Machine Learning Modeling for Predicting Infliximab Pharmacokinetics in Pediatric and Young Adult Patients With Crohn Disease: Leveraging Ensemble Modeling With Synthetic and Real-World Data

Background: Predicting infliximab pharmacokinetics (PK) is essential for optimizing individualized dosing in pediatric patients with Crohn disease (CD). Machine learning (ML) has emerged as a tool for predicting drug exposure; however, its development typically requires large datasets. This study aimed to develop an ML model for infliximab PK prediction by leveraging population PK model–based synthetic and real-world data. Methods: An initial ML model was trained using the XGBoost algorithm with synthetic infliximab concentration data (n = 560,000) generated from an established pediatric PK model. The prediction errors were assessed using real-world data, including 292 plasma concentrations from 93 pediatric and young adult patients with CD. A second XGBoost model, incorporating clinical features, was used to correct these errors. The performance of the model was evaluated using the root mean square error (RMSE) and mean prediction error (MPE). Results: The first ML model yielded RMSE and MPE values of 6.44 and 1.84 mcg/mL, respectively. The features of the second XGBoost model included the predicted infliximab concentrations, cumulative dose, and dosing interval duration. A 5-fold cross-validation demonstrated improved performance of the ensemble model (RMSE = 4.30 ± 1.09 mcg/mL, MPE = 0.21 ± 0.39 mcg/mL) compared with the initial model and was comparable with the Bayesian approach (RMSE = 4.81 mcg/mL, MPE = −0.67 mcg/mL). Conclusions: This study demonstrated the feasibility of combining synthetic and real-world data to develop an ML-based approach for infliximab PK prediction, potentially enhancing precision dosing in pediatric CD.