Researchers propose argumentation-based framework to structure ML-driven retinal diagnosis for human expert review

The paper introduces a framework that structures image-based diagnostic predictions using the Toulmin model of argumentation, which comprises six components: claim, grounds, warrant, qualifier, rebuttal, and backing. In this setup, a machine learning model generates a diagnostic claim, while a specialized biomarker extraction model provides the grounds. A MedGemma agent, equipped with medical knowledge, analyzes the warrant that links the grounds to the claim. The qualifier is derived from quantitative evaluations of both the warrant and grounds models, and a rebuttal is constructed using image similarity measures computed with MedSigLip. The authors argue that presenting these components to human experts enables a more informed and critical assessment of ML-generated diagnoses.