Population Pharmacokinetics of Artemether, Dihydroartemisinin, and Lumefantrine in Rwandese Pregnant Women Treated for Uncomplicated Plasmodium falciparum Malaria

Twenty-two Rwandese pregnant patients in their second (range for estimated gestational age [EGA], 15.7 to 27.6 weeks) or third trimester (range for EGA, 28.3 to 39.0 weeks) with a median EGA of 27.9 weeks were studied, and their blood plasma samples were analyzed for both LF and ARM-DHA. The treatment was well tolerated. The demographics of these patients are shown in Table 1.

A total of 363 blood plasma samples for LF and 387 blood plasma samples for ARM and DHA were analyzed. Nine samples out of these were excluded for LF, ARM, and DHA due to hemolysis. For LF less than 8% of the data were below the lower limit of quantification (BLLOQ), while for ARM and DHA 24% and 28%, respectively, were.

The final structural model for LF (Fig. 1) was a two-compartment disposition model with an absorption model consisting of 5 first-order transit compartments delivering the absorbed amount to the central compartment, with mean absorption time (MAT) of 4.04 h and terminal elimination half-life of ∼4 days (Table 2). This was consistent with previous studies reporting population PK of LF (15, 22).

A frequentist prior approach applied to fixed PK parameters of LF (i.e., clearance, volume of distribution, and absorption parameter) allowed previously reported PK information to support aspects that could not be characterized by the current study data alone, i.e., implementation of a peripheral compartment. The BLLOQ data were not included for LF in this analysis since the proportion was less than 10%. There was also no bias in BLLOQ data observed in the visual predictive checks (VPC) (Fig. 2) when these data were treated as missing or categorical data. Between-subject variability (BSV) was estimated for all structural parameters except for apparent clearance (CL/F), intercompartmental clearance (Q/F), and the volume of distribution of the peripheral compartment (V_p/F) of LF, which were estimated with poor precision (>50% residual standard error [RSE]). The final PK model structure for LF is depicted in Fig. 1A. The final PK parameter estimates are summarized in Table 2.

The addition of between-occasion variability (BOV) was significant for mean transit time (MTT) (change in objective function value [ΔOFV] = −15.1) and relative bioavailability (F) (ΔOFV = −124.2). The description of the outer data percentiles and hence model fit was improved when Box-Cox transformation of the distribution of the BSV for F was applied (ΔOFV = −120.85). Eta shrinkages for BSV computed for the final LF parameter estimates were 25.8% for V_p/F and 11.1% for MTT. Eta shrinkage for BOV ranged between 14.5% and 39.4% for MTT and 32.5% to 65.3% for F. Computed epsilon shrinkage (residual variability) for LF was 19.9%. The median area under the plasma concentration-time curve from 0 h to infinity (AUC_0–∞) for LF was 641 h · mg/liter (95% confidence interval [CI], 596 to 858 h · mg/liter). The predictive performance of the LF model is depicted in the VPC in Fig. 2.

The Laplacian estimation method was used to estimate model parameters for ARM-DHA. The final ARM-DHA model consisted of one-compartment disposition models for both ARM and DHA, with the assumption of irreversible in vivo conversion of ARM to DHA (Fig. 1B). An absorption model consisting of 2 transit compartments was superior to other explored absorption models.

From the observed data, the peak ARM concentration 2 h after the first dose was approximately 3 times (median 3.75) higher than that 2 h after the last dose, while this ratio was 0.87 for DHA. This suggested the possibility of a time-dependent decrease in ARM concentrations, i.e., enzyme induction of its own clearance, as previously described (17, 20, 23). The autoinduction effect (EFF) of the ARM concentration on the half-life of the susceptible enzymes, believed to represent CYP3A4, was modeled using a simple nonlinear relationship as described below (equation 1):where E_max represents the maximal autoinduction effect, EC₅₀ represents the ARM concentration when the autoinduction is half maximal, and CP is the ARM plasma concentration.

The estimated enzymatic half-life was 30.4 h, while the EC₅₀ was 9.37 nM. Typical clearance of ARM increased 43% at occasion six compared to occasion one.

Approaches to describe autoinduction of the clearance using alternative physiologically plausible models were also explored during the model building, such as (i) ARM being cleared via two elimination pathways, i.e., inducible and noninducible clearances (both estimated), where the inducible pathway produces DHA, and (ii) similar to approach i but with noninducible clearance producing DHA. These explorations did not improve model fit, as judged by the OFV change and visual diagnostics.

