Abramson Cancer Center Support Grant

University Of Pennsylvania, Philadelphia PA

Investigators

Linked publications & trials

Abstract

PROJECT SUMMARY/ABSTRACT Pediatric brain tumors (PBTs) are the leading cause of disease-related death in children in the United States, with low grade glioma being the most frequent diagnosis. The Children’s Brain Tumor Network (CBTN) represents a multi-institutional data and biosample sharing framework operating a standardized IRB-approved observational protocol with more than 5,000 enrolled participants and families. Complementing longitudinal data collection have been large-scale, NCI-sponsored molecular datasets underpinning one of the largest multimodal pediatric brain tumor cohorts recently released as part of the Childhood Cancer Data Initiative and the Gabriella Miller Kids First Data Resource, a data coordination center at the Children’s Hospital of Philadelphia. Despite these successes, challenges still exist for the rapid exchange and use of medical imaging data, as well as the association of imaging with other data modalities, and there remains the unmet need for scalable infrastructure to enable predictive analytics with imaging and multi-modal pediatric brain tumor data. Consequently, the translation of clinically acquired MRIs to research and AI/ML model building for predictive modeling of PBTs has been limited. The proposed work will establish a cloud-based federated learning (FL) infrastructure that leverages CBTN’s existing multi-institutional network of >30 hospitals and existing Kids First Data Resource cloud platforms for imaging, clinical, and genomic data, and implement a multi-modal prediction framework for progression-free survival risk in pLGG as a proof-of-concept. An open-source, extensible FL framework will be utilized with data privacy measures and user-friendly workflows for data curation and the association of subject-level data across modalities. The FL framework will be assessed for stability and generalizability in the context of real-world healthcare data collected clinically, as well as technical and model development factors that could potentially limit its broader use. This includes support for federated statistics to quantitatively characterize dataset heterogeneity between institutions (e.g., differences in data quality due to site-level imaging and molecular data acquisition protocols). Overall, the project’s goal is to equip research investigators with the ability to perform multi-institutional data analysis with state-of-the-art federated methods in a secure, high-computing, high-throughput environment, and minimize the related technical and operational burdens met by scientific teams. Through these efforts, the community will be better able to bridge advanced multi-modal analytics with the pediatric neuro-oncology research setting and to ultimately develop personalized treatments for better outcomes in children with cancer.