Inclusion and exclusion criteria
Volunteers aged 20–55 years and within 20% of their ideal body weight [IBW (kg) = height (cm) - 100) × 0.9] were enrolled in the study. Those who were not suitable on the basis of physical examinations and routine laboratory tests (blood hematology, biochemistry, prothrombin time, bleeding time, urinalysis) were excluded.
Study design and ethical considerations
The analysis was performed using data from a randomized, open-label, multiple-dose, two-period, two-treatment, comparative crossover study involving 38 healthy adult males. The clinical trial was conducted at Kyungpook National University Hospital Clinical Trial Center (KNUH CTC) to determine the bioequivalence and non-inferiority of two different triflusal formulations, as reported previously [4].
The study was conducted in accordance with the guidelines of Good Clinical Practice and the Declaration of Helsinki and its amendments at the KNUH CTC, and was approved by the institutional review board at KNUH. Written informed consent was obtained from all volunteers before their participation.
Study drugs and dosage regimen
Triflusal capsule (Disgren capsule, 300 mg), as the reference formulation, and triflusal EC capsule (Disgren enteric-coated capsule, 300 mg), as the test formulation were manufactured and provided by Myung-In Pharm. Co., Ltd (Seoul, Republic of Korea). The subjects received the test or reference formulation in multiple doses, followed by a 13-day washout period and subsequent administration of the alternative formulation. During each period, each subject received a single 900 mg oral loading dose on Day 1, followed by a 600 mg/day maintenance dose (given as two 300 mg capsules once daily) from Day 2 to 9.
Blood sampling procedures
For measurement of the plasma concentration of HTB (PK), whole blood samples were obtained at time 0 (pre-dose), and 24, 48, 96, 144, 168, 192, 192.5, 193, 194, 196, 199, 202, and 216 h after the first dose (from Day 1 to Day 10). For platelet aggregation measurements (PD), sampling was performed at time 0 (pre-dose), and 24, 48, 96, 144, 168, 192, 196, 202, and 216 h post-dose. The overall sampling schema is presented in Figure 1.
Blood (7 mL) for PK analysis was collected into tubes containing sodium heparin, and blood (3 mL) for PD analysis was collected into tubes containing 0.109 mmol/L sodium citrate. An aliquot (1 mL) of each plasma sample was placed in a microcentrifuge tube containing 0.1 mL 0.1 M HCl to prevent degradation of triflusal in the plasma. The remaining samples were chilled promptly on crushed ice to maintain the distribution ratio between triflusal and HTB [9].
HTB plasma concentration measurements
Concentrations of HTB, rather than triflusal, were measured in this study because the level of triflusal is too low to be detected after 4 h when administered orally, as reported by Lee et al. [4, 10]. Blood samples were analyzed by high-performance liquid chromatography coupled with tandem mass spectrometry (HPLC-MS/MS) [7]. Quantification was performed by multiple reaction monitoring (MRM) transitions. A C18 column was used to separate the components of the samples by chromatography. Linear calibration curves were analyzed in the range of 1 to 300 μg/mL HTB (coefficient of correlation, r = 0.999). The lower limit of quantification of HTB was 1 μg/mL. The intra-day and inter-day % CVs for HTB were each less than 15%. Figure 2 shows the plasma concentrations of HTB with respect to time.
Platelet aggregation assessment
The tubes containing sodium citrate were centrifuged (1,000 rpm, 10 min) and 500 μL of platelet-rich plasma (PRP) was collected as the supernatant. The PRP was transferred to microcentrifuge tubes, which were capped to prevent changes in pH upon exposure to air, and the samples were kept on ice. To obtain 500 μL of platelet-poor plasma (PPP), the remaining sample in the original tubes was centrifuged again (3,000 rpm, 10 min) [11]. Platelet aggregation in PRP was assessed using a Chronolog optical aggregometer (Model 490-4D; Chrono-log, Havertown, PA, USA). Cuvettes were lined up along the two rows of the aggregometer, filled with 250 μL of PRP or PPP, and warmed for 5 min at 37°C [12]. To measure the degree of platelet aggregation, 0.5 μL AA was added to the cuvettes as an agonist, and light transmission aggregometry was observed for 5 min. The aggregometer was calibrated with a cuvette containing PRP, equaling 0% light transmission, as the baseline and with a second cuvette containing PPP, equaling 100% light transmission, as the reference [13]. Based on analyses of the blank plasma samples, which were validated in three individuals, the intra-day and inter-day precisions were calculated as (SD/mean) × 100 (%) and ranged from 2.1 to 6.8% and from 2.1 to 7.6%, respectively [4].
Dataset
In the current study, PK and PD data from 34 healthy volunteers who received the reference drug were used for population PK-PD analysis. In total, 476 HTB plasma concentration data points and 340 platelet aggregation (%) data points were included. The latent variables used as candidate covariates were age, height, and weight, and levels of albumin, creatinine, serum aspartate transaminase (AST), and alanine transaminase (ALT). Creatinine clearance (CLCR), calculated using the Cockcroft-Gault equation, was also included.
Population PK-PD model development
The population PK-PD analysis was conducted by non-linear mixed effects modeling using NONMEM (ver. 6.2; Icon Development Solution, Ellicott City, MD, USA). The estimate and the between-subject variability for PK-PD parameters were investigated. The first-order conditional estimation (FOCE) method with interaction option was used whenever applicable [14]. The between-subject variability of each parameter of the basic model was applied using the exponential model:
where P
ij
is the value of the j
th parameter in the i
th individual, θ
j
is the typical value of the j
th population parameter, and η
ij
is a random variable for the i
th individual in the j
th parameter. It is assumed that η
ij
is normally distributed, with a mean of zero and a variance of ω2, and that η
ij
≈ N(0, ω
ij
2). The residual variability, consisting of intra-individual variability, experimental errors, process noise, and/or model mis-specifications, was modeled using additive (ϵ
add,ij
), proportional (ϵ
pro,ij
), and combined error structures. The combined additive and proportional error model on the following equation was initially applied to explain the gaps between the observed values and those predicted by individual PK parameters:
where C
ij
is the j
th observed value in the i
th individual, and ϵ
pro,ij
and ϵ
add,ij
are the residual intra-individual variability with a mean of zero and variances of and , respectively. A similar approach was made to the PD observations. The standard errors, which give information on the precision of the parameter estimates, were explored using NONMEM’s “$COVARIANCE” step.
