Exploring Large Language Models in Glioma Molecular Subtyping

New findings in European Radiology suggest large language models can extract molecular-subtype information from routine MRI report text, delivering preoperative molecular cues without additional testing.

In a multicenter retrospective series of adult-type diffuse glioma reports, knowledge-based prompting substantially improved molecular-subtype prediction compared with naïve prompting, offering a practical pathway to augment diagnostic workflows.

The team evaluated MRI report text from adult diffuse glioma cases and compared prompt strategies head-to-head, using molecular-subtype prediction accuracy as the primary endpoint. The cohort spanned multiple centers and common diagnostic categories, and the analysis showed that prompt content materially altered model outputs across tested LLMs.

Clinically, models performed markedly better with knowledge-based prompts and could extract actionable molecular markers from report language, including IDH status and related descriptors embedded in narrative findings. Larger or better-tuned models were less sensitive to prompt detail, while smaller or less-capable models gained most from explicit knowledge injection—indicating that LLMs can surface molecular signals already present in reports.