BSV was estimated for all structural parameters except for the E_max of the autoinduction effect of ARM, due to instability of the model. The addition of BOV on both MTT (ΔOFV = −12.6) and F (ΔOFV = −9.39) yielded significant improvements in the fit. The final model parameter estimates for both ARM and DHA are presented in Table 2, and the VPC of the ARM-DHA model is presented in Fig. 2.

During the model building, various body size descriptors (i.e., total body weight, ideal body weight [IBW], fat-free mass [FFM], and normal fat mass [NFM]) were implemented allometrically on structural model parameters of LF and ARM-DHA and explored. For both LF and ARM-DHA, none of the applied body size descriptors contributed to model improvement. Considering biological plausibility and previous population PK reports, total body weight, centered to median body weight, was implemented allometrically on clearances (raised to the power of 0.75) and volumes of distribution (raised to the power of 1.0).

Pregnancy has been reported to have significant impact on antimalarial drugs (15, 24). However, in this study, no pregnancy-related covariate effects were found for LF, ARM, or DHA. During the covariate analysis, observed parasitemia density on MTT, EGA on F (described linearly and with a spline), and dosing occasion (OCC) on F (described exponentially) for LF were all selected during the forward selection step (P ≤ 0.05; 1 degree of freedom); however, these were not maintained in the backward elimination step (P ≤ 0.01).

As for ARM-DHA, none of the covariates tested (time-varying parasitemia density and EGA) had any effect on the PK parameters except for baseline parasitemia density on MTT in the forward step, but again this parameter was not maintained during the backward elimination step (P ≤ 0.01). The final PK models for ARM-DHA and LF therefore incorporated only body weight implemented allometrically on clearances and volumes of distribution.

Efforts to simultaneously estimate the PK models for LF and ARM-DHA and explore the correlation between these two drugs' PK parameters, particularly as an effort to assess the correlation in the observed variability in the absorption parameters of both drugs, were not successful and not pursued further because (i) adding off-diagonal correlation elements to explore MTT correlations resulted in an unstable model, and (ii) implementation or correlations in F produced only very minimal and clinically nonsignificant improvements of fit (ΔOFV = −5.54). Moreover, simultaneous modeling substantially increased the computational time (>18-fold increase).

Since day 7 LF concentrations were not available for all patients (samples on exactly day 7 were available for 12/22 patients; samples between days 6 and 10 were available for 16/22 patients), model-based predictions were used to compare PK target attainment of this standard regimen and the alternatively proposed 5-day regimen for pregnant patients (13) (Fig. 3). The simulated median (range) day 7 LF plasma concentration after the standard dosing regimen was 709 (269 to 1,940) ng/ml (corresponding to 1,340 [509 to 3,670] nM), while for the alternative dosing regimen of 5 days this was 2,010 (769 to 4,580) ng/ml (corresponding to 3,801 [1,450 to 8,650] nM).

In terms of PK target attainment, for the standard dosing regimen 11.5% of patients achieved a day 7 concentration below the target concentration, while this was only 0.3% for the alternative dosing regimen.

This PK study was nested within a pharmacovigilance study reported elsewhere on the use of ACT in pregnancy during acute malaria (44). The study was conducted in the obstetrics and gynecology ward of Rwamagana district hospital in Rwanda. Rwamagana is in the eastern province of Rwanda, with mesoendemic malaria transmission intensity, and the study took place from June 2007 to July 2009. The study was approved by the Rwanda National Ethics Committee as study number IRB 00001497.

Briefly, eligible pregnant women older than 18 years with uncomplicated Plasmodium falciparum malaria, confirmed by light microscopy, were recruited once signed informed consent was obtained. Pregnancy was confirmed by a human chorionic gonadotropin pregnancy urine test, and gestational age was estimated by ultrasound. Enrolled patients were prescribed 4 tablets of a fixed oral combination of ARM and LF twice daily under supervision for 3 days (at 0 h [initial dose] and 8, 24, 36, 48, and 60 h) with a glass of milk and/or a small cake (a fatty meal) to enhance absorption. Novartis, Basel, Switzerland, provided the tablets, and each tablet contained 20 mg of ARM and 120 mg of LF.

