Clinical applications of machine learning methods on pharmaceutical development



Journal Title

Journal ISSN

Volume Title



The investigation of the effect of new drugs on patients involves mathematical modeling of complex nonlinear dynamical systems. Traditionally, simplistic pharmacokinetic (PK)/pharmacodynamic (PD) modeling has been employed for such investigations based on assumptions that may not represent the dynamical changes in the body accurately. Recent research, therefore, employs a data-driven approach to the problem by modeling an input-output nonlinear dynamical system (IO-NLDS) to predict the vital signs based on specific drug infusion rates. In specific cases considering only a few monitored vital signs, improved performance has been reported. The goal of our research is to employ such a data-driven approach to investigate the effects of new drug concentration for better prediction of clinical outcomes. Due to the enormous cost and long time required for actual clinical trials, current research is focused on extending the IO-NLDS model to include the variations expected in real-time noisy data by using a reliable statistical framework for simulation, prediction, and control of the variability observed in real data. Our research focused on including various neuronal networks in PK and PD modeling. Feedforward and Recurrent neural networks provide comparable results in PK and PD modeling with limited data from current clinical trials. The Recurrent neural network (RNN) is an extremely powerful sequence model that captures the dynamics of sequences via cycles in the network of nodes. Unlike standard feedforward neural networks, recurrent networks retain a state that can represent information from previous computations. The RNN can contribute to developing a model reflecting the relationship between the drug response and concentration of Drug X. Drug X is an orally administered drug designed to increase the blood concentrations of an endogenous molecule R. The mechanism of action of Drug X is modeled by Nonlinear Mixed Effects Modeling (NONMEM), and the various simulated intensive datasets are conducted for machine learning experiments. Reinforcement Learning (RL) is presented as a successful practice to facilitate the optimal dose selection while providing promising accuracy in matching the RNN model and NONMEM model.

Embargo status: Restricted to TTU only. TTU community may view by logging in with their eRaider (top right). Others may request access by clicking on the PDF link to the left.



Machine learning, Pharmacology