AI-007 - MACHINE LEARNING MODELING TO PREDICT INFLIXIMAB PHARMACOKINETICS IN PEDIATRIC PATIENTS WITH CROHN'S DISEASE

K. Irie¹, P. Minar^2,3, T. Mizuno^1,3; ¹Division of Translational and Clinical Pharmacology, Cincinnati Children's Hospital Medical Center, , , ²Division of Gastroenterology, Hepatology and Nutrition, Cincinnati Children's Hospital Medical Center, , , ³Department of Pediatrics, University of Cincinnati College of Medicine, , .

Background: Proactive therapeutic drug monitoring (TDM) has been shown to improve infliximab (IFX) treatment outcomes in pediatric inflammatory bowel disease (IBD). Bayesian estimation using a population (Pop) pharmacokinetic (PK) model is a useful approach to predicting IFX PK. Emerging research has shown that a hybrid algorithm combining machine learning (ML) and PopPK/pharmacodynamic (PD) modeling may provide better predictive performance than the conventional Bayesian approach. In this study, we developed an ML model to predict IFX PK using simulated and real-world data from children with Crohn’s disease (CD) to find an improved way to predict IFX trough concentrations.

Methods: The XGBoost model for predicting IFX concentrations was trained with simulated PK data (n=240,000) using R. The simulations were performed using a published pediatric IFX PopPK model (Xiong et al. 2021 CPT). Variability in clearance, weight, dose, dosing interval, and time at predicted concentration were considered for the simulations. The model was evaluated using newly simulated data (n=14,000) and real-world data (292 concentrations in maintenance phase) from 93 children with CD. The root mean square error (RMSE) and mean prediction error (MPE) were used for evaluation. The ML model’s performance was compared with the Bayesian approach. The analysis was completed in October 2023 .

Results: For the ML model, the RMSE and MPE were 1.3 µg/mL and < 0.001 µg/mL for the training and 1.4 µg/mL and 0.002 µg/mL for the test data, respectively. The RMSE and MPE for the newly simulated data were 1.4 µg/mL and -0.020 µg/mL, respectively, and relatively better than those for the Bayesian approach (RMSE=2.5 µg/mL, MPE=-0.7 µg/mL). The validation with real-world data indicates the comparable performance between the ML model (RMSE=5.7 µg/mL, MPE=1.0 µg/mL) and the Bayesian approach (RMSE=4.8 µg/mL, MPE=-0.7 µg/mL).

Conclusion: The ML model has been developed to predict IFX PK in children by leveraging simulated and real-world data. The validations demonstrated that the model adequately predicts the IFX PK using dose, dosing interval, weight, and an observed trough concentration. This study provides a foundation to further develop a hybrid PopPK-PD/ML strategy to facilitate model-informed precision dosing of IFX in pediatric IBD.