Model-based dose adaptation of capecitabine for prevention of severe hand-and-foot syndrome: in silico comparison with the standard method

Inès Paule (1), Michel Tod (1), Emilie Hénin (1), Benoit You (1), Brigitte Tranchand (1), Gilles Freyer (1,2), Pascal Girard (1)

(1) EA3738 CTO, Faculté de Médecine Lyon-Sud, Université Claude Bernard - Lyon 1, France;(2) Service d’Oncologie Médicale, Hospices civils de Lyon, France

Introduction:
Management of anticancer therapies is complex due to their narrow therapeutic indices (interval between minimum effective and toxic doses) and high inter-patient variability. With identical dosage, some patients may show no therapeutic response while others may experience severe side effects. One objective is to maximize anticancer effect for each patient without running unacceptable risks of severe toxicities. Individual dosage optimization is one of the tools to achieve this therapeutic goal.
Capecitabine (Xeloda®, Roche) is an orally taken prodrug of 5-fluorouracil (5FU), which is one of the most extensively used chemotherapeutic agents against solid tumours. As it is preferentially metabolized to the active molecule 5FU in the tumour tissue, capecitabine has the advantage of being less toxic to the healthy tissues [1], while having superior or at least non-inferior efficacy as compared to the intravenously administered 5FU associated to leucovorin (5FU/LV) [2, 4]. Although less toxic than 5FU, capecitabine shows toxicities as diarrhea and palmar plantar erythrodysesthesia, the so called hand-and-foot syndrome (HFS), which is much more frequently experienced by patients treated with capecitabine (54%) than those treated with 5FU/LV (6%) [3]. This syndrome manifests as a numbness or even desquamation of palm and sole skin, possibly disturbing daily activities. It is measured on an ordinal scale of severity from grade 0 (none) to grade 3.
While the common practice is to reduce doses by 25% after the 2nd occurence of severe toxicity (grade>=2), then by 50% after the 3rd episode, without taking into account other information about the patient, a more rational individualized dose adaptation approach might allow better control of toxicity and thus improve the therapeutic benefit. The idea is to determine the most appropriate dose for the next treatment cycle according to the prediction of individual toxicity risk, evaluated dynamically and by taking into account the particular patient’s characteristics.

Objectives:
To set up the methodology for individual dose adaptation on the basis of ordinal observations and evaluate its feasibility and performances, as compared to the standard approach, by randomized in silico clinical trials.

Methods:
HFS individual model
Individual prediction-based dose adjustment schemes for capecitabine were derived on the basis of individual HFS observations and a population longitudinal HFS toxicity model previously developed in [3], using a dataset of 595 metastatic or advanced colorectal cancer patients from two phase III studies[2, 4]. The mixed effects transitional and proportional odds model for longitudinal ordinal data links taken doses, basal creatinine clearance and previous toxicity to the risk of (the highest) HFS grade of the week. Due to absence of pharmacokinetic data, the drug effect is quantified by a kinetic-pharmacodynamic (KPD) model [5], whose particular feature is to relate the pharmacodynamic outcome (here, toxicity grade) to the taken drug doses without specifying the true PK model. The main idea of this model is to assume drug accumulation in a virtual effect (KPD) compartment and a mono-exponential pseudo-elimination.
To use this model for individual dose adaptation, we need to (i) estimate random individual effects of the patient, conditional on the observations of his past HFS toxicity (the estimation step), (ii) then to choose the new dose producing an acceptable risk of severe toxicity for the next cycle (the dose calculation step).
Estimation of random individual effects (ETAs)
1. Individual likelihood. Individual parameters were estimated by maximum a posteriori (MAP) estimator. For this, a specific likelihood function for the ordinal observations was derived.
2. Optimization. Local (simplex, quasi-Newton) and global (Adaptive Random Search) optimization methods, as well as Bayesian estimation by MCMC were tested.
Dose determination rule: the most suitable dose for the next cycle was considered to be the one that correspondes to a certain tolerable limit of severe toxicity risk (in 2 weeks).
In silico clinical trial protocol: One treatment cycle corresponded to 2500 mg/m2/day for 2 weeks, followed by 1 week rest. Trial duration was 30 weeks (10 cycles). Patients were assumed to drop out of the trial if they could not receive any dose for more than 6 consecutive weeks or if severe toxicity occurred for the 4th time.
The “Standard” protocol corresponded to the approved dose adaptation rules of capecitabine which are currently used. The alternative “Individual” protocol adjusted the dose according to the individual model estimations, without increasing the dose over the nominal dose. The “Individual+” protocol was as the “Individual” one, except that it allowed an increase of the dose up to 50% in patients not showing any toxicity at all. Those 3 protocols are described in Table I.

