Machine Learning algorithms for concentration prediction in anticancer drug therapy: Towards novel ap-proaches for individual dose adjustment

Introduction: In oncology, under- or overdosing often occurs due to the administration of drug doses that do not take into account the high inter-individual variability of pharmacokinetic processes. One option for dose individualisation is Therapeutic Drug Monitoring (TDM), i.e. individual dosing based on measured drug plasma concentrations. In this work, we compared a novel multimodal pharmacokinetic SciML (MMPK-SciML) approach with several classical ma-chine learning (ML) and population pharmacokinetic (PopPK) models regarding their predictive performance of drug plasma concentrations. Methods: A dataset of 541 fluorouracil (5FU) plasma concentrations from 156 patients as example for an iv admin-istration [1] and another dataset of 302 sunitinib concentrations from 47 patients as example for a po administration [2] were used for analysis. The datasets were divided into 80% training and 20% test data and the final results were ob-tained using 10-fold cross validation. The tested algorithms comprised of: • PopPK models [1,2] using different estimation methods (First Order Conditional Estimation with Interaction and Stochastic Approximation Expectation Maximization with Interaction), analyzed with NONMEM®. • Classical ML algorithms, such as Random Forest, Support Vector Machine, various Gradient Boosting techniques and simple Neural Networks. For these analyses, the training data was additionally augmented with 1000 simulat-ed patients. • A novel extension of a hybrid model [3] between Pop-PK and Scientific ML (SciML) that includes Variational Infer-ence (VI). In these models, compartmental PK models were coupled with neural networks to learn patient-specific offsets from population parameters, i.e., random effects, in a fully data-driven manner. Results: Our results demonstrate that a compartmental model structure is generally required to make accurate predic-tions of drug plasma concentrations. The classical ML models, which are fully data-driven, were not able to adequately learn key aspects of the data and the performance was not improved by training data augmentation. In the case of 5FU, MMPK-SciML had the most accurate drug concentration predictions of all methods, as shown by performance metrics and goodness-of-fit plots. For sunitinib, we again observed more accurate predictions than of classical ML methods, but slightly less accurate concentration predictions compared to PopPK. This difference is probably attributa-ble to the smaller size of the dataset and the higher complexity of the task. Conclusions and Outlook: Overall, MMPK-SciML has shown promising results in our work and has the potential to be used in the development of new dosing strategies. In a scoping review conducted by our team, we showed that Reinforcement Learning methods are increasingly being developed for automated dose individualisation of anticancer drugs. However, most algorithms still lack clinical utility because they are often based on generic models , depict simu-lated cellular interactions, and do not impose appropriate constraints on the proposed regimens. Therefore, in the next step of our project, we aim to develop Reinforcement Learning algorithms that are as close to clinical practice as possi-ble and to integrate the MMPK-SciML structure for concentration simulation.