This paper addresses the critical challenge of predicting future insulin dosages for diabetic patientsin Intensive Care Units (ICUs), a setting where precise glycemic control is crucial. The work proposes aunique methodology that sidesteps the complications associated with traditional glucose-insulin interactionmodels, due to data sparsity. We introduce a simplified version of Bergman’s Insulin-Glucose Interaction Modeland construct an extended dataset based on the MIMIC III database. This dataset includes 870 predictivefeatures encompassing demographic data, prior insulin administrations, and average glucose levels. The paperalso introduces a Reinforcement Learning approach, utilizing Deep Q-Learning, to optimize both the thetraining population and the feature selection for individualized predictions. We found that our CompositeMultilinear Regression Model algorithm outperforms the single-patient regression model in terms of MeanAbsolute Error (MAE). Specifically, the MAE values for the Type I and Type II Diabetic groups were 2.33 and3.68, respectively, significantly better than the single-patient regression model. The work contributes a novelapproach to insulin dose prediction, offering a promising pathway for more effective glucose management inICU settings.