Patients and data
For this work, raw data from a total of 150 capecitabine-naïve patients were pooled from two open, prospective multi-centered observational cohort studies. Both studies aimed at evaluating the effect of pharmaceutical care on adherence of capecitabine-treated patients and were approved by the ethics committee at the Faculty of Medicine of the University of Bonn [
17,
18]. A summary of the observed data can be found in Table
1. Capecitabine was administered orally twice daily as an intermittent regimen in 3-week cycles (14 days of treatment and seven-day break). Dose modifications, treatment interruptions and discontinuations were conducted at the sole discretion of the treating oncologists.
Table 1
Summary of observed data [
17,
18]
Patients analyzed (male/female) | 150 (39/101) |
Age (years), median (range) | 62 (28–93) |
Tumor entity | |
Colorectal cancer | 71 |
Breast cancer | 67 |
Other | 12 |
Therapy-related details | |
Capecitabine monotherapy | 71 |
Capecitabine combination therapy | 79 |
Absolute daily dose (mg), median (range) | 3000 (1000 – 5000) |
Number of observed cycles per patient, mean (range) | 5.2 (1 – 6) |
Number of patients with treatment interruptions | 33 |
Duration of treatment interruptions (days), median (range) | 8 (1 – 118) |
Number of treatment discontinuations | 56 |
Number of observed transitions between adverse event grades | |
0 → 0 | 254 |
0 → 1 | 93 |
0 → 2 | 41 |
0 → 3 | 7 |
1 → 0 | 26 |
1 → 1 | 125 |
1 → 2 | 44 |
1 → 3 | 9 |
2 → 0 | 8 |
2 → 1 | 34 |
2 → 2 | 69 |
2 → 3 | 12 |
3 → 0 | 2 |
3 → 1 | 6 |
3 → 2 | 9 |
3 → 3 | 22 |
Occurrence and severity of HFS were assessed by the patients using a questionnaire developed at the Department of Clinical Pharmacy at the University of Bonn. The description of HFS severity grades (0 to 3) was based on the descriptions provided by the CTCAE grades, version 3.0 [
1]. Grade 0 was described as the absence of symptoms, patients with grade 1 had minimal skin alterations (e.g. redness) without any pain. Grade 2 was described as skin reactions (e.g. fissures, blisters, swelling) and/or pain without impairment of activities of daily living and patients with HFS grade 3 had severe skin reactions (e.g. peeling, blisters, bleeding) and/or severe pain, including impaired activities of daily living. Patients were asked to complete the questionnaire after each conducted cycle. Therefore, up to six HFS grade assessments per patient were collected. Before starting capecitabine treatment, patients were considered asymptomatic.
Data analysis
This population pharmacodynamic analysis was performed using non-linear mixed effect modeling. Model parameters were estimated by the Laplacian method implemented in the software NONMEM 7.4.3 [
19]. The likelihood-ratio test was used to discriminate between nested models. The inclusion of an extra parameter or covariate required a statistically significant reduction (
p ≤ 0.01) of the objective function value (OFV) provided by NONMEM. Furthermore, visual predictive checks (VPC) assisted in model selection.
Implemented scripts in PsN (version 4.8.1) [
20,
21] were also used for model development and R (version 3.5.1) [
22] was used for visualization of results as well as generating random numbers for simulation analyses. Piraña (version 2.9.7) [
23] served as a front interface.
Model building
Since HFS can only be graded on a categorical scale, the probability of each grade was modeled with a proportional odds model which was extended with Markov elements. In this work, a minimal continuous-time Markov model (mCTMM) was applied to analyze the severity of HFS. The mCTMM was developed by Schindler and Karlsson and is a simplification of standard continuous-time Markov models [
24]. A compartmental structure with four compartments was used, with each compartment representing one HFS severity grade (0, 1, 2, and 3) [
7]. The probability of each grade was modeled as an amount in the respective compartment and described by differential equations in which solely transitions between adjacent states were considered (Eq.
