Neuroblastoma is characterized by marked biological and clinical heterogeneity, with major differences in treatment and survival across risk groups. More precise prognostic tools at diagnosis could therefore help identify patients who may benefit from treatment intensification or closer monitoring.
This multicenter study from the PRIMAGE cohort developed a machine-learning model combining clinical, molecular and MRI-derived radiomic features to predict overall survival and improve risk stratification. The model was developed in 513 patients and independently evaluated in 22 patients from different hospitals.
A random survival forest integrating eight clinical, molecular and imaging features achieved a C-index of 78.8% and a time-dependent AUC of 77.7% in the discovery test sets, increasing to 93.4% and 95.4%, respectively, in the preliminary external validation cohort.
Beyond prediction, the model stratified patients into low-, intermediate- and high-risk groups and showed better discrimination of high-risk patients than the standard INRG classification. Model interpretation also revealed clinically meaningful predictors: MYCN status and LDH were the strongest contributors, while MRI radiomics captured additional information on tumor heterogeneity and size.
Importantly, previous neuroblastoma radiomics studies had largely focused on predicting individual characteristics such as MYCN amplification or bone marrow involvement. This study extends radiomics to overall survival prediction and patient risk stratification, using a substantially larger multicenter cohort and preliminary independent external validation.
Lozano-Montoya J, Jimenez-Pastor A, Fuster-Matanzo A, Weiss GJ., Cerda-Alberich L, Veiga-Canuto D, Martínez-de-Las-Heras B, Cañete-Nieto A, Taschner-Mandl S, Hero B, Simon T, Ladenstein R, Marti-Bonmati L, Alberich-Bayarri A (2025) Risk stratification in neuroblastoma patients through machine learning in the multicenter PRIMAGE cohort. Front. Oncol. 15:1528836.