Pharmacokinetics and Pharmacodynamics of Intensive Antituberculosis Treatment of Tuberculous Meningitis

The most effective antituberculosis drug treatment regimen for tuberculous meningitis is uncertain. We conducted a randomized controlled trial comparing standard treatment with a regimen intensified by rifampin 15 mg/kg and levofloxacin for the first 60 days. The intensified regimen did not improve survival or any other outcome. We therefore conducted a nested pharmacokinetic/pharmacodynamic study in 237 trial participants to define exposure–response relationships that might explain the trial results and improve future therapy. Rifampin 15 mg/kg increased plasma and cerebrospinal fluid (CSF) exposures compared with 10 mg/kg: day 14 exposure increased from 48.2 hour·mg/L (range 18.2–93.8) to 82.5 hour·mg/L (range 8.7–161.0) in plasma and from 3.5 hour·mg/L (range 1.2–9.6) to 6.0 hour·mg/L (range 0.7–15.1) in CSF. However, there was no relationship between rifampin exposure and survival. In contrast, we found that isoniazid exposure was associated with survival, with low exposure predictive of death, and was linked to a fast metabolizer phenotype. Higher doses of isoniazid should be investigated, especially in fast metabolizers.

ng/mL for isoniazid and 8.00 to 4,830 ng/mL for ethambutol. Precisions and accuracies of quality control (QC) samples during the analysis of plasma clinical samples are shown in Table S6. The simultaneous quantification of the 4 analytes in human CSF was challenged by the low concentrations, a significant matrix effect, and coeluting endogenous interferences. For those reasons, only rifampin and isoniazid were quantified in CSF. The CSF samples were processed following the method described above for plasma, but with one additional step before the protein precipitation: 20 mL of a Bovine Serum Albumin (BSA) solution (at 200 mg/mL in MS grade water) was added to 100 mL of CSF samples. The measurable ranges (LLOQ to ULOQ) in CSF were 36.0 to 4,600 ng/mL for isoniazid and 32.0 to 1,600 ng/mL for rifampin. Precisions and accuracies of QC samples during the analysis of CSF clinical samples are shown in Table S6.

HPLC analysis of levofloxacin
Levofloxacin concentrations in plasma and CSF were measured by high-performance liquid chromatography (HPLC) with time-programmed fluorescence detection after solid-phase extraction according to in-situ published method (1). The LLOQ of both plasma and CSF samples were 20 ng/mL. Total precision of all quality control plasma and CSF samples (low, middle and high concentrations) was <5.21%.

PK/PD modelling
The population PK analysis was performed using nonlinear mixed-effects modelling in the software NONMEM (version 7.4, ICON Development Solutions, Ellicott City, MD, USA), compiled using gFortran (version 4.60). Perl-speaks-NONMEM (PsN; version 4.6.0) and R (version 3.2.0, http://www.r-project.org/) were used to evaluate the goodness of fit and output visualizations. The first-order conditional estimation method including η-ε interactions (FOCE-I) and the Laplace algorithm were used throughout the population PK and PD model-building procedure, respectively. Discrimination between models during the model building phase was based on standard visual diagnostics and the objective function value (OFV), calculated as proportional to twice the log-likelihood of the data. A reduction in OFV (∆OFV) of 3.84 and 6.64 was considered a significant improvement at p < 0.05 and p < 0.01, respectively, between two hierarchical models after inclusion of one additional parameter (one degree of freedom difference).

PK modelling
If the proportion of samples below the LLOQ was low (<5%), these LLOQ samples were omitted from the model development process (M1 approach) and not considered further (2). If the proportion of LLOQ samples was non-negligible (>5%), LLOQ concentration-data were evaluated by comparing the proportion of predicted and observed concentrations below the LLOQ, using categorical visual predictive checks. If model misspecification was present when omitting LLOQ data, the M6 (imputing the first concentration below LLOQ within a patient as half of the LLOQ) and M3 (maximizing the likelihood to predict censored data) methods were evaluated. Different disposition models, including one-and two-compartment disposition models, were tested for each drug. A flexible transit-absorption model, with a stepwise addition of a fixed number of 1-10 transitcompartments, was employed to describe the absorption process. One CSF compartment was added in the models (when CSF concentration was available) to describe the transfer process between the plasma and CSF, with a parameter PC (partition coefficient) to quantify the transfer between the central and CSF compartments. Inter-individual variability was added exponentially to all parameters, and assumed to be normally distributed with a zero mean and variance ω 2 . Relative bioavailability (F) was fixed to unity in the population, allowing quantification of the inter-individual variability in the absorption process. The residual unexplained variability, assumed to be normally distributed with a zero mean and variance σ 2 , was modelled as an additive error on log-transformed concentrations, which is approximately equivalent to an exponential residual error on an arithmetic scale. Simulation/evaluation-based diagnostics was used to identify concentration-time data outliers (simeval command in PsN) within each drug model. The final population PK models were re-estimated after removal of these outliers.

