Federated Learning Framework for Multimodal Survival Analysis in Pediatric Cancers

Overall Summary/Abstract Childhood cancers, although rare, significantly impact thousands of children globally each year. The complexity and variability of these cancers necessitate extensive research to understand their underlying mechanisms. Given the small number of cases and the diverse genetic and clinical profiles, collaboration across multiple institutions is crucial. The Beat Childhood Cancer (BCC) Research Consortium is a national collaborative group consisting of over 50 institutions and hospitals devoted to research and clinical trials in neuroblastoma, CNS tumors, sarcomas, and other rare solid tumors. Close to 1,000 tumors have been investigated for their genomic and transcriptomic profiles, clinical and imaging data, leading to the establishment of several clinical trials, including the recent FDA approval of DFMO for high-risk neuroblastoma. The BCC cohort is well-positioned to collaborate with other pediatric cancer research consortia and institutes to further the goal of eradicating childhood cancer. Federated learning (FL) offers a novel approach to facilitate such collaborations, allowing researchers to pool data and insights without compromising patient privacy and security. By enabling the analysis of distributed datasets, FL generates robust models and uncovers critical insights, leading to improved diagnostics and therapies for childhood cancers. In this P30 supplement project, we will focus on establishing an FL framework for neuroblastoma and pediatric central nervous system (CNS) tumors. Neuroblastoma is a highly heterogeneous cancer arising from multiple organ sites and primarily affecting young children. Despite advancements in understanding its genomic landscape, the prognosis for high-risk patients remains unfavorable. Pediatric CNS tumors are a diverse group of malignancies that occur in the brain and spinal cord, representing the second most common type of cancer in children. The brain's intricate structure and vital functions make these tumors particularly challenging to treat. The location of the tumor within the brain can greatly influence symptoms and treatment options, impacting critical areas responsible for movement, sensation, cognition, and other essential functions. Pediatric brain tumors encompass a wide range of subtypes, each with distinct biological behaviors and prognoses. Our central hypothesis is that a FL framework can enhance the integration and analysis of diverse pediatric cancer datasets, leading to the identification of multimodal survival predictors and therapeutic targets. We will test this hypothesis through three specific aims: Specific Aim 1: To finalize the implementation of a FL pipeline to securely centralize and analyze diverse datasets from multiple institutions. Specific Aim 2: To elucidate factors that influence survival in pediatric neuroblastoma using machine learning (ML) and FL. Specific Aim 3: To elucidate factors that influence survival in pediatric CNS tumors, including gliomas and medulloblastomas, using ML and FL. This study introduces a novel FL framework for pediatric cancer research, enabling secure, collaborative analysis of diverse datasets. This research promises to advance our understanding of pediatric cancers, improve patient outcomes, and pave the way for AI-driven decision-making applications in healthcare.