Artificial Intelligence Offline Reinforcement Learning for Insulin Titration in Type 2 Diabetes Under Variable Capillary Glucose Log Missingness: A Doubly Robust Evaluation of Healthcare Hypoglycaemia Safety

Offline reinforcement learning (RL) has been proposed as a method for deriving insulin titration policies from retrospective capillary glucose logs, but the safety of such policies is threatened by non-random missingness in those logs and by the absence of a principled estimator for hypoglycaemia risk under distribution shift. The framework combines an inverse-propensity-weighted (IPW) correction for log-missingness with a direct outcome model, yielding a doubly robust estimator of the hypoglycaemia rate under a candidate policy. A before-after-control-impact (BACI) design is used to characterise the estimator's power to detect clinically meaningful reductions in hypoglycaemia, with an intraclass correlation of 0.64, between-cluster SD of 6.0, and residual SD of 4.5. Power was estimated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design analysis indicates that 54 clusters are required to detect a 5-percentage-point reduction at 80% power. The contribution is a formal identification result for policy value under log-missingness, a doubly robust estimator with a stated influence function, and a design-characterisation that bounds the sample sizes at which offline RL insulin titration can be evaluated safely. The framework does not replace clinical validation; it specifies the conditions under which such validation is statistically feasible.