PaCE-RL: Context-Aware Reinforcement Learning for Personalized Glycemic Control in ICU Nutrition Transition

Parenteral-to-enteral nutrition (PN-to-EN) transitions in ICU patients cause abrupt, time-delayed glucose fluctuations, making existing insulin protocols and personalized treatments unreliable. We propose a novel Patient Context Encoder (PaCE) that generates embeddings for a downstream reinforcement learning (RL) policy to recommend insulin doses that keep post-EN glucose within 80-180 mg/dL. PaCE builds a context embedding by first combining static risk factors and clinical interventions via risk-conditioned modulation to form a modulated sequence, then learning delayed and cumulative temporal effects using learnable-offset and N-step convolutions, and finally fusing the outputs with attention. The embedding, together with transition features, forms the RL state. PaCE-RL outperforms baseline RL and state-of-the-art methods when evaluated on 15,562 patients, and matches physician doses in 83% of cases.