric patients with a histopathologically confirmed diagnosis of medulloblastoma were extracted from the Surveillance Epidemiology and End RESULTS: database (2000-2019) and split into training and validation cohorts in an 80:20 ratio (Figure 1). The Cox proportional hazards model was used to identify the univariate and multivariate survival predictors. Subsequently, a calculator with those factors was developed to predict 2-, 5-, and 10-year overall survival as well as median survival months for pediatric medulloblastoma patients. The performance of the calculator was determined by discrimination, calibration, and decision curve analysis (DCA). A total of 1,739 pediatric patients with medulloblastoma met the prespecified inclusion criteria. Fourteen variables, including age, sex, race, ethnicity, median household income, county attribute, laterality, histology, anatomical location, tumor grade, tumor size, surgery status, radiotherapy, and chemotherapy, were included in the calculator (). The concordance index was 0.757 in the training cohort and 0.762 in the validation cohort, denoting clinically useful predictive accuracy. Good agreement between the predicted and observed outcomes was demonstrated by the calibration plots. The DCA curves indicated that the developed model has a good clinical prognostic benefit for pediatric medulloblastoma patients, and could aid prognostication efforts in the clinical setting. An easy-to-use prognostic calculator for a large cohort of pediatric patients with medulloblastoma was established. Future efforts should focus on improving granularity of population-based registries and externally validating the proposed calculator.