Models were evaluated using diagnostic scatter plots, goodness-of-fit plots, and the log likelihood ratio test (LRT). The results for LRT were considered statistically significant if the objective function value (OFV) decreased by 3.84 (p = 0.05); this was used as a cut-off criterion for model improvement.
For PK, various absorption functions and distributional (1-compartment vs. multi-compartment) models were compared. When linking PK-PD relationship, two different models – binary probability and inhibition of build-up – were tried sequentially to find the best description on the PD data. The binary probability model uses a form of relationship similar to the Michaelis-Menten equation, to explain instantaneous PK-PD relationship. In this equation, C
pred,ij
is the estimated concentration over time for individuals, EC
50
is the plasma concentration corresponding to the half-maximal response, and γ is the shape factor. To implement this approach, the PD data (% platelet aggregation) was transformed to a binary format based upon the finding that the distribution of PD observation (the maximum platelet aggregation observed in each sample) was clearly bimodal (Figure 2). The values above 74% were regarded as no inhibition of platelet aggregation (IPA), and thus as DV = 0, whereas values below 74 were regarded as IPA (DV = 1), according to a guideline presenting the normal ranges for the degree (%) of aggregation in platelet-rich plasma [15]. To find a reliable initial estimate for this model, a logistic regression procedure was performed in advance. For a binary output variable, Y, we can model conditional probability P(Y = 1 | X = x) as a function of x [16]. Let π(x) be a linear function of x, π(x) = P(Y = 1 | X = x). The model can be written as π(x) = exp(β0 + β1x) / exp(β0 + β1x) + 1. Equivalently, by a logit transformation, the logistic regression model has a linear relationship, logit(π(x)) = log(π(x) / 1 - π(x)) = β0 + β1x, where β0 is the intercept coefficient and β1 is the slope coefficient, for which the sign is positive or negative if π(x) is a strictly increasing or decreasing function of x, respectively [17]. Using the frequency of DV = 1 at each time point, the intercept (β0) and slope (β1) were estimated using SAS software (ver. 9.2; SAS Institute, Inc., Cary, NC, USA). The inhibition of build-up model was tried to evaluate whether there is a time-delay between concentration and response in the form of continuous variable. The differential equation for platelet aggregation at a certain HTB concentration is as follows [18]:
where dR/dt is the rate of change of the measured response (R) over time, k
in
is the turnover input rate for production of the response, I
max
is the fraction representing the maximal capacity by which the drug can inhibit platelet aggregation (0 ≤ I
max
≤ 1), C is the HTB concentration, IC
50
is the median inhibition concentration, and k
out
is the turnover output rate of the response.
Covariate analysis
The parameter-covariate relationship, which explains why PK and PD vary among individuals, was explored using variables in the dataset. In the exploratory analysis, scatter plots of one covariate versus another were used to assess correlations between covariates and of parameters versus covariates. The numerical procedure used generalized additive modeling (GAM), as implemented in the Xpose library (ver. 4.4.0, Department of Pharmaceutical Biosciences of Uppsala University, Uppsala, Sweden) [19] of the ‘R’ software (ver. 2.15.3; R Foundation for Statistical Computing, Vienna, Austria). In this procedure, a response random variable, Y, is regressed on the individual covariates (Xj) according to the general equation:
where the s0 is an intercept, and s1(X1), · · ·, sp(Xp) are smooth functions that are estimated in a nonparametric fashion. In the equation, candidate covariates with the greatest decrease in the Akaike information criterion were selected as final covariates in the PK model when there was a reduction of at least 3.84 compared with the previous model, and the relationship between the random variable, η, for each parameter and the candidate covariates was improved. Otherwise, they were dropped from the model.
Graphical analyses using goodness-of-fit plots – including observed-versus-population-predicted, observed-versus-individual-predicted, and residual diagnostics – were also conducted to diagnose the degree of model development.
Model evaluation
Bootstrap was recruited to check the robustness of the PK-PD model and parameter estimates. The median value and 95% confidence interval for each parameter were obtained from 1,000 bootstrap dataset. For the PK model, a visual predictive check (VPC) was implemented to further evaluate the accuracy and predictive performance. A total of 1,000 replicate simulation was performed and the 50th (estimated population median), 5th, and 95th percentiles of the created concentrations plotted against time were compared with the overlaid observed concentrations. In addition, predicted concentration-response relationship is plotted with the predicted probability of each concentration data observed at the time of PD sampling.
Sample size determination
We considered the sample size for having an 80% power at a significance level of 0.05 for bioequivalence trials. It was calculated under the condition of the equivalence limit of 22.3%, the true mean difference of 10%, and the standard deviation of 24.2%. The calculation was performed by a formula suggested from Chow et al. [20]. Volunteers were randomly assigned to one of two sequences with simple randomization using SAS software (ver. 9.2; SAS Institute, Inc., Cary, NC, USA).
Statistical analyses
The demographic data were presented using descriptive statistical analysis performed with the standard SPSS package (version 12.0 for Windows, SPSS, College Station, TX, USA). The mean, standard deviation, and range were obtained for each variable.