Presentation (CTI)

OBJECTIVES: Neuroblastoma is one of the most common extracranial solid tumors in children and accounts for about 15% of pediatric cancer deaths. Substantial tumor heterogeneity among affected children leads to variable prognoses. Thus, this study aimed to develop and validate an artificial intelligence (AI)-based random survival forest (RSF) model to predict survival in pediatric neuroblastoma.
METHODS: We conducted a retrospective analysis of the Surveillance, Epidemiology, and End Results (SEER) registry (2004-2022) to predict cancer-specific survival among patients aged <18 years with a histologically confirmed, primary diagnosis of neuroblastoma or ganglioneuroblastoma. Features included demographic factors, clinical characteristics, and treatments. Categorical variables were pre-processed with one-hot encoding, and missing values were flagged to capture their predictive value. We derived and compared the RSF model with a Cox proportional hazards (CoxPH) model, evaluating performance through the concordance index (C-index), area under the receiver operating characteristic curve (AUC), integrated Brier score, and decision curve analysis (DCA). A complete-case analysis was conducted as a sensitivity analysis.
RESULTS: The study included 3,557 pediatric neuroblastoma patients. The RSF model outperformed the CoxPH model on the training (C-index: 0.84 vs 0.78) and test datasets (C-index: 0.80 vs 0.77) and demonstrated superior accuracy in predicting survival rates at 1, 3, and 5 years (AUC: 0.79-0.80 vs 0.75-0.77). The integrated Brier score and DCA of the RSF model were also favorable, confirming its calibration and clinical utility. Permutation feature importance of the RSF model revealed that chemotherapy status was most predictive of survival, followed by age, summary stage, tumor size, and radiotherapy status. The complete-case analysis demonstrated similar model performance (C-index: 0.79), confirming model robustness.
CONCLUSIONS: This study demonstrates the potential of AI-based approaches using publicly available data to predict pediatric neuroblastoma survival. With external validation, RSF models could inform clinical decision-making and improve patient outcomes.