Interpretability of Reinforcement Learning Anticoagulation Dosing Policies Under Warfarin–Antibiotic Interaction Shifts: A Healthcare Safety Estimation

Reinforcement learning (RL) policies for warfarin dosing are trained on interaction-free transition dynamics, yet deployment coincides with antibiotic co-prescription that shifts the dose-response relationship. the present analysis formalises the resulting safety problem as a distributional shift in the Markov decision process and derives an interpretability audit that estimates, without observational data, the probability that a policy's recommended dose exceeds a clinician-defined safety envelope after the shift. The audit combines a linearised warfarin–antibiotic interaction model with a before-after-control-impact (BACI) design estimator whose operating characteristics are characterised under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64). Power was estimated across 2000 design cells for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points; 56 clusters are required for 80% power at a 5-percentage-point effect. The audit yields a policy-level safety score, a per-state attribution of the score to interaction-sensitive features, and a design-level bound on the smallest interaction effect the audit can detect. The contribution is a formal bridge between RL interpretability and cluster-level safety estimation, not a clinical trial result.