Uncertainty-Aware Risk Stratification in Pediatric Low-Grade Glioma Using Multimodal Data.

BACKGROUND AND PURPOSE: Risk stratification in pediatric low-grade glioma (pLGG) remains challenging due to biological and clinical heterogeneity. We developed an uncertainty-aware multimodal survival framework that integrates deep learning features from T2-weighted MRI, molecular subtype, and clinical information. MATERIALS AND METHODS: Data from the Children's Brain Tumor Network included 360 subjects with imaging data and 493 with molecular subtype information derived from tumor tissue genomic profiling; clinical data were available for all patients. A pretrained deep learning model was fine-tuned for tumor segmentation using a pediatric brain tumor cohort (n=752) and subsequently used to extract imaging features from T2-weighted MRI. A regularized survival model integrating imaging and clinical features (clinico-ResNet; DL-M1) and a separate clinical-molecular survival model were trained and validated on discovery cohorts and evaluated on independent replication cohorts. A multimodal model (DL-M2), combining risk scores from the clinico-ResNet and clinical-molecular models through late fusion, was developed in the subsect of patients with both imaging and molecular data (n=294). Bootstrap resampling was used to quantify per-patient prediction uncertainty. RESULTS: DL-M1 achieved Harrell's C-indices of 0.73 (95% CI: 0.68-0.78) and 0.70 in the discovery and replication cohorts, respectively, with performance statistically comparable to a multiparametric radiomic pipeline (p>0.05). Adding molecular subtype information improved performance in the replication cohort (C-index: 0.68 vs 0.63, p=0.016) but not in the discovery cohort (0.79 vs 0.78, p=0.26). The multimodal model reclassified 18 patients in a manner consistent with established BRAF-associated prognostic biology. Uncertainty analysis showed a marked reduction in bootstrap confidence interval widths following late fusion with the molecular model, decreasing from a median of 2.37 to 0.57 in the discovery cohort and from 2.32 to 0.60 in the replication cohort, corresponding to relative reductions of 75.8% and 75.2%, respectively. CONCLUSIONS: Integrating molecular subtype into a clinical-imaging survival framework improved risk stratification in pLGG. Deep learning features extracted from T2-weighted MRI achieved performance comparable to a multiparametric radiomics pipeline without complex tumor segmentation. These findings support uncertainty-aware multimodal survival modeling as a streamlined and more transparent approach for pLGG risk stratification.