Simulating realistic patient profiles from pharmacokinetic models by a machine learning postprocessing correction of...

Area of Science:

• Pharmacokinetics
• Machine Learning
• Computational Biology

Background:

• Population pharmacokinetic (PopPK) models are crucial for drug development.
• Model misspecification, particularly residual unexplained variability (RUV), can limit the accuracy of PopPK models.
• Current methods for assessing model quality may not fully capture the extent of misspecification.

Purpose of the Study:

• To develop and validate a machine learning (ML) based postprocessing method to address model misspecification in PopPK.
• To generate realistic virtual patient profiles by correcting individual predictions.
• To introduce a metric for quantifying the degree of model misspecification.

Main Methods:

• Individual residual errors (IRES) from a PopPK model were modeled using supervised ML algorithms.
• Random Forest was identified as the optimal ML algorithm.
• A novel metric based on R-squared between IRES and ML-predicted IRES (IRES_ML) was developed to quantify misspecification.

Main Results:

• The ML postprocessing step successfully corrected individual predictions, as shown in diagnostic plots.
• Realistic virtual patient profiles were generated, effectively mitigating artifacts from elevated RUV, even in misspecified models.
• The R-squared metric correlated with the extent of model misspecification, validating its utility.

Conclusions:

• ML methods offer a powerful approach to enhance PopPK model quality by addressing RUV and misspecification.
• The proposed methodology provides a practical tool for generating reliable virtual patient data.
• The developed metric serves as a valuable diagnostic for assessing PopPK model adequacy.