Non-linear autoregressive exogenous modelling to predict individual haematotoxicity during chemotherapy
Marie Zech (1,2,3), Yuri Kheifetz (2), Markus Scholz (2,3)
(1) Center for Scalable Data Analytics and Artificial Intelligence (ScaDS.AI) Dresden/Leipzig, Germany, (2) Institute for Medical Informatics, Statistics and Epidemiology (IMISE), Leipzig University, Germany (3) Faculty of Mathematics and Computer Science, Leipzig University, Germany
Objectives: Cytotoxic cancer therapies frequently result in severe haematotoxic side-effects. Predicting a patient’s haematologic response to treatment is of high clinical relevance but is difficult due to high between-patient heterogeneity. A number of mechanistic and semi-mechanistic models of bone marrow haematopoiesis have been developed so far to solve this task [1-4]. However, the established models were not able to describe patients exhibiting irregular dynamics [1,2]. Here, we propose a data-driven hypothesis-free machine learning approach to model individual patient’s time courses and determine whether it can be an alternative to semi-mechanistic modelling.
Methods: We apply recurrent neural networks based on non-linear autoregressive exogenous (NARX) models to describe the highly non-linear dynamics of haematologic lineages under chemotherapy. To cope with the relative sparsity of individual patient data, we implement a transfer learning process. We learn an individuals dynamics with a network pre-trained on complex data, controlling the transfer process with the Bayesian information criterion. Further, we employ several model optimization and reduction methods to derive robust and parsimonious individual networks. We apply and test this framework based on a virtual patient population under different treatment scenarios. Virtual patient data are generated by simulating a semi-mechanistic model of Friberg et al. [3] for a hypercube of plausible model parameters and for different chemotherapy schedules. For each virtual patient, we train a personalized prediction model using NARX neural networks and analyse its prediction performance based on therapy scenarios not used for model training. Additionally, we apply this transfer learning approach on a second set of virtual patients exhibiting dynamics of cumulative neutropenia, which we simulate with the semi-mechanistic model of Henrich et al. [4]. This transfer closely resembles possible real- world applications, where a model learned on simulated data could be fine-tuned to more complex real individual data.
Results: We observe good generalization performances for both virtual patient populations. Our approach to reduce model complexity is successful, resulting in parsimonious individual networks. While the network performance depend on the complexity of the treatment scenario, there is no strong correlation between performance and network complexity. This suggests that the overfitting issue of learning the individual networks is successfully addressed.
Conclusions: We demonstrate that NARX modelling can provide robust predictions of an individual patient’s response to treatment, and therefore, can serve as an alternative to traditional population approaches. The transfer process to more complex dynamics with cumulative toxicity shows that the approach is able to capture the dynamics of a family of haematopoiesis models with largely different parameter settings and therapy schedules, while maintaining parsimonious individual network architectures.
References:
[1] Kheifetz Y, Scholz M (2019) Modeling individual time courses of thrombopoiesis during multi-cyclic chemotherapy. PLOS Computational Biology 15(3): e1006775. https://doi.org/10.1371/journal.pcbi.1006775
[2] Kheifetz, Y, Scholz, M. Individual prediction of thrombocytopenia at next chemotherapy cycle: Evaluation of dynamic model performances. Br J Clin Pharmacol. 2021; 87: 3127– 3138. https://doi.org/10.1111/bcp.14722
[3] Friberg, L. E., Henningsson, A., Maas, H., Nguyen, L., & Karlsson, M. O. (2002). Model of chemotherapy-induced myelosuppression with parameter consistency across drugs. Journal of clinical oncology: official journal of the American Society of Clinical Oncology, 20(24), 4713–4721. https://doi.org/10.1200/JCO.2002.02.140
[4] Henrich, A., Joerger, M., Kraff, S., Jaehde, U., Huisinga, W., Kloft, C., & Parra-Guillen, Z. P. (2017). Semimechanistic Bone Marrow Exhaustion Pharmacokinetic/Pharmacodynamic Model for Chemotherapy-Induced Cumulative Neutropenia. The Journal of pharmacology and experimental therapeutics, 362(2), 347–358. https://doi.org/10.1124/jpet.117.240309