level: research
a new framework applies the toulmin model of argumentation to machine learning predictions for retinal diagnosis. instead of just giving a yes or no answer, the system explains its reasoning step by step. it starts with a claim from an ml model, then provides grounds from a biomarker extraction model. a medical knowledge agent called medgemma acts as the warrant, linking the grounds to the claim. the qualifier shows how confident the system is, based on how well the grounds and warrant models perform. a rebuttal is built using image similarity from medsiglip, pointing out possible counterarguments.
the approach moves beyond traditional explainable ai methods by structuring the explanation like a human argument. each part of the toulmin model has a specific role: the claim is the diagnosis, the grounds are the image features, the warrant is the medical reasoning, and the rebuttal challenges the claim with similar but different cases. this makes the ai's decision process more transparent and easier for clinicians to follow. the system does not just highlight image regions but provides a full logical breakdown.
by combining specialized models for each argument component, the framework aims to support diagnostic assistance rather than replace human judgment. the medgemma agent brings in medical knowledge, while medsiglip helps find visually similar images that could lead to different conclusions. this structured output could help doctors spot when the ai might be wrong and understand why. the work shows how argumentation theory can bridge the gap between black-box predictions and trustworthy clinical tools.
why it matters: it gives clinicians a clear, step-by-step reasoning for ai diagnoses, making it easier to trust and verify machine learning in medical settings.