Dosage Individualization of Linezolid: Precision Dosing of Linezolid To Optimize Efficacy and Minimize Toxicity

ABSTRACT

The high interindividual variability in the pharmacokinetics (PK) of linezolid has been described, which results in an unacceptably high proportion of patients with either suboptimal or potentially toxic concentrations following the administration of a fixed regimen. The aim of this study was to develop a population pharmacokinetic model of linezolid and use this to build and validate alogorithms for individualized dosing. A retrospective pharmacokinetic analysis was performed using data from 338 hospitalized patients (65.4% male, 65.5 [±14.6] years) who underwent routine therapeutic drug monitoring for linezolid. Linezolid concentrations were analyzed by using high-performance liquid chromatography. Population pharmacokinetic modeling was performed using a nonparametric methodology with Pmetrics, and Monte Carlo simulations were employed to calculate the 100% time >MIC after the administration of a fixed regimen of 600 mg administered every 12 h (q12h) intravenously (i.v.). The dose of linezolid needed to achieve a PTA ≥ 90% for all susceptible isolates classified according to EUCAST was estimated to be as high as 2,400 mg q12h, which is 4 times higher than the maximum licensed linezolid dose. The final PK model was then used to construct software for dosage individualization, and the performance of the software was assessed using 10 new patients not used to construct the original population PK model. A three-compartment model with an absorptive compartment with zero-order i.v. input and first-order clearance from the central compartment best described the data. The dose optimization software tracked patients with a high degree of accuracy. The software may be a clinically useful tool to adjust linezolid dosages in real time to achieve prespecified drug exposure targets. A further prospective study is needed to examine the potential clinical utility of individualized therapy.

TEXT

Linezolid is an oxazolidinone antibacterial agent that is used for the treatment of infections caused by Gram-positive bacteria. The use of linezolid has increased in recent years because of its activity against drug-resistant organisms, its demonstrated clinical effectiveness, and its favorable pharmacokinetic (PK) properties (i.e., high oral bioavailability and extensive tissue distribution into infection sites with anatomical barriers) (1–3). Numerous studies have reported a high variability in the plasma drug exposure for patients receiving a standard fixed regimen of linezolid (600 mg every 12 h [q12h]) (4–7). Some of the PK variability of linezolid can be explained by fixed effects (e.g., body weight, renal and/or hepatic function, and the severity of illness) (4, 5, 7–9). However, a large portion of the observed variance remains unexplained. Variability in drug exposure leads to concentration-dependent therapeutic failure in some patients and an increased probability of toxicity in others (9–12). Precision dosing is potentially a way that clinical outcomes can be further optimized.

The principal aim of this study was to develop a population PK model of linezolid using a large number of hospitalized patients and then use this model to build algorithms that can individualize the dosing of linezolid. The performance of this algorithm was characterized in a separate cohort of patients.

RESULTS

Linezolid dosing and PK.

A total of 338 acutely hospitalized patients were included in the study. The demographic and relevant clinical data are summarized in Table 1. Linezolid was administered at the standard licensed regimen (i.e., 600 mg q12h) in 323/338 of patients (95.6%), at a dose of 300 mg q12h in 2/338 (0.6%), and at a dose of 600 mg q8h in 13/338 (3.8%) of patients. The intravenous (i.v.) and oral routes were used in 300 (88.8%) and 38 (11.2%) of patients, respectively. The median (range) duration of treatment was 11.0 (3 to 127) days. A total of 868 linezolid plasma concentrations obtained from 338 patients were included in the population analysis. The mean (standard deviation [SD]) number of observations per patient was 2.6 (1.7) with a range of 1 to 15. The means (SD) of the minimum concentration (C_min,ss) and maximum concentration (C_max,ss), both obtained at steady state, were 7.2 (9.5) and 19.7 (12.3) mg/liter, respectively. The first C_min,ss concentrations were subtherapeutic (<2 mg/liter) in 43.1% of the patients, in range (between 2 and 8 mg/liter) in only 28% of them and supratherapeutic (>8 mg/liter) in 28.9%. Figure 1 shows the linezolid plasma concentration-time profiles for all patients in this study.

TABLE 1.

