Cluster analysis: an alternative method for covariate selection in population pharmacokinetic modeling

Area of Science:

• Pharmacometrics
• Systems Biology
• Data Science

Background:

• Biological data often exhibits heterogeneity, necessitating models with numerous variables.
• High variable numbers can complicate identifying significant combinations to reduce inter-individual variability (IIV).
• Current methods may involve multiple single-variable introductions, which can be inefficient.

Purpose of the Study:

• To introduce a multivariate variable strategy using cluster analysis for reducing IIV.
• To stratify patient populations into homogeneous groups based on multiple covariates.
• To improve the interpretability and performance of pharmacokinetic models.

Main Methods:

• Exploratory multivariate analysis using cluster analysis to create a multivariate categorical covariate.
• Application of Euclidean distance and complete-linkage algorithm for stratification.
• Pharmacokinetic modeling using NONMEM for cortisol kinetics in 82 patients.

Main Results:

• Cluster analysis identified five distinct clusters based on BMI, obesity duration, and waist-hip ratio.
• These clusters were further grouped into three sub-populations representing non-obese, intermediate, and extreme obese individuals.
• The three-cluster model demonstrated improved performance and interpretability compared to the basic model, similar to a classical covariate model.

Conclusions:

• Cluster analysis provides an effective stratification strategy for biological data with high dimensionality.
• This approach enhances the interpretability of pharmacokinetic models by identifying distinct patient sub-populations.
• Cortisol stimulation and elimination rates vary significantly across different obesity strata.