shayhan-ameen/glycemic_control_Pn2En: Development of a Predictive Model for Glycemic Control in Critically Ill Patients Transitioning from Parenteral to Enteral Nutrition

The transition from parenteral nutrition (PN) to enteral nutrition (EN) in critically ill patients presents significant challenges in glycemic control, compounded by the absence of comprehensive clinical guidelines. To address this, we introduce a predictive model leveraging reinforcement learning, specifically an Offline Actor-Critic network, to optimize insulin dosing during this critical phase. This study analyzes retrospective data from 826 ICU patients across three South Korean teaching hospitals, resulting in 1,242 PN-to-EN feeding transitions. The research is conducted with the approval of the institutional review boards at each participating site, ensuring compliance with ethical standards, including patient consent and data anonymization. The model focuses on a crucial 48-hour window surrounding EN initiation, dynamically adjusting insulin doses to maintain blood glucose levels within the 80–180 mg/dL target range. Achieving an accuracy of 84.3%, the model demonstrates substantial efficacy across a diverse patient cohort, successfully managing 1,033 hyperglycemic events and correctly intervening in 14 hypoglycemic episodes. Its robust performance highlights its adaptability to complex metabolic dynamics and clinical utility in refining glycemic management strategies. Moreover, its minimal computational demands facilitate seamless integration into existing clinical systems, enhancing its practical viability for real-world deployment. This approach establishes a new benchmark for AI-driven decision support in critical care, optimizing metabolic control with high precision and efficiency. Full Changelog: https://github.com/shayhan-ameen/glycemic_control_Pn2En/commits/version-1