From ML Predictions to Informed Diagnostic Assistance Using the Toulmin Model of Argumentation

Diagnosis requires more than just predictions: Rather than relying solely on a machine learning model's diagnosis claim, the framework decomposes the diagnostic process into six structured components (claim, grounds, warrant, qualifier, rebuttal, and backing), enabling clinicians to critically evaluate each step rather than accepting the prediction at face value.

Explainability alone may be insufficient: The paper suggests that traditional explainable AI (XAI) methods have limitations, proposing instead an argumentation-based approach that provides a more comprehensive assessment by incorporating medical knowledge, confidence levels, and potential counterarguments to the diagnosis.

Multiple specialized models improve trustworthiness: The framework combines different AI components—biomarker extraction, medical knowledge reasoning (via MedGemma agent), and image similarity analysis (via MedSigLip)—rather than relying on a single model, allowing clinicians to verify the diagnosis through multiple independent lines of evidence.

Confidence levels matter as much as predictions: The "qualifier" component, which reflects the overall quantitative evaluation of the model's grounds and reasoning, is presented to experts, emphasizing that clinicians should know not just what the AI recommends but how confident it should be in that recommendation.

AI should support human judgment, not replace it: By presenting structured argumentation components including potential rebuttals (counterarguments), the system is designed to facilitate informed expert assessment rather than automate diagnostic decisions, keeping the clinician in an active, critical role.