Demographic and clinical characteristics of the included patients

Parametera	Result
Total patients, n	338
Male, n (%)	221 (65.4)
Median age, yr (range)	68 (21–98)
Median body wt, kg (range)	75 (37.5–193)
Median BMI, kg/m2 (range)	26.7 (13.4–70.8)
Critically ill patients, n (%)	225 (66.6)
Initial linezolid dose, n (%)
300 mg q12h i.v.	2 (0.6)
600 mg q12h oral	38 (11.2)
600 mg q12h i.v.	285 (84.3)
600 mg q8h i.v.	13 (3.8)
Median serum creatinine, mg/dl (range)*†	1.08 (0.17–13.6)
Median GFR, ml/min/1,73m2 (range)*†	63.2 (3.3–196.5)
GFR < 90 ml/min/1.73 m2, n (%)*	211 (62.4)
GFR > 130 ml/min/1.73 m2, n (%)*	19 (5.6)
Renal replacement therapy, n (%)	39 (11.5)
Serum albumin, g/dl (range)	2.6 (1.4–5.1)
Median total serum protein, g/dl (range)	5.3 (3–8.4)
Median total leukocytes, cells × 109/liter (range)	12.9 (0.32–102.5)
Liver cirrhosis, n (%)	25 (7.4)
Child Pugh A	4 (1.2)
Child Pugh B	13 (3.8)
Child Pugh C	8 (2.4)
Median hospital stay, days (range)	40.0 (3–753)
Mortality during linezolid treatment, n (%)	27 (8.0)
In-hospital mortality, n (%)	91 (26.9)

^aGFR, glomerular filtration rate. *, values at the beginning of linezolid treatment; †, estimated using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) formula (40).

FIG 1.

Mean population (A) and individual (B) predicted concentrations versus observed concentrations of linezolid in plasma. The broken line is the line of identity (i.e., observed = predicted concentrations).

Population PK model.

A three-compartment pharmacokinetic model that consisted of absorptive, central, and peripheral compartments was fitted to the data. Linezolid was administered as a bolus input into the absorptive compartment (gut) and as zero-order time delimited i.v. input into the central compartment over 1 h. Drug was cleared from the central compartment, and this was modeled as a first-order process.

Estimates for measures of central tendency, dispersion, and the 95% confidence intervals for the population PK parameters from the final model are shown in Table 2. Figure 2 shows the observed-predicted values using the median parameter values both before and after the Bayesian step. After maximum a posteriori probability-Bayesian estimation, a linear regression of observed versus predicted values had an intercept and a slope of −0.157 and 0.959, respectively, and a coefficient of determination of 0.96. The bias and imprecision were both acceptable (bias = 1.16 mg/liter; imprecision = 19 mg/liter). Covariates such as age, body weight, the presence of liver cirrhosis, and critical illness did not significantly improve the fit of the model to the data and were not considered for further model building.

TABLE 2.

Population pharmacokinetic parameters of linezolid^a

Parameter	Mean	SD	Median	95% CI
CL (liters/h)	7.720	7.191	5.126	3.809–6.267
Vc (liters)	33.690	17.697	31.493	29.047–35.218
Ka (h−1)	7.036	10.314	1.900	1.236–2.457
F	0.804	0.207	0.880	0.841–0.901
Kcp (h−1)	4.418	8.531	0.243	0.217–0.437
Kpc (h−1)	7.633	11.980	0.546	0.306–1.079

^aCL, clearance; V_c, volume of the central compartment; F, bioavailability; K_a, first-order absorption rate constant. K_cp and K_pc are the first-order intercompartmental rate constants. CI, confidence interval.

FIG 2.

Linezolid plasma concentration-time profiles for patients receiving oral or i.v. linezolid. Sampling was performed in the majority of cases after the second or third day of treatment.

Probability of target attainment.

