Abstract
Model-informed precision dosing (MIPD) utilizes pharmacokinetic/pharmacodynamic (PK/PD) models to optimize drug therapy. However, conventional MIPD often requires manual simulation and regimen selection, which are time-consuming and demand specialized expertise. Reinforcement learning (RL), in which an agent learns optimal decisions through iterative interactions with an environment, offers a scalable and automated alternative. In this study, we developed a model-informed Deep Q-Network (DQN) to personalize infliximab dosing for patients with Crohn's disease. The DQN was trained in a simulation environment incorporating a population PK model, inter-individual variability, and assay error. Virtual patients with randomly and independently sampled covariates from log-normal distributions were used to explore dosing strategies at Infusions 1, 3, and 4. Doses ranged from 1 to 10 mg/kg at Infusion 1 and from 1 to 20 mg/kg thereafter, with intervals of 4-12 weeks. The reward function prioritized achieving trough concentrations of 18-26 μg/mL before Infusion 3 and 5-10 μg/mL before Infusions 4 and 5, while penalizing overtreatment and additional infusions. The DQN policy converged after 80,000 episodes, yielding target attainment probabilities (PTAs) of 92.9% and 98.4% at Infusions 4 and 5, respectively, in 1000 virtual patients. High doses (11-20 mg/kg) were selected in only 0.2% of cases. At Infusion 4, 66.8% of patients received an 8-week interval, and 57.3% at Infusion 5. Retrospective real-world validation showed that patients whose actual doses matched DQN recommendations had trough levels significantly closer to target ranges. These findings support the feasibility of using DQN-based agents to enhance and automate infliximab individualized dosing in pediatric populations.