Diagrammatization and Abduction to Improve AI Interpretability With Domain-Aligned Explanations for Medical Diagnosis

Diagrams beat technical visualizations for medical AI: When medical students evaluated AI explanations for heart disease diagnosis, they strongly preferred clinically-relevant diagrams of heart murmurs over standard technical saliency maps, suggesting that explanations designed in domain experts' own language are more trustworthy and understandable.

AI can explain itself using expert reasoning: DiagramNet uses "abductive reasoning"—the same hypothesis-testing approach doctors use—to select the most likely diagnosis and explain it, rather than just showing which pixels the AI focused on, making the AI's decision-making process more transparent to medical professionals.

Better explanations can improve AI performance: The researchers found that DiagramNet not only provided clearer explanations than baseline models, but also achieved better diagnostic accuracy, suggesting that building AI systems with interpretability in mind from the start may actually make them work better.

One-size-fits-all XAI explanations fall short in high-stakes domains: Generic visualization techniques for explaining AI decisions often leave users confused and require additional reasoning to understand. For critical applications like medical diagnosis, explanations must be tailored to how domain experts actually think and communicate.

Trust requires alignment with expert knowledge: By incorporating medical ontology (the structured knowledge of how heart conditions relate to each other) and clinical reasoning processes into the AI system itself, the researchers created explanations that experts found more credible and easier to verify.