After the administration of the standard dose of linezolid (600 mg i.v. q12h), the probability of target attainment (PTA) for achieving 100% total drug T_>MIC in plasma for the different tested MIC values is shown in Fig. 3. With this fixed dose, an optimal PTA could only be achieved for isolates with a MIC of 0.5 mg/liter when estimating the 100% T_>MIC and for isolates with a MIC of 1 mg/liter when an AUC/MIC of ≥100 mg ⋅ h/liter was considered. In any case, an optimal exposure could not be achieved for isolates with a MIC value of 4 mg/liter, which is the current susceptibility breakpoint according to EUCAST (13). The regimen of linezolid needed to achieve a PTA of ≥90% for all susceptible isolates classified according to EUCAST was estimated to be as high as 2,400 mg q12h, which is four times higher than the maximum licensed linezolid dose. In addition, the PTAs for achieving trough concentrations within the therapeutic range (between 2 and 8 mg/liter) with different linezolid regimens (300 mg i.v. q12h, 600 mg i.v. q24h, 600 mg i.v. q12h, 600 mg i.v. q8h, and 1,200 mg i.v. q24) were determined and are shown in Fig. 4.

FIG 3.

PTA for achieving 100% T_>MIC and AUC/MIC ≥ 100 mg · h/liter in plasma of linezolid (600 mg/12 h administered as a 1-h infusion) during the third day of treatment (from 48 to 72 h after the start of the treatment).

FIG 4.

PTA for achieving trough concentrations within the therapeutic range (between 2 and 8 mg/liter) during the third day of treatment with different doses of linezolid (from 48 to 72 h after the start of the treatment).

Validation of the linezolid software dosing controller for predicting individual concentrations and dosage individualization.

The test data set from 10 patients that were used to assess the performance of the controller had the following demographic and clinical characteristics. Eight patients (80%) were male with a median (range) age of 74.5 (59 to 88) years. The median (range) body weight and body mass index (BMI) were 72.5 (60.0 to 100 kg) and 26.1 (21.3 to 31.6) kg/m², respectively. Six patients (60%) were critically ill, and all of them received a standard regimen of linezolid (600 mg q12h). Intravenous dosing was used in 9 (90%) patients, and only one received oral administration. There was a median of three observations (range, two to five) per patient, and the measured target concentrations of linezolid ranged from 0.8 to 36.6 mg/liter.

The algorithm was able to accurately track the observed PK of each of these 10 patients. The combined individual patient observed-versus-predicted concentrations of linezolid from all 10 patients are shown in Fig. 5. The r² value of the linear regression was 0.998, with a slope of 0.985 and an intercept of 0.175. The median (interquartile range) bias and percent bias for the predictions of the target concentrations were −0.1 (0.17) mg/liter and −0.5 (1)%, respectively. For the dosage prediction, the bias and percent bias were 101.5 (237.3) mg and 16.9 (39.6)%, respectively.

FIG 5.

Observed-versus-predicted linezolid concentrations of all 10 patients. The solid line is the line of identity (observed = predicted concentrations), and the dashed line is the linear regression line fitted to the pooled data with a slope of 0.985, an intercept of 0.175 and a R² value of 0.998.

Example of the clinical utility of the linezolid dose optimization software.

The potential utility of the controller is illustrated in Fig. 6, which depicts a typical patient for whom the standard linezolid regimen resulted in a trough concentration significantly higher than desired target concentration and associated with an increased risk of toxicity. The administration of a lower dose suggested by the algorithm was predicted to rapidly allow for the achievement of a linezolid trough concentration that was safe and effective (Fig. 6).

FIG 6.

Graph showing the potential advantage of a controller for the individualization of linezolid concentrations. A patient receiving a standard linezolid dose (600 mg i.v. q12h infused over 1 h) was sampled at 1 h (peak or C_max) and 12 h after linezolid dose administration (trough or C_min) on two occasions (past data in red color). The second observed trough concentration is much higher than the superior limit of the therapeutic range (8 mg/liter), being potentially toxic. The controller predicted two much lower maintenance doses of 394 and 215 mg i.v. q12h (a half of the initial dose) infused over 1 h to achieve a therapeutic trough concentration of between 5 and 7 mg/liter (future data in green color). The solid line represents the mean predicted concentration-time profile of the patients.

DISCUSSION

Therapeutic drug monitoring (TDM) is a well-established adjunct for some antimicrobial classes such as the glycopeptides and aminoglycosides (14). More recently, TDM has been used for other classes of drugs, such as the beta-lactams, to improve antimicrobial activity (especially in special populations), and to minimize the emergence of AMR (15). Despite numerous studies of linezolid PK demonstrating significant PK variability (16–19) and knowledge about drug exposures that are associated with safety and efficacy, active dose adjustment of linezolid is usually not considered (18).