1):
$$\begin{aligned} & \frac{{{\text{d}}P\left( 0 \right)}}{{{\text{d}}t}} = K_{10} \cdot P\left( 1 \right) - K_{01} \cdot P\left( 0 \right) \\ & \frac{{{\text{d}}P\left( 1 \right)}}{{{\text{d}}t}} = K_{01} \cdot P\left( 0 \right) + K_{21} \cdot P\left( 2 \right) - K_{10} \cdot P\left( 1 \right) - K_{12} \cdot P\left( 1 \right) \\ & \frac{{{\text{d}}P\left( 2 \right)}}{{{\text{d}}t}} = K_{12} \cdot P\left( 1 \right) + K_{32} \cdot P\left( 3 \right) - K_{21} \cdot P\left( 2 \right) - K_{23} \cdot P\left( 2 \right) \\ & \frac{{{\text{d}}P\left( 3 \right)}}{{{\text{d}}t}} = K_{23} \cdot P\left( 2 \right) - K_{32} \cdot P\left( 3 \right) \\ \end{aligned}$$
(1)
d
P(grade)/d
t represents the rate of change over time of the probability of experiencing grades 0, 1, 2 or 3,
P(grade) is the probability of experiencing one of the HFS grades,
Kgrade,grade+1 and
Kgrade,grade−1 are transition rate constants for worsening to higher grades and for recovering to lower grades, respectively.
When an observation event occurred, the amount in the compartment corresponding to the respective severity grade was set to 1 whereas the other compartments were set to 0 before the next observation. This introduced the Markov property. Between two observations, rate constants defined the transitions of probabilities between different grades. In an mCTMM, it is assumed that the transition rate between two consecutive grades is independent of the grade resulting in fewer model parameters than in other Markov models. Only the mean equilibration time (MET) was introduced as a constant parameter characterizing the transition rates across different grades. The transition rate constants govern the rate at which the probability of the adverse event severity distributes between two observations. They were defined as functions of the MET and the probabilities of the respective severity grades [
24].
The calculation of the probabilities experiencing one of the HFS grades was similar to a proportional odds model [
25]. Since four different HFS grades were considered, three probabilities had to be estimated. The fourth probability was defined as 1 minus the sum of the three others. Logit transformation was conducted to express the respective probability as a value within the interval between 0 and 1 (Eq.
2):
$${\text{logit}}\left( {P\left( {{\text{Gr}}_{ij} \ge n} \right)} \right) = \log \left( {\frac{{\left( {P\left( {{\text{Gr}}_{ij} \ge n} \right)} \right)}}{{1 - \left( {P\left( {{\text{Gr}}_{ij} \ge n} \right)} \right)}}} \right) = \alpha_{n} + g\left( {x_{i} } \right) + \eta_{i}$$
(2)
Gr
ij is the HFS grade for the
ith individual at the
jth occasion.
P(Gr
ij ≥ n) represents the probability that the HFS grade is greater than or equal to grade
n. This can be also defined as the cumulative probability of grade
n.
αn is the intercept on the logit scale and
g(
xi) represents a linear function on the logit scale which contains explanatory factors, such as drug exposure or covariates, such as age or sex. These factors are related to the probability experiencing HFS.
ηi represents the interindividual random effect for the
ith individual assuming a normal distribution with a mean of 0 and a variance of
ω2. To ensure that the cumulative probability of the respective next higher grade is lower, the following parametrization of the logit intercept was used (Eq.
3):
$$\alpha_{n + 1} = \alpha_{n} + b_{n + 1}$$
(3)
The parameter bn+1 is negatively constrained and has to be estimated in the model.
Using the inverse logit function (also called expit function),
\(P\left( {{\text{Gr}}_{ij} \ge n} \right)\) can be directly calculated as follows (Eq.
4):
$$P\left( {{\text{Gr}}_{ij} \ge n} \right) = \frac{1}{{1 + {\text{e}}^{{-\left( {\alpha_{n} + g\left( {x_{i} } \right) + \eta_{i} } \right)}} }}$$
(4)
Additionally, an interindividual variability (IIV) as an exponential function of the MET was included.
