Скачать или смотреть Amortized approach to non-linear mixed-effects modelling with neural density posterior estimation

Talk by Jonas Arruda at BioInference 2023.

https://bioinference.github.io/

Abstract: Non-linear mixed-effects models (NLMEs) are a powerful tool for analyzing heterogeneous populations, and are commonly used in a wide range of fields such as pharmacokinetics, single-cell analysis, econometrics, and engineering. However, the non-linearity in these models, combined with the presence of random effects, can make them computationally demanding and difficult to fit to large data sets and even infeasible for models including more complex stochasticity. This is in particular the case for uncertainty analysis.

Additionally, established methods often lead to non-reproducible results, making it challenging to find accurate parameter estimates. For the first time in the context of NLMEs, we use simulation-based neural posterior estimation, a novel machine learning-based approach, which has been developed to address general parameter estimation problems. This approach is based on invertible neural networks and has been shown to effectively handle complex models and large data sets.
This method is applied to first train entity-specific posterior distributions and then use these posteriors for estimating the population parameters of the NLMEs, thus improving the scalability and robustness of parameter estimation.

By leveraging these posteriors, researchers can perform more comprehensive uncertainty analysis and population model calibration, leading to more accurate and reliable results.