Monte Carlo simulations performed with the PK model used to build the control software suggested that the use of the standard fixed regimen of linezolid (600 mg q12h) did not allow the achievement of pharmacodynamic targets for isolates with an MIC up to 4 mg/liter, which is the ECOFF value for S. aureus and E. faecium. For these strains, higher dosages may be required to achieve favorable clinical outcomes. However, dosage escalation may lead to potentially toxic drug exposures in more than 30% of the patients. This limitation provides the principal justification for considering dosage individualization to achieve drug exposures of linezolid that are safe and effective.

Some clinicians support routine TDM of linezolid as a potential strategy to ensure an effective and safe use of this agent (20, 21). Although linezolid is generally well tolerated and may cause only mild gastrointestinal adverse events, its use has been also related to hematological toxicity, particularly in prolonged linezolid courses, in patients with liver disease, renal failure, or with a reduced baseline platelet count. Pea et al. concluded in a study that TDM may be useful for improving safety outcomes in adult patients with infections that require prolonged treatment with linezolid and that dosage adjustments based on plasma concentrations may help to avoid drug-induced adverse events (21). Algorithms to enable dosage adjustment for linezolid have not been developed or used, and this has been an impediment for routine use of TDM for linezolid. This study provides the necessary tools to begin addressing that problem.

Here, an algorithm was developed to achieve concentration targets for linezolid that are safe and effective. The final PK model was a three-compartment linear model with an absorptive compartment, which is consistent with previously developed PK population models of linezolid (22, 23). Since our primary aim was to develop an algorithm for dosage individualization, we selected the simplest structural model without extensive covariate building.

The algorithm that has been constructed to facilitate individualized dosing needs to be further characterized in a prospectively study, as has previously been done with other antimicrobial agents (24, 25). Regimen planning should be performed by an experienced clinician with due consideration for factors with spurious results (e.g., suboptimal compliance and medication errors). However, the software may be a clinically useful tool for a decision support tool for precision dosing of linezolid as a way to optimize the use of linezolid in patients with serious Gram-positive infections.

MATERIALS AND METHODS

Study design.

All patients ≥18 years and receiving linezolid for > 72 h undergoing routine TDM in the Hospital del Mar, Barcelona, Spain, between January 2011 and June 2018 were eligible for inclusion. The study was approved by the Comitè Etic d’Investigació Clínica del Parc de Salut Mar (2016/6987/I).

Treatment with linezolid.

Patients received linezolid because of a documented or suspected infection with a Gram-positive pathogen. Linezolid was used at the discretion of the treating physician. All patients initially received the standard recommended dosage (600 mg every 12 h, i.v. and/or orally).

Therapeutic drug monitoring of linezolid.

TDM of linezolid plasma concentrations was performed as a part of routine clinical practice. In the majority of our patients, the first plasma sampling was obtained at ≥48 h after treatment initiation, when the steady state has been presumably achieved. Blood samples (4 ml) were collected immediately prior to the administration of the next dose (C_min,ss) and then 30 min after the end of a 1-h i.v. infusion or at 2 h after oral dosing (C_max,ss). In some patients, an intensive plasma sampling was performed. According to the TDM results, the initial dose of linezolid was modified to obtain a C_min,ss within the therapeutic range of efficacy 2 to 8 mg/liter.

Bioanalysis of linezolid concentrations.

Blood samples were collected in heparinized tubes, which were immediately centrifuged (3,000 × g for 10 min at 4°C) to obtain plasma, which was then stored at −80°C until bioanalysis. Linezolid concentrations were determined using a validated high-performance liquid chromatography (HPLC) method with minor modification of a previously described method (26).

Briefly, for each plasma sample, 100 μl was mixed with 100 μl of methanol and then vortexed for 10 s. The mixture was then centrifuged for 5 min at 15,000 × g in a refrigerated centrifuge. A total of 50 μl of the supernatant was injected. The HPLC equipment was an Alliance e2695 (Waters Cromatografía, S.A., Barcelona, Spain). The mobile phase consisted of a mixture of sodium acetate buffer (pH 3.4) and acetonitrile (80:20 vol/vol) delivered in an isocratic flow at 1.3 ml/min. The chromatogram run time was 10 min, and the ultraviolet detector wavelength was set at 250 nm. The assay response was linear (coefficient of linearity >0.99) over the dynamic range of the assay (0.5 to 30 mg/liter in plasma). The limit of quantification was 0.5 mg/liter in plasma. Imprecision values were <15% over the entire range of calibration standards, and accuracy was within the range of 85 to 115% for all concentrations.

