A Hybrid Model Associating Population Pharmacokinetics with Machine Learning: A Case Study with Iohexol Clearance...

Area of Science:

• Pharmacometrics
• Machine Learning in Drug Development
• Clinical Pharmacokinetics

Background:

• Maximum a posteriori Bayesian estimation (MAP-BE) with limited sampling strategies can introduce uncertainty in individual pharmacokinetic parameter estimation.
• Combining population pharmacokinetic (PopPK) models with machine learning (ML) algorithms shows promise for enhancing estimation accuracy.

Purpose of the Study:

• To evaluate a hybrid ML/PopPK approach for improving individual iohexol clearance estimation.
• To assess the performance of ML algorithms in correcting MAP-BE estimates derived from limited sampling.

Main Methods:

• Developed Xgboost and glmnet models to predict estimation errors of MAP-BE clearance using limited sampling (0.1, 1, 9h).
• Trained models on simulated data and validated in an independent dataset and 36 real patients.
• Compared hybrid ML-corrected MAP-BE estimates against reference values derived from dense sampling.

Main Results:

• The hybrid approach reduced root mean squared error (RMSE) by 29% (glmnet) and 24% (Xgboost).
• Reduced the percentage of profiles with >20% prediction error by 60% (glmnet) and 40% (Xgboost).
• Observed a significant decrease in eta-shrinkage, indicating improved individual estimation precision.

Conclusions:

• The hybrid ML/PopPK algorithm offers a substantial improvement over MAP-BE alone for individual iohexol clearance estimation.
• This approach enhances accuracy and reduces bias in pharmacokinetic parameter estimation, particularly with limited data.