Clinical Pharmacokinetics (2021) 60:943–953

https://doi.org/10.1007/s40262-020-00971-2

ORIGINAL RESEARCH ARTICLE

A Model‑Informed Method for the Purpose of Precision Dosing

of Isoniazid in Pulmonary Tuberculosis
Stijn W. van Beek1   · Rob ter Heine1 · Jan‑Willem C. Alffenaar2,3,4 · Cecile Magis‑Escurra5 · Rob E. Aarnoutse1 ·
Elin M. Svensson1,6 on behalf of the Isoniazid Precision Dosing Group

Accepted: 28 November 2020 / Published online: 22 February 2021

© The Author(s) 2021

Abstract
Background and Objective  This study aimed to develop and evaluate a population pharmacokinetic model and limited
sampling strategy for isoniazid to be used in model-based therapeutic drug monitoring.
Methods  A population pharmacokinetic model was developed based on isoniazid and acetyl-isoniazid pharmacokinetic data
from seven studies with in total 466 patients from three continents. Three limited sampling strategies were tested based on
the available sampling times in the dataset and practical considerations. The tested limited sampling strategies sampled at
2, 4, and 6 h, 2 and 4 h, and 2 h after dosing. The model-predicted area under the concentration–time curve from 0 to 24 h
(AUC​24) and the peak concentration from the limited sampling strategies were compared to predictions using the full phar-
macokinetic curve. Bias and precision were assessed using the mean error (ME) and the root mean square error (RMSE),
both expressed as a percentage of the mean model-predicted AUC​24 or peak concentration on the full pharmacokinetic curve.
Results  Performance of the developed model was acceptable and the uncertainty in parameter estimations was generally
low (the highest relative standard error was 39% coefficient of variation). The limited sampling strategy with sampling at 2
and 4 h was determined as most suitable with an ME of 1.1% and RMSE of 23.4% for AUC​24 prediction, and ME of 2.7%
and RMSE of 23.8% for peak concentration prediction. For the performance of this strategy, it is important that data on both
isoniazid and acetyl-isoniazid are used. If only data on isoniazid are available, a limited sampling strategy using 2, 4, and
6 h can be employed with an ME of 1.7% and RMSE of 20.9% for AUC​24 prediction, and ME of 1.2% and RMSE of 23.8%
for peak concentration prediction.
Conclusions  A model-based therapeutic drug monitoring strategy for personalized dosing of isoniazid using sampling at
2 and 4 h after dosing was successfully developed. Prospective evaluation of this strategy will show how it performs in a
clinical therapeutic drug monitoring setting.

* Stijn W. van Beek

Stijn.vanBeek@radboudumc.nl Key Points 
1
Department of Pharmacy, Radboud Institute for Health Model-informed precision dosing may aid in optimizing
Sciences, Radboud University Medical Center, Geert
Grooteplein zuid 10, 864, 6500 HB, Nijmegen, tuberculosis treatment
The Netherlands
We developed a population pharmacokinetic model for
2
School of Pharmacy, Faculty of Medicine and Health, isoniazid and its metabolite based on the largest and
University of Sydney, Sydney, NSW, Australia
most diverse dataset thus far and thoroughly validated its
3
Westmead Hospital, Sydney, NSW, Australia suitability for purposes of dose individualization
4
Marie Bashir Institute of Infectious Diseases and Biosecurity,
University of Sydney, Sydney, NSW, Australia
We found that sampling at 2 and 4 h with data on both
5
isoniazid and acetyl-isoniazid, or sampling at 2, 4, and
Department of Respiratory Diseases, Radboud University
Medical Center, Nijmegen, The Netherlands
6 h without the metabolite could be used to predict expo-
6
sure (area under the concentration–time curve)
Department of Pharmacy, Uppsala University, Uppsala,
Sweden

Vol.:(0123456789)

