Linguistically Communicating Uncertainty in Patient-Facing Risk Prediction Models

Uncertainty communication requires different approaches for patients than for AI experts: Traditional explainability methods designed for developers and specialists are insufficient when communicating with patients. Healthcare AI systems need specially tailored natural language explanations that account for patients' different knowledge levels and information needs.

Four key communication challenges exist in patient-facing risk prediction: The paper identifies specific difficulties in communicating (1) model performance, (2) confidence levels, (3) reasoning behind predictions, and (4) "unknown knowns"—factors the model wasn't trained to recognize—all of which require careful linguistic framing for patient understanding.

Natural language presentation matters as much as accuracy: Beyond just making predictions, how uncertainty is linguistically presented and formatted significantly affects whether patients actually understand the information. The paper emphasizes that evaluating understandability is a critical but often-overlooked component of responsible AI in healthcare.

In-vitro fertilization outcome prediction serves as a concrete test case: The authors apply their framework to IVF outcome prediction, a high-stakes medical decision where patients need clear communication about both what the model can and cannot reliably predict, demonstrating real-world applicability of uncertainty communication principles.