Data collection.

The following data were collected from each included patient: age, gender, weight, renal function (serum creatinine and serum urea at the start of linezolid treatment), albumin, total serum protein, the presence of liver cirrhosis and severity (Child Pugh score) (27), the need for renal replacement therapies, and linezolid data (dose, route of administration, and TDM data).

Population pharmacokinetic modeling.

Population pharmacokinetic modeling was performed using the nonparametric adaptive grid (NPAG) algorithm embedded within Pmetrics (28, 29). One-, two-, and three-compartment models were fitted to the data. Elimination from the central compartment and intercompartmental transfer were modeled as first-order processes. Data were weighted using the inverse of the estimated assay variance and additional process noise.

Age, gender, actual body weight, serum creatinine, serum urea, serum albumin, total serum proteins, the presence of liver cirrhosis, critical illness, or the need of renal replacement therapies were evaluated as covariates using stepwise linear regression. Potential covariates were separately entered into the model and retained if their inclusion resulted in a statistically significant improvement in the log-likelihood value and/or in the observed-predicted plots.

The fit of each model to the data was assessed using a linear regression of observed-predicted values both before and after the Bayesian step. The mean prediction error and the mean bias-adjusted squared prediction error were used to assess bias and imprecision, respectively. Models were compared by calculating twice the difference in log-likelihood values.

A three-compartment linear model, with an absorption compartment, and first-order input and clearance from the central compartment, best described the data. The structure of the final PK mathematical model fitted to the study data was described by the following equations.

$$\text{dX}\left(\text{1}\right)/\text{dt\hspace{0.17em}}=\hspace{0.17em}-K_a\times \text{X}\left(\text{1}\right)$$

$$\text{dX}\left(\text{2}\right)/\text{dt\hspace{0.17em}}=\hspace{0.17em}{K}_a\times \text{X}\left(\text{1}\right)\hspace{0.17em}–\hspace{0.17em}{K}_{\text{cp}}\times \text{X}\left(\text{2}\right)\hspace{0.17em}+\hspace{0.17em}{K}_{\text{pc}}\times \text{X}\left(\text{3}\right)\hspace{0.17em}–\hspace{0.17em}\left(\text{CL}/V_{\text{c}}\right)\times \text{X}\left(\text{2}\right)$$

$$\text{dX}\left(\text{3}\right)/\text{dt}=\hspace{0.17em}{K}_{\text{cp}}\times \text{X}\left(\text{2}\right)\hspace{0.17em}–\hspace{0.17em}{K}_{\text{pc}}\times \text{X}\left(\text{3}\right)$$

Compartments 1, 2, and 3 denote the absorptive, central and peripheral compartments, respectively, with X(1), X(2), and X(3) representing the amount of drug (mg) in each respective compartment. CL is clearance from the central compartment (liters per hour); V_c (liters) is the volume of the central compartment, and K_cp and K_pc are first-order intercompartmental transfer rate constants (h⁻¹).

Other pharmacokinetic calculations.

The plasma AUC for each patient was estimated using the Bayesian posterior parameter estimates using the trapezoidal method embedded within Pmetrics. The average AUC_0–24 (AUC_0–24av) was calculated by dividing the cumulative AUC of each patient by the total time in hours and multiplying by 24 h.

Targets of linezolid exposure and toxicity.

The therapeutic range for linezolid for efficacy has been defined as a C_min,ss of 2 to 8 mg/liter (11, 12, 17, 21, 30, 31). An optimal pharmacokinetic/pharmacodynamic (PK/PD) ratio for linezolid antibacterial efficacy has been defined as a 100% time >MIC (i.e., C_min,ss ≥ MIC) in clinical studies (32). A trough concentration of linezolid of ≥2 mg/liter is a value often used in clinical practice and it has been associated with >80% probability of bacterial eradication (20). A C_min,ss > 8 mg/liter was used as the upper bound because it has been associated with a higher risk of thrombocytopenia in clinical studies (11, 12). However, it has to be considered that a much lower cutoff point (a trough concentration of linezolid of 0.19 mg/liter) has been associated with a 50% maximal mitochondrial toxicity in an in vitro hollow-fiber infection model system (33).

