O-23: Graded-evidence reasoning AI with calculators and education — a deterministic, compounding-evidence AI diabetes educator for type 1 diabetes, validated via 1.5 million patient-days of real-world data
J Pemberton, A Kohli, P Hayes, F DebongAbstract
Introduction
Published AI for diabetes education targets type 2 diabetes with no safety framework, evidence grading or real-world validation. No system combines a structured evidence-graded knowledge base with deterministic educational tools for type 1 diabetes.
Objectives
To build and validate a dual-component educational system (evidence-graded wiki plus six deterministic calculators) for type 1 diabetes across the paediatric to adult lifespan.
Methods
We built a version-controlled clinical wiki (67 concept pages, 453 source summaries, 900+ evidence items) graded A to E, ingesting ISPAD 2024 and ADA 2026. Six paired deterministic calculators cover activity, exercise insulin-on-board, hypoglycaemia and hyperglycaemia treatment, AID setting review, exercise planning and alcohol. Both were validated against a real-world type 1 diabetes dataset (10,000 individuals, 2013 to 2025, 1.5 million patient-days) through 33 pre-specified assessments. Safety used 77 general, 18 adversarial and 12 jailbreak cases.
Results
Of 33 assessments: 18 confirmed (10 fully, 5 with nuance, 3 partial), 8 reversals, 7 deferred or not completable; 17 reversals of assumptions logged overall. Key reversals: step benefit plateau at 4,000 to 5,000 (not 10,000); sleep regularity correlated more strongly with time in range than exercise volume (77.0% vs 71.3%, n=251, p<0.001); female pen users with high exercise had the worst time below range (4.3% vs 2.2% in males, n=694, p<0.001). Safety: 0/77 genuine failures; 18/18 adversarial and 12/12 jailbreak attempts resisted.
Conclusions
A dual architecture pairing an evidence-graded knowledge base with deterministic educational calculators, validated at population scale, surfaced contradictions to clinical assumptions. A paired consultancy stream funds free public access (Robin Hood model). It offers a responsible-AI template for paediatric and adult type 1 diabetes education. Next: paediatric cohort ingest and external validation.