Olga Teplytska (*1), Diego Valderrama (*2), Luca Marie Koltermann (1), Elena Trunz (3), Alina Pollehn (3), Eduard Schmulenson (1), Achim Fritsch (1), Jan Hasenauer (4,5), Holger Froehlich (2,6), Ulrich Jaehde (1)
(1) Department of Clinical Pharmacy, Pharmaceutical Institute, University of Bonn, Germany, (2) Fraunhofer Institute for Algorithms and Scientific Computing SCAI, Sankt Augustin, Germany, (3) Institute of Computer Science II, Visual Computing, University of Bonn, Germany, (4) Hausdorff Center for Mathematics, University of Bonn, Germany, (5) Life & Medical Sciences-Institute (LIMES), University of Bonn, Germany, (6) Bonn-Aachen International Center for Information Technology (b-it), University of Bonn, German
Introduction:
In oncology, under- or overdosing often occurs due to the administration of drug doses not taking 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. Usually, population pharmacokinetic methods are used for TDM requiring empirical models. In this work, we use two real data sets for two different anticancer drugs and compare different machine learning (ML) and population pharmacokinetic (Pop-PK) methods to accurately predict drug plasma concentrations. It is hypothesized that ML algorithms can outperform Pop-PK models because they are more expressive and can learn better from the available data.
Objectives:
To develop ML algorithms that can accurately predict plasma concentrations of anticancer drugs and compare them to classical Pop-PK models.
Methods:
A dataset of 549 fluorouracil (5FU) plasma concentrations from 157 patients as example for an iv administration [1] and another dataset of 308 sunitinib concentrations from 47 patients as example for a po administration [2] were used for analysis. Because all the 5FU samples were taken at steady-state, each 5FU measurement was considered as an individual patient. This simplification has not been applied to the sunitinib data. The datasets were divided into 80% training and 20% test data and the final results were obtained using 10-fold cross validation. The tested algorithms comprised of:
- Pop-PK models [1,2] using different estimation methods (First Order Conditional Estimation with Interaction and Stochastic Approximation Expectation Maximization), analyzed with NONMEM®.
- Classical ML algorithms, such as Random Forest, Support Vector Machine, various Gradient Boosting techniques and simple Neural Networks.
- A novel extension of a hybrid model [3] between Pop-PK and Scientific ML (SciML) that includes Variational Inference (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:
In the case of 5FU, we observed a more than threefold difference in favor of SciML: the average Root Mean Squared Error (RMSE) for the Pop-PK methods was 0.32, ranged from 0.33 to 0.61 for the classical ML methods, while we obtained values from 0.07 to 0.1 for SciML. For sunitinib, the RMSE was 13.90-13.94 for the Pop-PK methods, 19.51-25.15 for the classical ML methods and 16.40-17.99 for SciML.
Conclusions and Outlook:
Our results show that SciML significantly outperforms Pop-PK models and classical ML algorithms in the case of 5FU, but that classical Pop-PK methods perform better for sunitinib. In general, it is apparent that a compartmental model structure is required for accurate drug plasma concentration prediction due to the complexity of the data.
As future work, we plan to additionally augment the training data for the classical machine learning methods to assess whether their accuracy can be further enhanced. In the next step of the project, we aim to integrate the SciML architecture for concentration prediction into a reinforcement learning framework for dose optimisation of 5FU and sunitinib.
References:
[1] Schmulenson E, Zimmermann N, Müller L, Kapsa S, Sihinevich I, Jaehde U. Influence of the skeletal muscle index on pharmacokinetics and toxicity of fluorouracil. Cancer Med. 2023 ;12:2580-2589.
[2] Diekstra MH, Fritsch A, Kanefendt F, Swen JJ, Moes D, Sörgel F, Kinzig M, Stelzer C, Schindele D, Gauler T, Hauser S, Houtsma D, Roessler M, Moritz B, Mross K, Bergmann L, Oosterwijk E, Kiemeney LA, Guchelaar HJ, Jaehde U. Population Modeling Integrating Pharmacokinetics, Pharmacodynamics, Pharmacogenetics, and Clinical Outcome in Patients With Sunitinib-Treated Cancer. CPT Pharmacometrics Syst Pharmacol. 2017 ;6:604-613.
[3] Valderrama D, Ponce-Bobadilla AV, Mensing S, Fröhlich H, Stodtmann S. Integrating machine learning with pharmacokinetic models: Benefits of scientific machine learning in adding neural networks components to existing PK models. CPT Pharmacometrics Syst Pharmacol. 2024; 13: 41-53.
Acknowledgements:
This work was partially funded by Federal Ministry of Education and Research within the projects „BNTrAinee“ (funding code 16DHBK1022).
* Olga Teplytska and Diego Valderrama share the first authorship of this work.
Reference: PAGE 32 (2024) Abstr 11119 [www.page-meeting.org/?abstract=11119]
Poster: Methodology – AI/Machine Learning