Just before first drug administration (time < 0 h), blood was collected for biochemistry and the remaining plasma was used as the first, pre-drug administration (time < 0 h) PK sample. Thereafter, PK samples were collected through venipuncture. Blood samples were collected at 2 and 4 h after each dose and just before doses 2, 3, 4, 5, and 6 to measure trough concentrations. Additional samples were taken at 0.25 h after doses 1 and 2. Samples were collected at 6, 8, and 12 h after the last dose, in addition to a scheduled sample at day 7 after initiation of treatment. Patients were admitted until the last sample was taken and parasite clearance achieved. A total of 4 ml of blood per sampling time was collected in glass lithium-heparin vacuum tubes with gel. The sodium heparin tubes were at room temperature (18°C to 25°C) prior to use. Samples were centrifuged without delay and the plasma was separated and frozen at −70°C. Samples were shipped on dry ice to the Department of Clinical Pharmacology, Mahidol-Oxford Tropical Medicine Research Unit, Bangkok, Thailand, for drug measurements. The laboratory is accredited according to ISO15189 and ISO15190. Drug concentrations of ARM and DHA were measured in plasma using high-performance liquid chromatography coupled with tandem mass spectrometry (45); meanwhile, LF was measured using automated solid-phase extraction (46). The LLOQ were 1.43 ng/ml for both ARM (4.79 nM) and DHA (5.03 nM) and 24.86 ng/ml (47 nM) for LF. Samples that showed signs of extensive hemolysis were excluded from analysis. Quality control samples of LF, ARM, and DHA at three levels (low, middle, and high) were analyzed within each batch of clinical samples to ensure precision and accuracy during routine clinical drug measurements. The coefficients of variation for all analytes were lower than 5% for all quality control samples, which is well below the required precision of ±15% according to U.S. FDA regulatory guidelines (47, 48). Parasitemia density assessment was conducted by microscopy using Field's method, i.e., thin or thick blood film counts of asexual parasites and gametocytes every 8 h (±1 h) following the first dose administration until 72 h postdose (48, 49). After discharge from the hospital, samples for thick and thin blood films were collected on outpatient basis on days 7, 14, 21, 28, 35, 42, 49, and 56 or on any other day when clinically indicated.

Estimation and simulation were performed using nonlinear mixed-effects modeling in the software NONMEM 7.3 (ICON Development Solutions, Ellicott City, MD) (50). Postprocessing, diagnostics plots and automation were performed using Perl-speaks-NONMEM (PsN, 4.5.3) (51), Xpose (4.5.3) (52), Pirana (2.9.2) (53), and R (3.2.4) (54).

The molar units of LF, ARM, and DHA concentration were transformed to their natural logarithms for this modeling analysis. All the BLLOQ data for ARM and DHA were included and explored with the application of the likelihood-based M3 method for censored observations using the Laplacian estimation method (55, 56). Conversely, LF BLLOQ data were not included, as they made up <10% of the total data, and the FOCEI (first-order conditional estimation with interaction) estimation method was used. The unexplained residual error was estimated using an additive error model on the logarithmic scale for all drugs, which equates to an exponential error model on an arithmetic scale. In the case of ARM and DHA, a separate additive error model was used for each analyte. Different structural absorption (first-order, first-order with transit compartment, and sequential absorption) and distribution (one-, two-, and three-compartment) models were explored for all drugs.

Because of LF data sparseness, particularly during the elimination phase, informative priors based on a previous study (15) were applied to all parameter estimates (57). The chosen prior model explored PK properties of LF in pregnant and nonpregnant women with uncomplicated Plasmodium falciparum malaria in Uganda. The final model had pregnancy retained as a significant covariate on intercompartmental clearance. Hence, the frequentist prior estimation for LF's Q/F was recalculated to represent estimation for pregnant women in this study. The typical relative F was implemented as a fixed parameter for the parent analyte, i.e., LF and ARM (100% relative bioavailability). The stochastic model implemented consisted of BSV modeled as shown below (equation 2), BOV, and residual variability. Individual parameters for both drugs were modeled as lognormally distributed around the population estimate, except for F of LF. Box-Cox transformation (58) was explored for the distribution of BSV on F as shown below (equation 3) to assess formally the assumption that PK parameters are lognormally distributed.where P_i represents the individual parameter estimate, P_pop represents the typical parameter estimate for the population, η represents the BSV, and λ represents the estimated Box-Cox transformation factor.

ARM is known to exhibit an autoinduction of its own clearance (17). Enzyme kinetics was included in the ARM-DHA PK model, and an enzyme turnover model used previously by Hassan et al. and Smythe et al. was also adapted in this study (37, 59). ARM and DHA, expressed as molar concentrations, were characterized simultaneously assuming complete and irreversible in vivo conversion of ARM into DHA.

