Mihaly Leiwolf1,2,3, Francois Riglet1, Adrien Tessier4, Sylvain Fouliard1
1Quantitative Pharmacology, Translational Medicine, Servier, 2Inserm U1070 PHAR2, University of Poitiers, 3Inserm U1248 P&T, University of Limoges, 4Clinical Pharmacology, Translational Medicine; Servier
Introduction: Identifying clinically relevant relationships between covariates and parameters is a key step in population pharmacokinetics (PopPK) modeling. Integrating covariates into such PK models allow for optimizing dosing regimens for specific sub-populations. The reference method to select covariate is the stepwise covariate model (SCM) building process, which is slow and computationally heavy. Machine Learning (ML) is increasingly applied in pharmacometrics (PMx) for its ability to handle large datasets efficiently, making covariate selection faster and more effective[1, 2]. Objective: The objective of the work is to compare ML algorithms and current PMx approaches for covariate model selection based on simulated scenarios and to develop a ML based approach method complementary to the classical PMx methods. Methods: The evaluation is performed on a simple structural PopPK model with a first order absorption, a single distribution compartment and a first order elimination. Six ML methods are compared with SCM and its faster version, SCM+[3]: three methods based on Random Forest (RF) algorithm (RF, VSURF and Boruta), two penalized regression methods (Lasso and Ridge) and Artificial Neural Networks (ANN). A new method is developed based on the combination of three others: RF, Lasso and Ridge to increase precision. These methods are applied to target variables, i.e. observations to be described by ML methods, from the model: random effects ? and Empirical Bayesian Estimates (EBE) of volume (V) and clearance (CL) parameters. The primary endpoint of analysis is the sensitivity of evaluated methods, and the secondary endpoint is precision. A threshold function for normalized covariate importance scores is determined to maximize the performance criteria of each method. The different methods are evaluated on 1500 simulated datasets of 200 subjects divided into 15 scenarios, including: 10 scenarios with a different number of covariates to evaluate and effect size of 0.4 or 0.8, 2 scenarios with correlated covariates, 2 high shrinkage scenarios, and a scenario without true covariates to specifically assess specificity. The 100 simulated dataset per scenario are obtained using the Stochastic Simulation and Estimation (SSE) tool of Perl-speaks-NONMEM (PsN). PsN is also used for SCM building process while all ML methods are R package applications. Results: Four methods have maximum sensitivity: RF, Lasso, Ridge, and their combination. Overall, VSURF and ANN are the least effective. The methods least sensitive to covariate correlation are VSURF, Lasso, and the combination of RF, Lasso, and Ridge, while RF and Boruta’s performance decreased the most in these scenarios. All methods are sensitive to shrinkage, with Lasso and SCM having the highest sensitivity in this scenario. Therefore, in case of high shrinkage, Lasso is the most suitable, as it achieves results like those of SCM. ML methods tend to select false positives, whereas SCM and SCM+ tend to select false negatives in scenarios involving correlated covariates. Thereby, combination of RF, Lasso, and Ridge used as a preselection method before SCM evaluation, highly reduces the number of models evaluated compared to classical PMx approaches (by 80% compared to SCM and by 64% compared to SCM+). Conclusions: This combined ML-SCM method would make covariate selection faster and more efficient without the risk of overlooking important covariates. Nevertheless, an extensive verification on other PK and on PK-PD models, on real data, is required before daily application of these ML methods in PMx.
[1] McComb M., Bies R., Ramanathan M. Machine learning in pharmacometrics: Opportunities and challenges. Br. J. Clin. Pharmacol. 2021;88:1482–1499. doi: 10.1111/bcp.14801. [2] Terranova, N., Venkatakrishnan, K. & Benincosa, L.J. Application of Machine Learning in Translational Medicine: Current Status and Future Opportunities. AAPS J 2021;23, 74. doi.org/10.1208/s12248-021-00593-x [3] Svensson RJ, Jonsson EN. Efficient and relevant stepwise covariate model building for pharmacometrics. CPT Pharmacometrics Syst Pharmacol. 2022; 11: 1210-1222. doi:10.1002/psp4.12838
Reference: PAGE 33 (2025) Abstr 11645 [www.page-meeting.org/?abstract=11645]
Poster: Methodology – AI/Machine Learning