PI-081 - PHARMACOMETRIC MACHINE LEARNING: INTEGRATING NEURAL NETWORKS FOR FLEXIBLE, ADVANCED COVARIATE ANALYSIS

A. Elmokadem¹, M. Weins², H. Husband¹, S. Castillo¹, K. Utsey¹, T. Knab³, D. Kirouac⁴; ¹Metrum Research Group, CT, United States, ²Metrum Research Group, CT, USA, ³Meterum Research Group, CT, USA, ⁴Metrum Research Group, Coneticut, USA.

Background: Integrating covariates in traditional pharmacometric modeling involves complex steps and assumptions about covariate-parameter relationships, which may oversimplify or miss important clinical variables impacting pharmacology and response. Deep Compartment Modeling (DCM) addresses these challenges by using artificial neural networks (ANN) to characterize these relationships in a single step, eliminating the need for covariate selection and enabling complex data inputs and non-linearities. We demonstrate DCM in pharmacometric workflows using two open-source platforms: A Julia-based workflow incorporating hierarchical random effects through Bayesian analysis, and an R workflow using Keras and TensorFlow for handling larger datasets.

Methods: Population Pharmacokinetic (PK) data with clinically relevant covariates was synthesized using a two compartment model to test the DCM frameworks. For the Julia-based workflow, we assessed whether it could identify covariate relationships and quantify interindividual versus residual variability (random effects) from a small dataset (10:20 train:test split). The R workflow was trained on a large (10,000 subject) dataset with non-linear covariates, exploring various network architectures with minimal code using Keras.

Results: The Julia framework (a hierarchical DCM) successfully identified PK model parameters, random effects, and ANN weights while quantifying uncertainty. This was evident from Bayesian model diagnostics and posterior predictive checks (PPCs) which accurately characterized the training and testing of PK datasets. The R workflow effectively captured non-linear covariate relationships, highlighted by Shapley additive explanations (SHAP), while also estimating the Residual Unexplained Variability (sigma) conditional on covariates.

Conclusion: Neural networks can be integrated with traditional pharmacometric models using several free open-source programming languages. Both Julia and R environments are suitable platforms, but there are tradeoffs regarding development speed, built-in capabilities, and documentation. DCM simplifies the covariate modeling process and uncovers complex, non-linear relationships in computationally efficient workflows.