The dynamics of the enzyme compartment over time was expressed as shown below (equation 4); A_ENZ is the amount of enzyme in the enzyme compartment, K_ENZ is the first-order degradation rate constant of the enzyme, and EFF is the link between ARM concentration and its enzyme pool through increase in enzyme production rate. Linear and nonlinear relationships (E_max model) describing the effect of ARM concentrations on the induction of its own clearance were explored.

The enzyme concentration was initialized at 1 in order to normalize it to unity at baseline; i.e., the zero-order production rate of the enzyme was set to K_ENZ. This (enzyme) then modulates the preinduced ARM clearance (equation 5).

The model was parameterized in such way that the enzyme half-life (t_1/2ENZ) was estimated as shown below (equation 6).

The body size descriptor covariates (total body weight, IBW, FFM, and NFM), EGA, observed baseline parasitemia density, observed time-varying parasitemia density, temperature, and dosing occasion (OCC; i.e., each dose given was considered single dosing occasion) were considered for exploration of covariate analysis for LF, ARM, and DHA based on biological plausibility and previous findings.

For body size descriptors, different covariate implementations were explored: allometric scaling using total body weight, allometric scaling using IBW (60) (equation 7), allometric scaling using FFM (equation 8), and allometric scaling using NFM (equation 9), with Ffat representing the contribution of fat mass normalized to the FFM estimated for (i) CL/F, Q/F, volume of distribution of central compartment (V_c)/F, and V_p/F for LF and (ii) apparent clearance and volume of distribution of central compartment for ARM (CL_ARM/F and V₂/F) and DHA (CL_DHA/F and V₃/F) (61).where WHS_max is 37.99 kg/m² and WHS₅₀ is 35.98 kg/m², which represent the maximal and half-maximal weight-for-height standards, respectively.

All size descriptors were scaled to their respective medians (i.e., total body weight [59 kg], FFM_median, IBW_median, and NFM_median) on PK parameters using allometric power exponents of 0.75 for clearances (CL/F, Q/F, CL_ARM/F, and CL_DHA/F) and 1 for volumes of distribution (V_c/F, V_p/F, V_ARM/F, and V_DHA/F).

An example implementation for body descriptor FFM on CL/F is shown below (equation 10):

Later, all chosen covariates were explored with body weight maintained on clearances and distribution volumes using a standard allometric function when evaluating linear, exponential, and power relationships for the other covariates, which were normalized to their median values in the population.

Stepwise covariate modeling was applied for all continuous covariates using P values of 0.05 (ΔOFV > 3.84; 1 degree of freedom) in the forward step and 0.01 (ΔOFV > 6.63) in the backward step (62). EGA was additionally explored as a categorical covariate (trimester 2 versus 3) using a forward inclusion cutoff ΔOFV of >5.99 (2 degrees of freedom).

The final separate LF and ARM-DHA PK models were subsequently evaluated as a combined model in a simultaneous fit, to explore correlations between the PK parameters of both drugs, in particular bioavailability and absorption rate. Parameter correlations were explored using the variance-covariance matrix.

Monte Carlo simulations (n = 2,200) were performed using the final LF population PK model for up to 15 days after the first dose to assess and compare PK target attainment after a standard dosing regimen (80 mg of ARM/480 mg of LF twice daily for 3 days) and an extended alternative dosing regimen (80 mg of ARM/480 mg of LF twice daily for 5 days) in this particular population (13). The predicted median and 5th and 95th percentiles for LF concentration were extracted from their simulated distributions to assess descriptively the day 7 concentration target attainment (280 ng/ml).

Model selection was guided by physiological plausibility, plausible parameter estimates, precision of parameters, visual diagnostics, and OFV, computed by NONMEM as minus twice the log likelihood where a drop in OFV of 3.84 or more was considered a significant (P = 0.05) improvement between two hierarchical models after inclusion of one additional parameter (1 degree of freedom). VPCs were performed (1,000 simulations) to evaluate the predictive performance of developed models. The 5th, 50th, and 95th percentiles of obtained data for the respective PK model were overlaid with the model-predicted 95% confidence intervals of the same percentiles. Diagnostic plots were used to evaluate the overall goodness of fit by plotting log-transformed observed drug concentrations against the population fitted and the individually fitted log-transformed concentrations. The reliability of individual parameter estimates and goodness-of-fit plots were also assessed through eta and epsilon shrinkages. Estimates of parameter imprecision were obtained from a recently developed procedure, sampling importance resampling (SIR) (63).