ID #56 Can AI inform decision-making in pediatric neuro-oncology?
Pournima Navalkele, Trevor McGuireAbstract
Background
Pediatric central nervous system (CNS) tumors are the leading cause of cancer-related mortality in children. Given their complexity, rarity, and clinical variability, designing individualized treatment plans is challenging. Artificial intelligence (AI) and large language learning models (LLMs) have the bandwidth to integrate neuroimaging, tumor genomic profiling, and patient-specific data to curate personalized treatment recommendations, thus supporting clinical decision-making and potentially improving the outcomes for pediatric CNS tumors.
Methods
Patients were consented to include their de-identified data. A retrospective chart review was conducted for three different pediatric tumor types: [1] diffuse midline glioma, H3K27M-altered (highly-aggressive, poor prognosis), [2] posterior fossa ependymoma, PFA (highly-recurrent, poor prognosis), and [3] sacrococcygeal chordoma (good prognosis). The primary outcome measure was accuracy of AI-proposed treatment recommendation compared to the standardized treatment protocol for the respective tumor-type. Three LLMs representing both deep-learning (Microsoft Copilot, ChatGPT-5) and short-form (Perplexity AI) generative software were selected based on popularity and ease of access.
Results
An analysis of each tumor case revealed that all three LLMs can generate a basic treatment outline, similar to the standardized protocol. While Microsoft Copilot and ChatGPT-5 were able to identify specific radiotherapy and chemotherapy plans, Perplexity AI struggled to generate treatment plans of similar specificity. For recurrent or progressive tumors, LLMs could not devise personalized treatment strategies.
Conclusion
Taken together, deep-learning LLMs, such as ChatGPT-5, can dependably generate treatment outlines for pediatric CNS tumors. LLMs provide a potentially significant benefit to pediatric neuro-oncologists in supporting clinical decision-making, and to families by providing education and insight regarding treatment planning. Further studies are warranted to validate these findings with higher case volumes, recurrent and progressive tumors, and newer iterations of LLMs. In the future, clinicians may need to consider the social implications of AI adoption on the ever-evolving patient-provider dynamic.