Fast screening of covariates in population models empowered by machine learning

Area of Science:

• Pharmacometrics
• Computational Statistics
• Machine Learning

Background:

• Population modeling in pharmacometrics (PMX) aims to identify significant covariate relationships.
• Selecting relevant covariates is a complex and time-consuming task in PMX.

Purpose of the Study:

• To evaluate the utility of supervised machine learning (ML) algorithms for covariate selection in PMX.
• To compare the performance and efficiency of ML methods against traditional PMX approaches.

Main Methods:

• Applied three ML methods (Random Forest, Neural Networks, Support Vector Regression) to NONMEM empirical Bayes estimates.
• Assessed ML model performance using Receiver Operating Characteristic (ROC) curves and F1 scores.
• Compared ML methods with stepwise covariate modeling (SCM) and COSSAC using simulated pharmacokinetic data.

Main Results:

• ML methods achieved comparable or superior F1 scores to SCM and COSSAC.
• ML methods demonstrated significantly faster computational speeds compared to SCM.
• Effect size of covariates impacted performance, but correlations did not.

Conclusions:

• ML methods enhance the efficiency of population covariate model building, especially for large datasets or complex models.
• ML provides rapid initial covariate screening, complementing traditional PMX for clinical relevance assessment.
• These findings support the integration of ML into PMX workflows for improved model development.