Comparison of Machine Learning and traditional Population Pharmacokinetic of Gabapentin in Neuropathic Pain
Year: 2026
Type: conference-paper
Abstract: Objectives: Compare the machine learning (ML) algorithms - genetic algorithm (GA) and exhaustive search (EX) - for population parameters estimation and covariate identification, using a published gabapentin population pharmacokinetics (PopPK) model as reference. Methods: The application of EX and GA to identify optimal models was previously described.[ 1, 2] The reference PopPK model, developed using MONOLIX (2019R2, Lixoft, France), employed stochastic approximation expectation minimization algorithm (SAEM) and was best described by a 1-compartment model with first order absorption (Ka), lag time, linear elimination, correlation between clearance (CL) and volume (V) and the estimated glomerular filtration rate (eGFR) as a covariate on CL.[ 3] A simulated dataset was constructed based on the published parameters typical values and between subject variability (BSV): CL = 14.7 L/h (0.369), V = 140 L (0.402) , KA = 1.12 h-1 (0.494), lag time = 0.316 h (0.496), and CL-V correlation of 0.782. eGFR was included as a covariate on CL (0.611). Residual variability was modeled using a proportional error model (0.181). Individual demographic data was included in supplemental information. A total of 554 observations were included in the dataset. ML models were run via Pirana (version 24.9.2, Certara) and NONMEM (version 7.5.1) using first-order conditional estimation with interaction (FOCEI). The model space (n=6,144 models) included the following dimensions of candidate features: 1-compartment model, effect of covariates (weight, eGFR, HbA1c, neuropathic pain Group) on CL or V, BSV on Ka, V and CL, lag time, CL-V correlation and residual error models. The GA included a local “bit-flip” search, iterating 1- and 2-bit array changes until fitness no longer improved. The EX algorithm enumerates and assesses every combination to find the optimal solution. Model fitness was assessed by penalties added to the OFV for other non-optimal outcomes, including failure to converge, failed covariance step, high correlation in the covariance matrix (> 0.95), overparameterization and condition number > 1000. Results: The EX search identified the “true” best model: 1-compartment, lag time, first order absorption, linear elimination, weight as covariate on V. GA required a local 2-bit to identify the same model. Parameter estimates were close to the simulated model, despite structural differences. Conclusions: Although eGFR was a covariate on clearance in the reference model, this correlation was taken into consideration when simulating the dataset, therefore not included as covariate for the simulated or the ML models. The ML-identified model included weight on V, improving the model for fitness criterium used; however, it would not be considered in the traditional approach due to high relative standard error (RSE%) for the covariate estimate (>100%.) This comparison highlights how modeling assumptions, covariate parametrization and fitness criterium can influence final model structure in the ML-based PopPK. ML-driven methods can efficiently explore large model spaces and identify plausible population pharmacokinetic models; however, evaluation of the final as being fit for purpose (e.g., reasonable RSE values) remains essential.
DOI: 10.70534/hxld2949
Language: English
Field: Medicine
Domain: Health Sciences
Open Access status: gold
OpenAlex ID: W7131290396