The Toulmin model for auditing medical AI diagnoses

A machine learning model that analyzes a retina and returns a label such as diabetic retinopathy is right much of the time, but it does not explain why it makes that claim or what would make the system change its mind. A group of researchers proposes in arXiv an architecture that breaks that diagnosis into six parts, the same ones Stephen Toulmin defined in 1958 to analyze any argument: claim, grounds, warrant, qualifier, rebuttal and backing.

The idea is simple to state and uncommon in clinical practice: instead of accepting the label the network produces, the system presents it as a claim that must be supported with structured evidence. The paper From ML Predictions to Informed Diagnostic Assistance Using the Toulmin Model of Argumentation distributes each part of the argument across different models.

How the argument is divided

The grounds come from a model specialized in extracting biomarkers from the image: what is actually visible in the retina. The warrant, that is, the bridge between those biomarkers and the claim, is assessed by a MedGemma agent equipped with medical knowledge. The qualifier comes from the joint quantitative evaluation of the grounds and warrant models, and works as a measure of how much confidence the conclusion deserves. The rebuttal is built with image similarity measures computed with MedSigLip, looking for similar cases that could contradict the diagnosis.

The result is not a black box with a probability next to it, but a dossier: here is the claim, here the evidence that supports it, here the clinical reasoning that connects them, here the degree of certainty and here what would cast doubt on it. Everything is handed to the physician, who keeps the final word and decides whether to accept, qualify or discard the conclusion. The stated goal is not to automate the diagnosis, but to give the specialist enough material to criticize it.

Why it matters

Explainability in medical AI usually stops at heatmaps that point to which pixels the network looked at. They are useful, but they do not say whether the reasoning is correct, only where the focus was. The argumentative approach changes the question: instead of where did the model look, it proposes does this conclusion hold. A physician can review the warrant, challenge the qualifier or examine the rebuttal without interpreting the internal activations of a neural network. It is a subtle but decisive difference when a medical decision is at stake.

There is an architectural detail that is interesting beyond medicine. The system does not rely on a single huge model that does everything, but on several specialized models, each responsible for one part of the argument, coordinated by an agent that applies domain knowledge. It is the same pattern we see in agent deployments with Claude Code and MCP servers: splitting a complex task into bounded, verifiable roles instead of trusting a single answer.

Who it is useful for

It matters for teams building clinical decision support that hit the regulatory and trust barrier: a diagnosis a clinician cannot audit rarely reaches production. It also matters for anyone designing high risk systems where the traceability of the reasoning weighs as much as the accuracy. The paper is a proof of concept focused on retinal diagnosis, not a product, and the authors themselves frame it as a framework to assist the expert, not to replace them.

We find the skeleton more valuable than the specific case. Structuring a model's output as a reviewable argument, with its grounds and its rebuttal made explicit, is a discipline many LLM deployments would appreciate outside the medical field. It remains to be seen whether the cost of orchestrating four models pays off against lighter explainability methods, something this work does not yet resolve.