Generation and Application of Avatars (Digital Twins) in Pharmacometric Modelling
Estelle Chasseloup (1), Andrew Hooker (1) and Mats O. Karlsson(1)
(1) Dept of Pharmaceutical Biosciences, Uppsala University, Uppsala, Sweden.
Objectives: The concept of Avatars or Digital Twins is well established in health, engineering, and systems biology. It is the creation of a digital representation of a physical or a biological system in order to explore and control its behaviour [1,2]. In this work we aim to explore aspects of generating and utilizing avatars for pharmacometric population models, accounting for clinical relevance. We used a neutropenia model as example.
Methods: A previously published pharmacodynamic model describing the neutrophil count after a single administration of docetaxel and its associated database (3553 observations for 601 IDs) were used [3]. In oncology the main clinical endpoints to monitor the haematological toxicities after a chemotherapy treatment are the count of the white blood cells before treatment (at baseline, B), at maximal depletion (nadir, N), and at the expected time of the next dose (return, R). The avatars were selected among model simulations according to the closeness of the simulated individuals to these clinically relevant criteria. All the combinations of these three clinical endpoints were tested (e.g. N, BN, and BNR). The closeness was defined at the individual level as an error margin (10, 20, 30 or 50%) around the observed data values for the three clinically relevant time points. To ensure at least one avatar for each ID for the different combinations tested we varied the number of simulations (1000 or 10000). NONMEM v7.3.0 and R v3.3.0 were used for the simulations and the selection process respectively.
Results and discussion: In this work we defined avatars for population models as a subset of simulated individuals based on their closeness w.r.t the observations at clinically relevant criteria. White blood cell count vs. time profile plots show that the avatars profiles are closer to the individual observations than the whole of the simulations. Bivariate plots of the random inter-individual parameters show significant differences between the avatars, the EBEs and the estimated theoretical distributions. We also show that 10000 simulations were not enough to guarantee an avatar for each ID for stringent (10%) or less stringent (50%) error margins. The avatar properties could be interesting in different areas of pharmacometrics, e.g. refined clinical trial simulations, quantifying individual predictive performance for model based dose adjustment in clinical practice, measurement of subject uniqueness, as a model diagnostic, and coherence of the simulations with the model and not only one but multiple observed variables.
Conclusions: An efficient method to select clinically relevant avatars for population models was implemented. Avatars can give nuanced information regarding the ability of a model to simulate data similar to the observed data. A dynamic avatar selection instead of a static one (i.e. after all the simulations were done) used in this work would guarantee an avatar for each subject, even for stringent criteria, using less simulation resources. Avatars properties could then be investigated in different areas of pharmacometrics.
References:
[1] Parrott, A. & Lane, W. Industry 4.0 and the digital twin. Deloitte University Press (2017). 11‘Grand.’ Webster’s Third New International Dictionary, Unabridged. 2018. Web. 18 Feb. 2018.
[2] Spanakis, M.,
Spanakis, E., G., Kafetzopoulos, D., Sakkalis, V., Tsiknakis, M., Marias, K., Dong, F. (2015). MyHealth Avatar with the generated virtual profiles from clinical trials. Poster presented at the PAGE (Population Approach Group of Europe) conference of 2015.
[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, 20(24), 4713-4721.