After building the base model, the effects of dose and time on the MET and the logit intercepts were tested. Here, dose was tested as a time-varying covariate between therapy cycles. Moreover, a covariate analysis was performed. Continuous (patient’s age) as well as categorical covariates (sex, tumor entity and concomitant chemotherapy) were included based on their statistical significance of reducing the OFV, i.e. improving the model fit. For one additional parameter in the model the OFV had to decrease by at least 6.64 which corresponds to a
p value ≤ 0.01 in the case of one degree of freedom. Additionally, adherence was tested as a covariate. It was measured using an electronic medication event monitoring system (MEMS™) [
17,
18] and assessed as pooled overall adherence per patient over the course of therapy. Patients were allocated to one of three groups (Overall adherence > 100%, 90–100% or < 90%).
Model evaluation
To assess the model fit, visual predictive checks for categorical data were used. 95% confidence intervals (CI) were generated from 1000 dataset simulations based on the observed dataset and superimposed by the observed proportions of patients experiencing the individual HFS grades over time.
In addition, model robustness as well as precision and bias of parameter estimates were evaluated by a non-parametric bootstrap analysis without stratification. Median and 95% CI of parameter estimates were derived from 1000 replicate datasets obtained from sampling individuals from the original dataset with replacement.
Simulation study
The developed model was used to perform a simulation study based on 1000 virtual patients to assess the appropriateness of the standard dosing regimen for capecitabine monotherapy of 1250 mg/m
2 twice daily and the proposed dose adjustments based on HFS severity according to the summary of product characteristics (SmPC) [
15]. Since no information of body surface areas (BSA) of the patients from the observational studies [
17,
18] was provided, random BSA values were generated using the rnorm function in R. BSA means and standard deviations were obtained from published data [
26]. Two simulation approaches were performed: (1) A simulation was performed in 1000 virtual patients with the above-mentioned starting dose of 1250 mg/m
2 for six cycles without dose adjustments. (2) A step-wise simulation was performed in the same 1000 patients with the same dose and total simulation duration as in (1). When meeting the criteria for dose adjustment according to the SmPC [
15], the capecitabine dose was adjusted after each conducted cycle. To have an equal number of patients in both simulation scenarios, patients for whom a treatment discontinuation would be recommended were kept in the analysis. After adjusting the dose, the simulation of the subsequent cycle was performed. The HFS grade corresponding to the highest simulated probability was used to assess toxicity.
The ability of the model to predict individual HFS severity was assessed by a simulation of patients with the same characteristics as in the original dataset. Therefore, the included random effect parameters were estimated by a Bayesian approach up to a certain cycle. Then, the HFS severity of the subsequent cycle was simulated based on the Bayesian estimates and covariate effects. This approach was conducted for predictions of cycle 2 up to cycle 6. Since Markov models can only predict the probability for each toxicity grade but not the grade itself, the grade corresponding to the highest probability was compared to the respective observed HFS grade. All grades were allocated to one of the following two groups: The first group consisted of HFS grades ≥ 2 which were classified as clinically relevant since dose reductions or treatment interruptions are conducted at grade 2 or higher [
15], the second group consisted of HFS grades 0 and 1. For the first group, a positive predictive value (PPV) was calculated. It indicated the ability of predicting clinically relevant HFS:
$${\text{PPV}} = \frac{{N\,{\text{true}}\,{\text{predicted}}\,{\text{events}}\,{\text{with}}\,{\text{grade}}\, \ge 2}}{{N\,{\text{total}}\,{\text{predicted}}\,{\text{events}}\,{\text{with}}\,{\text{grade}}\, \ge 2}}$$
(5)
The ability of predicting the absence of toxicity ≥ grade 2 was assessed by calculation of a negative predictive value (NPV) within the second group:
$${\text{NPV}} = \frac{{N\,{\text{true}}\,{\text{predicted}}\,{\text{events}}\,{\text{with}}\,{\text{grade}}\, \le 1}}{{N\,{\text{total}}\,{\text{predicted}}\,{\text{events}}\,{\text{with}}\,{\text{grade}}\, \le 1}}$$
(6)
Since patients were considered asymptomatic before starting therapy, predicted HFS grades at baseline were not included for calculation of both NPV and PPV.