Argtumour: integrating large language models and computational argumentation to discuss treatment options for high-grade glioma
OA Location
Author(s)
Type
Journal Article
Abstract
AIMS
High-grade gliomas are aggressive brain tumours with poor prognosis, and patients diagnosed with such tumours often face difficult treatment decisions. While medical guidelines can provide information on the available treatment options, they cannot address individual concerns or offer personalised recommendations. Large language models can provide AI-based language interpretation; Argumentation is a logic-based technique that provides explainability. In this study, we introduce ArgTumour, an interactive, explainable AI system that combines large language models (LLMs) and argumentation to support patient decision-making through summarising treatment options for glioblastoma (GBM) and their justifications.
METHODS
We extracted information on GBM management from NICE guidelines, evidence reviews, and a patient information document, using LLMs to identify treatment options and generate structured arguments for and against them. These arguments were evaluated for faithfulness using the IBM Granite Guardian 8B model, which assesses alignment with the original sources. A clinical expert reviewed the argument structures, and system performance was assessed by comparing confidence scores for NICE-supported vs. non-supported treatments. We have also started some early qualitative evaluation of the system through preliminary patient and public engagement sessions.
RESULTS
Our system identified 14 main treatment options for GBM, extracting 159 supporting and opposing arguments. The Granite Guardian model found 77% of arguments to be well-supported by the used sources, indicating good overall faithfulness. Among the 14 options, 6 aligned with NICE guidelines, with an average system confidence in these options of 73%, while non-NICE-supported options all received a confidence score of 0%.
CONCLUSION
Our findings highlight the potential for using LLM-based argumentation systems in medical decision support, providing more personalized and explainable recommendations. This approach allows us to mine medical guidelines to produce explainable arguments for and against different treatment options, where NICE recommended options are scored much more highly than those that are not.
High-grade gliomas are aggressive brain tumours with poor prognosis, and patients diagnosed with such tumours often face difficult treatment decisions. While medical guidelines can provide information on the available treatment options, they cannot address individual concerns or offer personalised recommendations. Large language models can provide AI-based language interpretation; Argumentation is a logic-based technique that provides explainability. In this study, we introduce ArgTumour, an interactive, explainable AI system that combines large language models (LLMs) and argumentation to support patient decision-making through summarising treatment options for glioblastoma (GBM) and their justifications.
METHODS
We extracted information on GBM management from NICE guidelines, evidence reviews, and a patient information document, using LLMs to identify treatment options and generate structured arguments for and against them. These arguments were evaluated for faithfulness using the IBM Granite Guardian 8B model, which assesses alignment with the original sources. A clinical expert reviewed the argument structures, and system performance was assessed by comparing confidence scores for NICE-supported vs. non-supported treatments. We have also started some early qualitative evaluation of the system through preliminary patient and public engagement sessions.
RESULTS
Our system identified 14 main treatment options for GBM, extracting 159 supporting and opposing arguments. The Granite Guardian model found 77% of arguments to be well-supported by the used sources, indicating good overall faithfulness. Among the 14 options, 6 aligned with NICE guidelines, with an average system confidence in these options of 73%, while non-NICE-supported options all received a confidence score of 0%.
CONCLUSION
Our findings highlight the potential for using LLM-based argumentation systems in medical decision support, providing more personalized and explainable recommendations. This approach allows us to mine medical guidelines to produce explainable arguments for and against different treatment options, where NICE recommended options are scored much more highly than those that are not.
Date Issued
2025-09-02
Date Acceptance
2025-03-31
Citation
Neuro-Oncology, 2025, 27 (Supplement_2), pp.ii26-ii26
ISSN
1522-8517
Publisher
Oxford University Press (OUP)
Start Page
ii26
End Page
ii26
Journal / Book Title
Neuro-Oncology
Volume
27
Issue
Supplement_2
License URL
Publication Status
Published
Date Publish Online
2025-09-02