Qurated: From ML Predictions to Informed Diagnostic Assistance Using the Toulmin Model of Argumentation
Why Your AI Diagnosis Tool Needs an Argument, Not Just an Answer
A prediction is not an argument. When an ML model tells a clinician "diabetic retinopathy: 87% confidence," it hands over a conclusion with no reasoning attached—and confidence scores are not reasoning. This is the central failure mode of clinical AI: we've optimized for accuracy while starving the human expert of the structure they need to actually evaluate that accuracy in a specific case.
The fix isn't more explainability heatmaps. It's argumentation.
The Core Insight: Borrow Structure From 2,000 Years of Logic
Stephen Toulmin's model of argumentation—built for law and rhetoric, not machine learning—maps almost perfectly onto what a diagnostic AI should be saying:
- Claim: The diagnosis itself ("this retina shows diabetic retinopathy")
- Grounds: The evidence—specific biomarkers detected in the image
- Warrant: The medical reasoning connecting evidence to diagnosis
- Qualifier: How confident we should be, and why
- Rebuttal: What would make this diagnosis wrong
- Backing: The domain knowledge underwriting the warrant
A single confidence score collapses all six of these into one number. That's not interpretability—it's compression that destroys the information a clinician needs most.
The Framework: Divide Labor, Not Just Compute
The paper's architecture assigns each Toulmin component to a specialized agent, rather than asking one model to do everything:
- A biomarker-extraction model generates the grounds — concrete visual evidence, not abstract saliency maps.
- A medically-tuned language agent (MedGemma) supplies the warrant — it reasons about why those biomarkers support the claim, using medical knowledge rather than pattern-matching alone.
- The qualifier emerges from evaluating both models jointly — confidence isn't a single softmax output; it's a function of how reliable the grounds and the warrant are together.
- A rebuttal is constructed automatically via image similarity (MedSigLip) — the system actively surfaces similar-looking cases that had a different diagnosis, forcing consideration of the counterargument.
The output isn't a label. It's a structured case file the clinician can interrogate.
The Mental Model: Assistance vs. Automation
Most clinical AI implicitly asks: "Should the machine replace the judgment?" This framework asks a better question: "How do we make the machine's judgment inspectable?"
Think of it as the difference between a black box and a brief. A brief doesn't just state a verdict—it presents evidence, reasoning, and the strongest counterargument, so a judge can rule with full context. That's the standard clinical AI should meet, not "trust the number."
Why This Generalizes Beyond Retinal Imaging
The Toulmin decomposition isn't domain-specific. Any high-stakes ML prediction—radiology, credit scoring, content moderation—can be forced through the same six-part structure:
- What's the claim?
- What's the visible evidence (grounds)?
- What's the reasoning connecting evidence to claim (warrant)?
- What would falsify this (rebuttal)?
If your AI system can't answer these four questions, it's producing outputs, not arguments—and outputs alone shouldn't drive decisions where being wrong costs something.
The Actionable Takeaway
Before deploying any high-stakes predictive model, force it through the Toulmin test: separate the what (claim) from the why (warrant) from the how sure (qualifier) from the what if I'm wrong (rebuttal). If your architecture can't generate these as distinct, inspectable outputs, you've built a prediction engine—not a decision-support system.