Insulin4RL: Real-Time Insulin Management in the Intensive Care Unit for Offline Reinforcement Learning

A research team has introduced Insulin4RL, a new healthcare dataset designed for offline reinforcement learning that preserves the irregular timing of real clinical data, according to a paper submitted in 2026 [1]. The dataset is derived from MIMIC-IV, a publicly available critical care database, and contains more than 375,000 labelled decisions across 12,209 patients who required insulin infusion titration in an intensive care unit [1][2]. Unlike many existing electronic health record datasets used in machine learning, Insulin4RL retains the naturally irregular inputs and actions found in actual clinical trajectories. The authors argue that the common practice of discretising records into fixed, regular time intervals creates fictional representations of complex clinical scenarios and undermines the generalisability of retrospective model evaluations [2]. The paper provides baseline performance metrics using model-free offline reinforcement learning and establishes a standardised evaluation protocol based on fitted Q-evaluation [2]. The work was submitted to arXiv on 17 June 2026 [1]. The release of task-specific clinical datasets has become a focus in applied machine learning, where researchers examine whether new data complements existing resources for training models [4]. The authors of the Insulin4RL paper suggest several areas for future research that could be addressed using the resource, though specific directions were not detailed in the abstract [2].