pyDarwin machine learning algorithms application and comparison in nonlinear mixed-effect model selection and...

Area of Science:

• Pharmacometrics
• Computational Biology
• Machine Learning Applications

Background:

• Forward addition/backward elimination (FABE) has been the traditional method for population pharmacokinetic (PPK) model selection.
• The need for more efficient and robust model selection techniques in PPK is growing.

Purpose of the Study:

• To evaluate five machine learning (ML) algorithms as alternatives to FABE for PPK model selection.
• To compare the efficiency and robustness of ML algorithms combined with local downhill search strategies.

Main Methods:

• Investigated Genetic Algorithm (GA), Gaussian Process (GP), Random Forest (RF), Gradient Boosted Random Tree (GBRT), and Particle Swarm Optimization (PSO).
• Combined ML algorithms with one-bit or two-bit local downhill searches for systematic feature exploration.
• Used an exhaustive search of 1,572,864 models as the gold standard for robustness.

Main Results:

• All ML algorithms, when paired with a two-bit local search, successfully identified the optimal PPK model.
• GA, RF, GBRT, and GP identified the optimal model using only a one-bit local search.
• Gaussian Process (GP) was the most efficient, examining 495 models, while Particle Swarm Optimization (PSO) was least efficient (1710 models). GP required the longest computation time (2975.6 min), whereas GA was fastest (321.8 min).

Conclusions:

• Machine learning algorithms, particularly Gaussian Process, offer efficient and robust alternatives to FABE for PPK model selection.
• The choice of ML algorithm and local search strategy significantly impacts efficiency and computational time in PPK model building.