The AUC/MIC has also been identified as the relevant PK/PD index for linezolid in animal and clinical studies (33, 34). Values of AUC/MIC of 100 and 300 have been suggested for efficacy and toxicity, respectively (34). Again, in the previously described hollow-fiber model a much lower value of AUC of about 150 mg ⋅ h/liter was related to mitochondrial toxicity (33).

Monte Carlo Simulations predicting optimal exposure and probability of toxicity.

Monte Carlo simulations (n = 1,000) of plasma concentrations were used to calculate the 100% time >MIC and the AUC₂₄/MIC of 100 at steady state (i.e., on day 3 of treatment, from 48 to 72 h posttreatment initiation) after the administration of a fixed regimen of 600 mg q12h i.v. The simulated AUC from 48 to 72 h were also compared to the individual predicted AUC after the Bayesian step during the same period of time.

Linezolid software dosing controller construction.

The final population PK linezolid model was incorporated into BestDose (29, 35). This software has been previously validated to individualize dosages of other antimicrobials such us vancomycin, voriconazole, piperacillin, and antiretroviral therapy (24, 25, 36–39).

The structural mathematical equations of the model relating input (linezolid dosing) to output (linezolid plasma concentrations) and the discrete joint probability distribution of pharmacokinetic parameters (support points) were included in the controller. These support points are the prior Bayesian sets of PK parameters that can be used to explain the future observed concentrations of the patients. The software used user-supplied data regarding dosing (dose and timing of drug administration) and the observed concentrations (past data) to find the least biased and most precise dosage regimen relative to a target concentration (future dosage), as previously described (25). The nonparametric collection of discrete support points for linezolid obtained from the previously designed population pharmacokinetic model was used as a “population prior” distribution of the values of the PK parameters. This distribution is the basis for the individualization of the linezolid dose by a multiple-model stochastic control. When information about a new patient (dosing history of linezolid and measured concentrations) is entered to BestDose, the probabilities of each support point are updated to new “Bayesian posterior” probabilities depending on the ability of each point to describe the PK behavior. After that, the software is able to calculate the dose that minimized the weighted squared error between the real target concentration from the parameter values of each point in the Bayesian posterior and the target.

Linezolid software dosing controller validation.

To validate the performance of the software, a separate data set obtained from 10 patients receiving linezolid and undergoing TDM was used. We included all measured linezolid concentrations of each patient except the last one, which was used to assess the predictive performance of the controller and used as the target concentration for dosage identification.

The performance of the model was first tested by visually comparing the observed versus last predicted concentrations of each individual patient. In addition, all combined observed versus predicted concentrations from all 15 patients were plotted, and a linear regression analysis was performed. In a second stage, the predicted optimal dosage calculated by BestDose to obtain the last target concentrations was compared to the real administered dose of linezolid, which was 600 mg in the majority of our patients.

Finally, the bias and percent bias for both target concentrations and target dose were calculated. The bias target was equal to the predicted concentration or dose minus the actual concentration or dose. The percent bias target was calculated as follows: (predicted concentration or dose – actual target concentration or dose)/actual target concentration or dose.

The final model was also evaluated graphically and statistically by visual predictive checks (VPCs) and normalized prediction distribution errors (NPDE) (see Fig. S1 and S2 in the supplemental material). A total of 1,000 data sets were simulated using the final population model parameters. For the VPCs, the 5th, 50th, and 95th percentiles of the simulated concentrations were processed using the R platform, plotted against elapsed time, and compared to observed concentrations. NPDE results were summarized graphically by default as provided by the NPDE R package (version 1.2) using (i) a Q-Q plot (where Q is quantile) of the NPDE and (ii) a histogram of the NPDE. The normal distribution of NPDEs (P 0.115 in the Shapiro-Wilk normality test) confirmed the adequacy of the model for dosing simulations.