Accounting for Insulin-on-board in Large Language Model Advice about Pediatric Hyperglycemia

LLMs tested for pediatric type 1 diabetes caregiver questions sometimes give insulin-dosing advice without warning about insulin-on-board, with safety-failure rates varying by meal and insulin context.

• Why it matters: Caregivers may rely on LLMs for hyperglycemia guidance, but the safety risk of missing insulin-on-board warnings is not well characterized for pediatric type 1 diabetes.
• What they did: The study evaluated four LLMs via programmatic endpoints and two via chat interfaces, measuring how often outputs advised dosing insulin without clear insulin-on-board accounting under different meal and insulin-history context.
• The result: Potential safety failures occurred across all six LLMs, and the frequency changed with added context, implying patient-facing use needs guardrails to reduce insulin-on-board-related errors.

AI summary of the abstract · may contain errors · not medical advice · read the paper.