Marie Steinacker (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 treatment in cancer therapy is frequently accompanied by dose-limiting haematotoxic side-effects. Predicting an individual’s risk holds significant clinical relevance, but proves to be challenging due to high between-patient heterogeneity. To solve this task, several (semi-) mechanistic models of bone marrow haematopoiesis have been developed [1-3], but a subset of patients exhibiting irregular dynamics could not be described [1,2]. In this regard, we proposed a data-driven hypothesis-free machine learning approach to model individual time courses [4]. Here, we provide a comparative analysis of our approach on the basis of real-world patient data collected in the framework of the NHL-B study [5,6]. We examine different machine learning approaches for time series to derive personalized predictions and explore the potential of combining biological knowledge and data-driven approaches.
Methods: We apply non-linear autoregressive networks with exogenous inputs (NARX) to describe the highly non-linear dynamics of haematologic lineages under chemotherapy. We study both feed-forward networks and gated recurrent units (GRU) as internal architectures for the NARX network. We also compare prediction performances of purely data-driven machine learning approaches, biologically motivated semi-mechanistic models [3] and combinations of it based on transfer learning techniques. To avoid over-fitting, we employ a combination of model optimization and reduction methods, as well as augmentation of training data. For individual patient data, we train personalized prediction models and analyse their prediction performances based on later treatment data not used for model training.
Results: Among the examined machine learning models, the NARX network based on a GRU architecture performed best. For scenarios where the semi-mechanistic model fails to make good predictions, the improvement by the network can be substantial, if there is enough data provided for the first treatment cycles of the patient. For scenarios, where the semi-mechanistic model is in good agreement with the patient’s data, the network performed comparable. For patients with a lower number of measurements, our transfer approach showed an improvement in comparison to pure machine learning or semi-mechanistic model-based approaches. Contrary, if the purely data-driven approach performs well and the semi-mechanistic model does not, the transfer learning approach could be detrimental for prediction performances. Our employed optimization methods proved to be successful in avoiding over-fitting.
Conclusions: We demonstrate that NARX modelling can provide robust predictions at an individual scale. However, prediction performances strongly depend on the amount of training data available per patient, in particular during the first therapy cycle. We recommend at least three well-spaced measurements which ideally cover the decline and recover phase of blood count dynamics, as well as the nadir. We also demonstrated that transfer learning could ameliorate the lack of data to some extent.
References:
[1] Kheifetz Y, Scholz M. Modeling individual time courses of thrombopoiesis during multi-cyclic chemotherapy [published correction appears in PLoS Comput Biol. 2019 Jun 4;15(6):e1007110]. PLoS Comput Biol. 2019;15(3):e1006775. Published 2019 Mar 6. doi: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(8):3127-3138. doi:10.1111/bcp.14722
[3] Friberg LE, Henningsson A, Maas H, Nguyen L, Karlsson MO. Model of chemotherapy-induced myelosuppression with parameter consistency across drugs. J Clin Oncol. 2002;20(24):4713-4721. doi:10.1200/JCO.2002.02.140
[4] Steinacker M, Kheifetz Y, Scholz M. Individual modelling of haematotoxicity with NARX neural networks: A knowledge transfer approach. Heliyon. 2023;9(7):e17890. Published 2023 Jul 5. doi:10.1016/j.heliyon.2023.e17890
[5] Pfreundschuh M, Trümper L, Kloess M, et al. Two-weekly or 3-weekly CHOP chemotherapy with or without etoposide for the treatment of young patients with good-prognosis (normal LDH) aggressive lymphomas: results of the NHL-B1 trial of the DSHNHL. Blood. 2004;104(3):626-633. doi:10.1182/blood-2003-06-2094
[6] Pfreundschuh M, Trümper L, Kloess M, et al. Two-weekly or 3-weekly CHOP chemotherapy with or without etoposide for the treatment of elderly patients with aggressive lymphomas: results of the NHL-B2 trial of the DSHNHL. Blood. 2004;104(3):634-641. doi:10.1182/blood-2003-06-2095
Reference: PAGE 32 (2024) Abstr 10772 [www.page-meeting.org/?abstract=10772]
Poster: Methodology - New Modelling Approaches