Multicenter retrospective study of offline reinforcement learning for insulin dosing during parenteral to enteral nutrition transition in ICU patients

The transition from parenteral nutrition (PN) to enteral nutrition (EN) in critically ill patients presents significant challenges in glycemic control, yet existing guidelines lack specific recommendations for managing insulin dosing during this period. In this multi-center retrospective study, we apply offline reinforcement learning (RL) to learn an insulin-dosing policy for the PN-to-EN transition, using a reward function defined by the clinically accepted target range of 80–180 mg/dL for blood glucose levels (BGLs). This study analyzes retrospective data from 826 ICU patients across three South Korean teaching hospitals, resulting in 1242 PN-to-EN feeding transitions. In the observed clinical data, only 31.8% of post-EN BGL measurements fell within the target range, whereas 67.95% exceeded 180 mg/dL, indicating that current clinical practice frequently fails to achieve adequate glycemic control during this transition. To address this gap, we model each transition event as a one-step dosing decision using a patient-specific state that incorporates the pre-EN BGL, minute-level carbohydrate exposure from concurrent PN and EN, and an insulin sensitivity indicator based on the insulin-to-glucose ratio. A reward function derived from the post-EN BGL measured 4 hours after EN initiation is used to train an offline actor–critic policy entirely from retrospective data, without requiring online patient interaction. Under leave-one-subject-out (LOSO) validation, the model achieved a directional consistency of 84.3%, a quadrant-based metric that evaluates whether the model’s recommended insulin adjustment aligns with the expected direction given the observed post-EN glycemic outcome. Performance was driven predominantly by hyperglycemic events (92.7%) and was lower in non-hyperglycemic events (66.6%). Sensitivity analyses showed stable performance across three patient-wise validating schemes (81.8–84.3%), but marked sensitivity to the reward specification (55.1–84.3% depending on the reward function). Representative case studies further illustrate model recommendations across diverse PN-to-EN scenarios; however, interpretation in hypoglycemia-prone settings remains limited by the small number of low glucose events. Formal off-policy evaluation was not performed; accordingly, this work should be regarded as a feasibility study rather than as evidence of clinical efficacy. To our knowledge, this is the first application of offline reinforcement learning to insulin dosing specifically during the PN-to-EN nutritional transition using real-world multi-center ICU data.