Population Pharmacokinetics of Dexmedetomidine in Critically Ill Patients

All studies included adult patients who were initially intubated, mechanically ventilated and expected to require light to moderate sedation for at least a further 24 h. The main exclusion criteria were (1) acute severe intracranial or spinal neurological disorder due to vascular causes, infection, intracranial expansion or injury; (2) uncompensated acute circulatory failure at time of randomisation (severe hypotension with mean arterial pressure <55 mmHg despite volume and pressors); (3) severe bradycardia (heart rate <50 beats/min); (4) atrioventricular-conduction block II–III (unless pacemaker installed); (5) severe hepatic impairment (bilirubin >101 μmol/L); (6) burn injuries and other injuries requiring regular anaesthesia or surgery; (7) use of centrally acting α₂ agonists or antagonists (e.g. clonidine, titzanidine, apraclonidine and brimonidine) within 24 h prior to randomisation; or (8) investigators’ own judgement.

The patients received an initial infusion of 0.7 μg/kg/h for 1 h. Thereafter, the dosing was titrated to clinical effect to maintain patients in the pre-defined target sedation range (Richmond Agitation and Sedation Scale 0 to −3 in all cases) using fixed dose levels ranging between 0.2 and 1.4 μg/kg/h. The maximum duration of treatment was 14 days.

Blood samples were taken at the following times: baseline, 1 h (±15 min) after starting study treatment and every day at approximately the same time until the end of study treatment. Additionally, two follow-up samples were taken at 24 and 48 h after the end of study treatment.

Concentrations of dexmedetomidine in EDTA plasma samples were determined with high-performance liquid chromatography–tandem spectrometry (HPLC-MS/MS; Shimadzu Prominence HPLC, Kyoto, Japan) and mass spectrometric detection (AB Sciex API4000 mass spectrometer, Toronto, ON, Canada), as previously described [17]. The lower limit of quantification was 0.02 ng/mL. The within- and between-run precision of the assay (coefficient of variation) was within 7.5 % in the relevant concentration range. Deuterated medetomidine was used as the internal standard.

As part of the safety monitoring, the values of aspartate aminotransferase (AST), alanine aminotransferase (ALT), bilirubin, creatinine clearance [18] and albumin were measured at baseline and on days 2, 4, 6, 9 and 14 after start of study drug infusion, and at 48 h post-dose. The baseline values of these markers were used for each patient as predictors of dexmedetomidine pharmacokinetics. If no baseline data were available for a patient, the average of all measurements from that patient was used. If no measurements from any time point were available for a patient, the median value of the whole population was substituted for the covariate value of that patient.

Data were analysed using the NONMEM^® software (version 7.2; ICON Development Solutions, Ellicott City, MA, USA) [19] with Intel Visual Fortran 11 compiler and Perl-speaks-NONMEM [20]. The model was fitted to data using the stochastic approximation expectation/maximisation algorithm [19], with 5,000 burn-phase iterations and 2,000 accumulation-phase iterations. The standard errors were calculated with importance sampling [19], using 20 iterations with 3,000 samples per subject and two degrees of freedom because of the sparseness of the data. One-compartment and two-compartment models with first-order elimination were tested before the inclusion of covariates to describe the time–concentration data.

Between-subject variability was modelled using log-normal distributions of individual parameter values, as shown in Eq. 1:where P _i is the parameter value of the ith subject, θ _pop is the typical (median) value of this parameter and η is a random variable with mean of zero and variance of ω ². Residual error was implemented as a combination of additive and proportional residual errors.

A full random-effects covariate model was used to quantify the relationship between pharmacokinetic parameters and covariates [21]. Briefly, a full random-effects covariate model uses random effects to both quantify the variability in pharmacokinetic parameters and to describe the individual values of observed covariate values. The covariate values are included in the dataset as observations. A full covariance matrix is estimated for the random effects, which means that the correlations between pharmacokinetic parameters, correlations between covariates, and the correlations between pharmacokinetic parameters and covariates are estimated. Advantages of this approach are that (1) it is not sensitive to correlated covariates, which means that all potential covariates can be included in the model; and (2) it may be more stable than a covariate model based on fixed effects. The full random-effects covariate model was considered the most appropriate approach for this project because many of the covariates were correlated. Furthermore, the dexmedetomidine dosing is based on dose titration by response. Because of this, there was considered to be no need to provide dosing guidelines based on covariates, unless dramatically altered pharmacokinetics could be associated with any single covariate. The following covariates were included in the model: body weight, age, creatinine clearance, AST, ALT, bilirubin and albumin. Log-normal distributions were used to describe the between-subject variability in covariate values. Prognostic indicators such as Simplified Acute Physiology Score (SAPS) were not applied in all studies and so could not be used as a covariate.

As an additional verification step, the strongest covariate relationships that were observed in the full covariate model were also tested for significance one at a time with the Likelihood Ratio Test. Briefly, a model without any covariates or covariances between random effects was used as the base model. Candidate covariates were included as predictors with a power model, as shown in Eq. 2:where COV _i is the individual value of the covariate, COV _std is the reference value of covariate, and θ _exp is an estimated parameter signifying the relationship between the covariate and the parameter. This way, p-values could be calculated for the significance of the covariates by comparing objective function values. The difference in objective function values between nested models is chi-square distributed. The estimation method used in this step was QRPEM [19] using 80 iterations with 3,000 samples per subject.

A subset of the whole dataset was used for an additional analysis of steady-state concentrations (C _ss). Based on previous work [11], samples taken from patients after 15 h (five half-lives) of continuous infusion with a constant infusion rate (R _inf) were considered to be at steady state. From these samples, the linearity of dexmedetomidine pharmacokinetics in doses up to 1.4 μg/kg/h was assessed.

