Structured and Unstructured (Hybrid) Modeling in Precision Medicine

One of the key objectives in precision medicine is to determine the right dose for the individual patient at the right time so that the desired therapeutic effect is achieved. The focus of this work is on modeling of pharmacokinetic/ pharmacodynamic data to facilitate the achievement of this goal. One novelty of our approach is to use structured models, such as physiologically-based compartment models and un-structured models, such as artificial neural networks or Gaussian Processes in a hierarchical fashion. The reason for using a hierarchical structure is that there are available well-established empirical compartmental and mechanistic physiologically based models, which do not explicitly account for various predictive covariates such as co-administered drugs or different laboratory measurements such as total protein, blood urea nitrogen, or urine output. Thus, we extend the structured models with the second hierarchical layer of an un-structured model and utilize the unstructured model to capture the effects of those covariates. Secondly, we employ Bayesian inference which allows direct quantification of uncertainty in the model predictions. Thirdly, utilization of Bayesian inference for the unstructured models (specifically Bayesian neural networks) allows the determination of important predictive covariates such as serum creatinine, blood urea nitrogen, or urine output.