Handling inter-occasion variability in model-based therapeutic drug monitoring
Joćo A. Abrantes (1), Siv Jönsson (1), Mats O. Karlsson (1), Elisabet I. Nielsen (1)
(1) Department of Pharmaceutical Biosciences, Uppsala University, Uppsala, Sweden
Objectives: The potential of dose individualization using therapeutic drug monitoring (TDM) is known to be challenged by high magnitudes of inter-occasion variability (IOV) [1, 2]. This analysis aims at comparing different approaches to handle IOV in a TDM context, using a population PK model for coagulation factor VIII (FVIII) [3].
Methods: Using the original model containing inter-individual variability (IIV) and IOV (ORIG), models with varying magnitudes of IOV on PK parameters (0-50%; IOV models) and corresponding models re-estimated without IOV (IIV models) were derived. The IOV models were used to simulate PK parameters and steady-state FVIII activity for 1000 severe haemophilia A patients at 4, 24 and 48 h after administration (30 IU/kg) on 4 occasions (OCCs). Empirical Bayes parameter estimates (EBEs) were obtained based on the simulated data using the IOV and corresponding IIV model, varying the amount of data included (only OCC 1, OCC 1+2, etc.). The individual doses aiming at the FVIII activity target (0.01 IU/mL at 48 h) were calculated as: i) and ii) EBEs by IOV model and dose obtained including only IIV (IOV1 approach) or IIV+IOV (IOV2 approach) etas; iii) EBEs by IIV model (IIV approach). The FVIII activity for the subsequent OCC was predicted using the estimated dose and simulated (true) parameters (real-world TDM setting). The performance was quantified as percentiles of the predicted 48 h FVIII activity. Simulations and estimations were conducted using NONMEM 7.3.
Results: The individual predicted doses resulted in low bias (median FVIII predictions for the subsequent OCC 2 to 4 ~0.010 IU/mL). The IOV1 approach was the most precise method and similar performance was noted for the IIV approach when IOV was <20%, while the IOV2 approach showed higher imprecision. Further, all approaches showed improved precision with increasing amount of data available (except IOV>IIV). The 2.5th and 97.5th percentiles of FVIII predictions using all approaches (IIV, IOV1, IOV2) based on simulations from the ORIG model were 0.0024-0.032 IU/mL (1 OCC) and 0.0027-0.027 (3 OCCs); the same values for the CL IOV 50% model were 0.00011-0.482 IU/ml (1 OCC) and 0.000059-0.52 (3 OCCs).
Conclusions: The IIV and IOV1 approaches showed similar performance in dose individualization when IOV was low (<20%), for remaining scenarios IOV1 was superior. If employing an IOV model in Bayesian forecasting, IOV etas should not be used in the calculation of the individualized dose.
References:
[1] Wallin, JE et al. Basic Clin Pharmacol Toxicol. 2010 Mar;106(3):234-42.
[2] Karlsson, MO & Sheiner, LB. J Pharmacokinet Biopharm. 1993 Dec;21(6):735-50.
[3] Björkman, S et al. Eur J Clin Pharmacol. 2009 Oct;65(10):989-98.