Briefly, the analysis consisted of calculating the clearance (CL) of dexmedetomidine based on these single observations, and plotting the calculated CLs against R _inf. The C _ss of a drug are dependent only on R _inf and CL of the drug (Eq. 3) [22].Therefore, the CL of the drug can be calculated as the drug R _inf divided by the C _ss of the drug. The observed CL versus R _inf was plotted and a linear model was fitted. The main interest was whether the slope of CL, as a function of R _inf, is different from zero. If the metabolism of dexmedetomidine became saturated at higher doses, then a negative trend would be visible in the plot of CL versus R _inf.

A total of 527 patients were included in the study. The overwhelming majority (96 %) of patients were Caucasian. The mean age (± standard deviation) was 62 (±15) years and there were more males (65 %) than females in the study population, which is typical of ICU settings [23]. Other relevant demographic factors have been summarised in Table 1. A total of 21 covariate values were missing, and population median values were substituted in their place.

There were a total of 2,144 dexmedetomidine concentrations above the limit of quantification available in the dataset. Figure 1a presents the number of samples taken after different R _inf. In total, there were 47 plasma samples taken during dexmedetomidine infusion, which contained dexmedetomidine concentrations below the limit of quantification. Furthermore, 95 % (n = 458) of the samples taken at baseline, and 72 % (n = 634 out of 875) of the samples taken at 24 and 48 h after end of infusion, were below the limit of quantification. Samples below the limit of quantification were excluded from the analysis (M1 method introduced by Beal [24]). The M3 method, which consists of treating the samples below limit of quantification (BLQ) as censored and maximising the likelihood for them being censored, was also tried. The M3 method resulted in similar parameter estimates [less than 1 % difference in population estimates of CL and volume of distribution (V _d)], but the NONMEM^® software failed in calculating the standard errors of parameters when the M3 method was used. Therefore, the M1 method was considered the most appropriate method for treating the BLQ observations.

Some atypically high concentrations of dexmedetomidine were encountered. The highest concentration was 383 ng/mL. There were four other unexpectedly high concentrations: 89, 82, 56 and 48 ng/mL. These samples were reanalysed and the same results were observed. These outlier concentrations occurred in four distinct patients, and they were mostly not preceded nor followed by atypically high dexmedetomidine concentrations.

One possible explanation for the high concentrations encountered is that the blood sample could have been taken downstream from the study drug infusion, for example from the same arm. To test this hypothesis, a semi-quantitative metabolite analysis was made for all samples of all the individuals from whom dexmedetomidine concentrations higher than 30 ng/mL were observed. The analysis showed no increase in metabolite concentrations during or after the high dexmedetomidine concentrations (data not shown). Based on this evidence, it seems that these high dexmedetomidine concentrations were not likely to reflect the true venous concentrations in patients and the concentrations were excluded from the model-building process. The final model was run both with and without these concentrations and both results are reported.

Figure 1b presents the time–concentration data for all individuals without the concentrations above 30 ng/mL. The numbers of blood samples per patient are summarised in Fig. 2a. The average duration of treatment was 2 days 14 h. The treatment durations of the patients are summarised in Fig. 2b.

Dexmedetomidine time–concentration data were best described with a one-compartment model. Although the two-compartment model resulted in a decreased objective function value (p < 0.001), the parameter values were highly dependent on initial estimates and resulted in implausible results, such as the distribution half-life ranging between 32 s and 1.5 h depending on the run. The previously reported distribution half-life of dexmedetomidine is 6 min [8, 9], and the blood sampling in the current study was sparse. Therefore, the data were not considered adequate to identify the parameters of a two-compartment model, and the one-compartment model was chosen.

The typical CL of dexmedetomidine was 39 L/h and V _d 104 L. The final model (with and without outlier concentrations) parameter estimates and standard errors are presented in Table 2. The inclusion of outlier concentrations had little impact on the estimates of CL (38 L/h) or V _d (108 L). However, there were increases in between-subject variability and standard errors of estimates when the outlier concentrations were included. The observations versus model predictions are shown in Fig. 3a, b.

The correlations between pharmacokinetic parameters and covariates are reported in Table 3. The strongest correlation (31 %) was found between body weight and CL; other covariates with higher than 10 % correlation to CL were AST and bilirubin (both with inverse correlation). Only albumin had a correlation to V _d that had a relative standard error less than 100 % (inverse correlation of −12.7 %, standard error of 9.41 %).

When body weight was included as a predictor of CL into the base model, there was a significant improvement (p < 0.001) in the model, and the estimate of θ _exp was 0.76, indicating an almost linear relationship. AST and bilirubin were also significant predictors of CL (p < 0.01 and p < 0.05) and their respective estimates of θ _exp were −0.067 and −0.091. Albumin was a significant predictor of V _d (p < 0.001) with an θ _exp estimate of −0.49. Table 4 presents the predicted changes in CL and V _d for the 2.5th and 97.5th percentiles of the covariate values.

A total of 643 observations (out of the 2,144) at R _inf equal to or below 1.4 μg/kg/h were included in the analysis of C _ss (Fig. 4a, b). One steady-state sample was considered an outlier and excluded because of being over 30 ng/mL. The calculated CLs versus dexmedetomidine R _inf are presented in Fig. 4c. Although the linear model estimated a CL slope significantly different from zero (p < 0.05), the estimate of slope was slightly positive. Based on this, the CL of dexmedetomidine does not decrease at higher doses, and it seems that no saturation of metabolism is evident in continuous dexmedetomidine infusions up to 1.4 μg/kg/h.