Toulmin Argumentation Framework Boosts Interpretability of Retinal ML Diagnoses

A new paper from researchers Anca Marginean and Adrian Groza (arXiv:2607.09664) tackles the black-box problem in medical AI by applying the Toulmin model of argumentation to retinal diagnosis. Rather than accepting a machine learning model's prediction at face value, their framework decomposes the diagnostic process into six interpretable components: claim (the diagnosis), grounds (biomarker extraction from images), warrant (why biomarkers support the claim), qualifier (confidence based on quantitative evaluation), rebuttal (counterarguments using image similarity), and backing (broader medical knowledge). This structure transforms a single ML output into a transparent, argument-driven assessment.

The architecture uses specialized AI agents for each component. A biomarker extraction model provides the grounds. A MedGemma agent—equipped with medical knowledge—analyzes the warrant linking those grounds to the claim. The qualifier is determined by overall quantitative evaluation of both warrant and grounds models. Crucially, rebuttals are constructed using MedSigLip, a model that computes image similarity to challenge the original diagnosis. All components are presented to the human expert, enabling a more informed and critical assessment. This approach moves beyond traditional XAI methods by providing a structured argumentation framework, making it easier for clinicians to trust or contest AI-driven diagnoses.