2021/03/02 by Édouard Ollier, Ollier, Edouard
Biochemistry, Genetics and Molecular Biology · Chemistry · Mathematics · #62-08 #Computation (stat.CO) #FOS: Computer and information sciences #G.3 #Gene expression and cancer classification #Methodology (stat.ME) #Spectroscopy and Chemometric Analyses #Statistical Methods and Inference
paper · pdf · doi:10.48550/arxiv.2103.01621
openalex publication_date 2021/03/02 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Nonlinear Mixed effects models are hidden variables models that are widely\nused in many fields such as pharmacometrics. In such models, the distribution\ncharacteristics of hidden variables can be specified by including several\nparameters such as covariates or correlations which must be selected. Recent\ndevelopment of pharmacogenomics has brought averaged/high dimensional problems\nto the field of nonlinear mixed effects modeling for which standard covariates\nselection techniques like stepwise methods are not well suited. The selection\nof covariates and correlation parameters using a penalized likelihood approach\nis proposed. The penalized likelihood problem is solved using a stochastic\nproximal gradient algorithm to avoid inner-outer iterations. Speed of\nconvergence of the proximal gradient algorithm is improved using component-wise\nadaptive gradient step sizes. The practical implementation and tuning of the\nproximal gradient algorithm are explored using simulations. Calibration of\nregularization parameters is performed by minimizing the Bayesian Information\nCriterion using particle swarm optimization, a zero-order optimization\nprocedure. The use of warm restart and parallelization allowed computing time\nto be reduced significantly . The performance of the proposed method compared\nto the traditional grid search strategy is explored using simulated data.\nFinally, an application to real data from two pharmacokinetics studies is\nprovided, one studying an antifibrinolytic and the other studying an\nantibiotic.\n