Model evaluation
Basic goodness-of-fit diagnostics were used to evaluate systematic errors and model misspecification for all population PK models. The sampling importance resampling (SIR) approach (3) was used to calculate parameter uncertainty in the final models (samples = 2,000, resamples = 1,000). The overall predictive performance of the final models were evaluated using simulation-based diagnostics (i.e. predictioncorrected visual predictive checks, n = 2,000 simulations).

Handling of data below the LLOQ
For rifampin, 42/1,249 (3.4%) of plasma samples and 55/708 (7.8%) of CSF samples were below the LLOQ. M1 and M6 approaches were applied to the plasma and CSF LLOQ samples, respectively, and the categorical visual predictive check for censored data showed good agreement between predicted and observed data below the LLOQ. For isoniazid, 118/1,249 (9.4%) of plasma samples and 61/708 (8.6%) of CSF samples were below the LLOQ. The M1 approach was applied to both plasma and CSF LLOQ samples, and the categorical visual predictive check of censored data showed good agreement between predicted and observed data below the LLOQ. For levofloxacin, 10/500 (2.0%) of plasma samples and 6/217 (2.3%) of CSF samples were below LLOQ, and omitted accordingly (M1 approach). For ethambutol and pyrazinamide, 7/584 (1.2%) and 24/1,063 (2.3%) of plasma samples were below the LLOQ, and omitted accordingly (M1 approach).

SUPPLEMENTARY REFERENCES
(1) Van Toi, P.     F is the relative bioavailability. CL/F is the elimination clearance. V/F is the central volume of distribution. MTT is the mean transit time. Q/F is the intercompartmental clearance. VCSF/F is the CSF volume of distribution. PC is the partition coefficient between central and CSF compartment. CrCl is the creatinine clearance. CrCl was included on parameter CL using a linear model ( = • (1 + • ( − 106))), and is the typical CL value of the population. RUV is the additive residual error on log scale. Cmax is the peak concentration. AUC is the area under the concentration-time curve. SIR is the sampling importance resampling. Population estimates in the table are given for a "typical" patient with free fat mass of 70 kg. The calculation of IIV (interindividual variability) and RSE (relative standard error), as well as secondary parameters refer to Table 1.

Secondary parameters
Day 14 Cmax (mg/L) 40.7 (4.9-106.9) Day 14 AUC (h×mg/L) 378.6 (73.3-1,349.2) F is the relative bioavailability. CL/F is the elimination clearance. V/F is the central volume of distribution. MTT is the mean transit time. AST is the aspartate aminotransferase. AST was included on parameter CL using an exponential model( = • •( −30) , and is the typical CL value of the population). RUV is the additive residual error on log scale. SIR is the sampling importance resampling. Population estimates in the table are given for a "typical" patient with free fat mass of 70 kg. Cmax is the peak concentration. AUC is the area under the concentration-time curve. The calculation of IIV and RSE, as well as secondary parameters refer to Table 1.  ( Figure S1. Graphical overview of the structural pharmacokinetic model for rifampin. CSF is the cerebrospinal fluid. F is the relative bioavailability. CL is the elimination clearance. Vc is the plasma volume of distribution. MTT is the mean transit time. Emax is the maximum increase in enzyme formation rate. kenz is the enzyme degradation rate. EC50 is the plasma concentration corresponding to 50% of Emax. Fmax is the maximum increase in relative bioavailability and ED50 is the dose corresponding to half of the Fmax. Q is the inter-compartmental clearance. VCSF is the CSF volume of distribution. PC is the partition coefficient between central and CSF compartment.       The full tree and the final tree are shown in A and B, respectively. The average predicted response (yval) is shown in each node. The response at the root node was assumed to be 1. A higher predicted response is associated with a higher risk of death. The numerators indicate number of deaths, and the denominators represent the total number of patients in each node. The percentage in each node is the fraction of patients in the node vs. total patients.