(T-057) Hybrid Population PK-Machine Learning Model Approach to Predict Infliximab Concentrations in Pediatric Patients with Crohn's Disease

Phillip Minar, na – ASSOCIATE PROFESSOR, Division of Gastroenterology, Hepatology and Nutrition, Cincinnati Children’s Hospital Medical Center; Jack Reifenberg, na – student, University of Cincinnati School of Medicine; Tomoyuki Mizuno, na – ASSOCIATE PROFESSOR, Division of Translational and Clinical Pharmacology, Cincinnati Children’s Hospital Medical Center

Objectives: Bayesian estimation using a population pharmacokinetic (PK) model is a valuable tool for individualizing dosages, particularly when faced with limited sample availability. However, the accuracy of such predictions can be compromised by various factors, including misspecification in prior information and uncertainties in PK changes. In this study, we developed a hybrid machine learning (ML) and population PK-based Bayesian approach to enhance the predictive performance of infliximab (IFX) concentrations using the Bayesian estimation in pediatric patients with Crohn's Disease.

Methods: Residuals between predicted IFX trough concentrations by Bayesian estimation using a pediatric IFX population PK model [1] and observed concentrations were computed using real-world data from pediatric patients with Crohn’s disease (292 concentrations from 93 patients). To predict the degree of residuals, we employed various ML algorithms, including linear regression, random forest, support vector regression, neural networks, and XGBoost, and evaluated their performance. Features for the ML models were identified based on the Pearson correlation coefficient (r) between the residuals and patient-related data, such as laboratory results, PK parameters, and dosing history. The predictive performance of the ML model-corrected Bayesian estimation was assessed using root mean square error (RMSE) and mean prediction error (MPE) with 5-fold cross-validation.

Results: For Bayesian estimation alone, the RMSE and MPE were 4.8 µg/mL and -0.7 µg/mL, respectively. We observed associations between the residuals and the observed IFX concentrations used for Bayesian estimation (i.e. prior trough) (r=-0.47, p< 0.001), the deviation to typical clearance (ETA for CL) (r = -0.140, p = 0.0168), erythrocyte sedimentation rate (ESR) (r = -0.15, p = 0.015), anti-IFX antibodies (r = -0.19, p=0.001), change in weight (r = -0.16, p = 0.005), and change in ESR (r = 0.23, p < 0.001) during the last dosing interval. Among the evaluated algorithms with selected features, the XGBoost model demonstrated the best predictive performance. The RMSE and MPE of the XGBoost model-corrected Bayesian estimation were 3.72 ± 0.88 µg/mL and -0.01 ± 0.54 µg/mL, respectively, with 5-fold cross-validation.

Conclusions: The ML model-corrected Bayesian estimation significantly improved predictive performance in predicting IFX trough concentrations in real-world data, surpassing the performance of Bayesian estimation alone. This hybrid population PK-ML approach may offer a framework for continually enhancing the predictive performance of Bayesian estimation through adaptive learning from new clinical data for dose individualization.

Citations: [1] Xiong et al., Real-World Infliximab Pharmacokinetic Study Informs an Electronic Health Record-Embedded Dashboard to Guide Precision Dosing in Children with Crohn's Disease. Clin Pharmacol Ther. 109(6):1639-1647, 2021