Scientific machine learning for predicting plasma concentrations in anti-cancer therapy

Abstract

A variety of classical machine learning approaches have been developed over the past ten years with the aim to individualize drug dosages based on measured plasma concentrations. However, the interpretability of these models is challenging as they do not incorporate information on pharmacokinetic (PK) drug disposition. In this work we compare well-known population PK modelling with classical and a newly proposed scientific machine learning (SciML) framework, which combines knowledge on drug disposition with data-driven modelling. Our approach lets us estimate population PK parameters and their inter-individual variability (IIV) using multimodal covariate data of each patient. A dataset of 549 fluorouracil (5FU) plasma concentrations as example for an intravenously administered drug and a dataset of 308 sunitinib concentrations as example for an orally administered drug were used for analysis. Whereas classical machine learning models were not able to describe the data sufficiently, the proposed model allowed us to obtain highly accurate predictions even for new patients. Additionally, we demonstrated that our model could outperform traditional population PK models in terms of accuracy and greater flexibility when learning population parameters if given enough training data.