Evaluation of machine learning methods for covariate data imputation in pharmacometrics

Area of Science:

• Pharmacometrics
• Clinical Research
• Data Science

Background:

• Missing data in clinical research reduces statistical power and can bias results.
• Handling missing covariate data is crucial for pharmacometric analyses, including population pharmacokinetic and pharmacodynamic studies.

Purpose of the Study:

• Introduce and evaluate two machine learning (ML) methods for imputing missing covariate data in pharmacometrics.
• Compare the performance of ML methods against established statistical imputation techniques.

Main Methods:

• Simulated missing data scenarios based on a prior pharmacometric analysis.
• Compared listwise deletion, mean imputation, standard multiple imputation (Norm), predictive mean matching (PMM), random forest (RF), and artificial neural networks (ANNs).

Main Results:

• Both traditional and ML-based imputation methods performed well, with some limitations.
• Predictive mean matching (PMM), random forest (RF), and artificial neural networks (ANNs) demonstrated the best performance with minimal bias.
• ML methods achieved results comparable to the established PMM method.

Conclusions:

• Machine learning methods are suitable for imputing missing covariate data in pharmacometric settings.
• ML approaches offer flexibility for complex, nonlinear relationships, especially with optimized parameters.
• ML methods provide a valuable alternative to traditional statistical imputation techniques.