Multimodal integration of pathomics and clinical features improves recurrence risk prediction in pediatric...

Accurate prediction of recurrence in pediatric medulloblastoma remains challenging with clinical variables alone. This study evaluated whether integrating deep learning (DL)-derived pathomic features could improve recurrence risk stratification.

Medulloblastoma patients aged below 18 years treated at Zhujiang Hospital during 2015-2025 were enrolled and allocated into training (70%) and validation (30%) cohorts. Quantitative tissue features were extracted from hematoxylin-eosin stained sections using ResNet-18 architecture. Three XGBoost-based models were constructed: clinical, pathomics, and multimodal. Predictive performance was assessed through receiver operating characteristic (ROC) and precision-recall curve (PRC) analyses. Prognostic stratification was validated via Kaplan-Meier estimation with log-rank comparison.

The cohort comprised 109 patients, with 33 (30.3%) developing recurrence. The multimodal framework demonstrated superior discriminative performance in validation, achieving an area under the curve (AUC) of 0.891 (95% CI: 0.770-0.981) with average precision of 0.775 (95% CI: 0.508-0.962). The model effectively stratified patients into distinct prognostic subgroups, with three-year recurrence-free survival of 100% in low-risk versus 45.5% in high-risk groups (P < 0.001).

Integrating DL-derived pathomic features with clinical variables significantly improves recurrence risk stratification in pediatric medulloblastoma. Enhanced monitoring of high-risk patients may facilitate early recurrence detection and timely salvage therapy.

A multimodal framework integrating deep learning-derived pathomics features from H&E-stained slides with clinical parameters significantly improves recurrence risk prediction in pediatric medulloblastoma (AUC = 0.891), outperforming clinical-only and radiomics-based models. This study demonstrates that routine histopathology slides contain quantifiable prognostic information beyond conventional microscopic assessment, offering potential as an accessible and objective complementary tool for risk stratification. The model effectively stratifies patients into distinct risk groups (3-year RFS: 100% vs 45.5%), potentially facilitating personalized surveillance strategies-intensified monitoring for high-risk patients and reduced imaging burden for low-risk patients.