Proof-of-concept simulation and exploration of statistical power
For the proof of concept, three arms were simulated with 10.000 patients per arm. In order to explore the statistical power, 100 replications of trials with 300, 400 and 600 patients per arm were simulated. Wilcoxon test was used to estimate the significance of reduction of toxicity duration. The simulations were run in Trial Simulator 2 (Pharsight®) [6], and required the writing of specific Fortran subroutines for individual parameter estimation and dose calculation that were linked to the simulator. Bayesian estimation was performed within WinBUGS software [7].

Results:
Optimization algorithm comparison. Simplex optimization method showed to be as accurate as Adaptative Random Search (ARS) and much faster. Quasi-Newton algorithm was slightly less accurate and slower than simplex. However, the precision of the estimates was not excellent. Large confidence intervals of estimates given by Bayesian estimation indicated identifiability issues for this particular model. However, tests performed with modified values of (dose-related) model parameters showed that it is possible to obtain rather accurate individual parameter estimates of an ordinal data model when the response (ie toxicity grade) is highly reactive to the input (doses)).
The proof-of-concept simulation showed that individual model-based adaptation would result in reduction of severe HFS toxicity incidence by 13%, of its average duration by 1.6 weeks (12 days); as well, an average reduction of HFS grade 1 duration by 3.3 weeks (23 days), as compared to the standard approach would be expected. Moreover, continuous control of severe toxicity risk resulted in earlier detection of patients intolerant to capecitabine, and therefore the mean treatment duration was 6 weeks shorter than with standard adaptation.
“Individual+” protocol simulation results suggested that patients without any toxicity could benefit of dose increase up to +50% without increase in severe HFS toxicity (29% population concerned).
A clinical trial comparing “Standard” and “Individual” dose adaptations should include 600 patients per arm to achieve at least a 90% statistical power for a significant (alpha=0.05) reduction of severe HFS duration.

Conclusions:
Individualized dose adaptation on the basis of ordinal observations, using the developed methodology, showed to be feasible and beneficial. In silico results indicate that in the case of hand-and-foot syndrome induced by capecitabine, severe toxicity incidence may be reduced by 13% and its mean duration by 12 days. Moreover, estimation of individual toxicity risk showed to be especially beneficial for allowing early detection of patients intolerant to capecitabine (at high risk of severe toxicity) and therefore better determination of the optimal moment to switch to another treatment.
There are several limitations to this work. Firstly, judgement of adaptation strategies is limited because impact on anti-cancer efficacy and other toxicities could not be evaluated. It should be considered that individual adaptation leads to 18% reduction of drug exposure as compared to the standard adaptation. The second restriction of this dose adaptation is related to the model which seems to assume inertia of HFS toxicity. This may be due to cumulative nature of the drug or model producing some bias for toxicity recovery. Nevertheless, this work shows that individual dose adaptation of oral anticancer drugs, performed on the basis of ordered categorical data, should be beneficial and feasible in clinical routine.

References:
[1] Blesch, K. S. et al. Clinical pharmacokinetic/pharmacodynamic and physiologically based pharmacokinetic modeling in new drug development: the capecitabine experience. Invest New Drugs 21, 195-223 (2003).
[2] Hoff, P. M. et al. Comparison of oral capecitabine versus intravenous fluorouracil plus leucovorin as first-line treatment in 605 patients with metastatic colorectal cancer: results of a randomized phase III study. J. Clin. Oncol. 19, 2282-2292 (2001).
[3] Hénin, E. et al. PAGE 2006; abstract No 929.
[4] VanCutsem, E. et al. Oral capecitabine compared with intravenous fluorouracil plus leucovorin in patients with metastatic colorectal cancer: results of a large phase III study. J. Clin. Oncol. 19, 4097-4106 (2001).
[5] Jacqmin, P. et al. Modelling response time profiles in the absence of drug concentrations: definition and performance evaluation of the K-PD model. J Pharmacokinet Pharmacodyn 34, 57-85 (2007).
[6] Pharsight Trial Simulator v2.1.2 User's Guide. Pharsight Corporation, Mountain View, CA . 2001.
[7] Lunn, D.J., Thomas, A., Best, N., and Spiegelhalter, D. (2000) WinBUGS -- a Bayesian modelling framework: concepts, structure, and extensibility. Statistics and Computing, 10:325--337.

Reference: PAGE 17 (2008) Abstr 1260 [www.page-meeting.org/?abstract=1260]

Oral Presentation: Applications