944 S. W. van Beek et al.

1 Background develop and evaluate a model-based TDM approach for INH

in adult patients with pulmonary TB using a population-PK
Although the disease burden of tuberculosis (TB) is falling, model suitable for dose individualization in programmatic
it remains one of the top ten causes of death globally and treatment.
is the leading cause of death from a single infectious agent
[1]. The average global treatment success rate for drug-
susceptible TB was estimated at 85% in 2017 [1]. Despite
this high success rate, it still means that one out of every 2 Methods
seven patients treated for TB has an unfavorable treatment
outcome. This shows that a significant part of the patients 2.1 Data
does not respond adequately to treatment. A possible reason
for not responding to treatment is suboptimal anti-TB drug The data originated from seven previously conducted stud-
exposure [2]. ies, including data from both clinical trials and TDM [14,
Isoniazid (INH) is one of the pillars on which treatment 23–28]. All studies included patients with pulmonary TB.
of drug-susceptible TB is based and is listed as an essential The study data are presented in Table 1. In total, data of
medicine by the World Health Organization [3]. Low INH 466 patients from three continents were used to develop
concentrations compared to the median in a population have the PK model. These PK data consisted of 2546 INH and
been reported to be common and are often a result of high 2236 acetyl-INH concentrations collected during an interval
inter-individual variability (IIV) in the pharmacokinetics range of 0–24 h. Of these concentrations, 6.0% of the INH
of this drug [4–7]. A large part of this variability can be and 10.9% of the acetyl-INH observations were measured
attributed to the polymorphic N-acetyltransferase 2 (NAT2) as ‘below the limit of quantification’ (BLQ), which ranged
enzyme that catalyzes the acetylation of INH into its acetyl- from 0 to 27.3% depending on the study.
INH metabolite [8]. Four studies involved intensive PK sampling with at least
Large inter-individual variability in pharmacokinet- nine samples taken after a dose of INH [14, 24, 26, 27], two
ics is one of the motives and prerequisites for therapeutic studies involved limited sampling with two to three sam-
drug monitoring (TDM) [2]. Therapeutic drug monitoring ples taken [23, 28], and one study involved limited sampling
is an approach to personalize drug dosing by measuring a with one sample taken for all except nine individuals who
patient’s drug exposure and comparing it to a drug-specific were sampled intensively [25]. Five studies only included
target level and adjusting the dose if needed. The exposure measurements of a single occasion [14, 24–27], one study
is approximated by sampling at drug-specific planned times. included some measurements of second occasions [28], and
This strategy is called a limited sampling strategy (LSS). one included measurements for second occasions for all indi-
The area under the concentration–time curve over a dosing viduals [23]. One study included data that were collected
interval (0–24 h, AUC​24) in relation to a minimal inhibitory during TDM practice [28]. Five studies included PK data
concentration has been proposed to be the most important on both INH and acetyl-INH [24–28], while the other two
predictor of treatment outcome for the first-line TB drugs studies included PK data on INH [14, 23].
[9–12]. However, many different pharmacokinetic (PK) tar- The analysis of INH and acetyl-INH concentrations was
get suggestions exist [2, 9, 13–16]. Some of these have a performed in The Netherlands for all studies, using either
peak concentration (Cmax) target while others use AUC​24. ultra-performance liquid chromatography with ultraviolet
Lacking anything more precise, a population median AUC​ detection or liquid chromatography–tandem mass spec-
24 value of INH corrected for acetylator status could serve
trometry [29]. Both methods were extensively validated
as a PK target as it is known that the standard dose is gener- in accordance with European Medicines Agency and US
ally effective [2]. Food and Drug Administration guidelines. Accuracy was
We propose the use of a model-based TDM method using between 99.4 and 108.8% and between 95.6 and 111.3%
individual predictions from a population PK model for INH. and the intra- and inter-assay coefficients of variation
Using a model-based method for TDM has several advan- (CVs) were less than 12.6% and 10.5% for the ultra-per-
tages: it can use flexible sampling times, it allows exposure formance liquid chromatography with ultraviolet detection
predictions following dose adjustments, and it can use data and liquid chromatography–tandem mass spectrometry
from previous sampling occasions for the same patient in methods, respectively, dependent on the concentration.
the prediction. Available models for INH are often based The lower limits of quantification for INH and acetyl-INH
on data from a single-center study with a limited sample for the ultra-performance liquid chromatography method
size limiting the generalizability of their use for dose indi- were 0.0253 mg/L and 0.162 mg/L, and these were 0.0452
vidualization [4, 5, 17–22]. The aim of this study was to mg/L and 0.0450 mg/L for the liquid chromatography-
tandem mass spectrometry method. Performance of the
A Model-Informed Method for Dosing of Isoniazid 945

Table 1  Overview of the included study data

Semvua [23] Tostmann [24] Burhan [25] Aarnoutse [26] Magis-Escurra Boeree [27] Van Beek [28] Total
[14]

Included 23 40 180 62 14 97 52 466

patients
Observed data points
 Total 92 619 534 1189 154 1804 390 4782
 INH 92 309 263 619 154 898 211 2546
 Acetyl-INH 0 310 271 570 0 906 179 2236
Data points BLQ, n (%)
 Total 2 (2.2) 54 (8.7) 1 (0.2) 40 (3.4) 42 (27.3) 249 (13.8) 7 (1.8) 395 (8.3)
 INH 2 (2.2) 27 (8.7) 0 (0) 5 (0.8) 42 (27.3) 72 (8.0) 4 (1.9) 152 (6.0)
 Acetyl-INH – 27 (8.7) 1 (0.4) 35 (6.1) – 177 (19.5) 3 (1.7) 243 (10.9)
Typical sam- 2 and 6 h Full PK, 9 2 h, full PK Full PK, 11 Full PK, 11 Full PK, 9 2, 4, and 6 h –
pling hours samples for 9 patients samples samples samples
0–24 h 0–24 h 0–24 h 0–24 h
Samples, 4 (4–4) 8 (4–9) 1 (1–13) 10 (7–11) 11 (10–11) 9 (5–10) 4 (1–8) –
median
(range)
Number of 2 1 1 1 1 1 1 (2 for 14 –
occasions patients)
Bioanalysis UPLC-UV UPLC-UV UPLC-UV UPLC-UV LC-MS/MS UPLC-UV LC-MS/MS –
method
Country of Tanzania Tanzania Indonesia Tanzania The Tanzania and The –
origin ­Netherlandsa South Africa ­Netherlandsa
Setting Clinical trial Clinical trial Clinical trial Clinical trial TDM Clinical trial TDM –

BLQ below limit of quantification, h hour, INH isoniazid, LC-MS or liquid chromatography-mass spectrometry, PK pharmacokinetics, TDM
therapeutic drug monitoring, UPLC ultra-performance liquid chromatography
a
 While the patients from these studies were treated in The Netherlands, the TB population in The Netherlands is very heterogenous with many
different ethnical backgrounds

assays was externally evaluated by participation in an the acetyl-INH data were added to the model. One- and two-
international proficiency testing program [30]. compartment disposition models were explored to describe
the pharmacokinetics of INH and acetyl-INH. A well-stirred
2.2 Software liver model was tested instead of first-order elimination to
describe the first-pass effect of INH. The molecular weight
R version 3.4.3 was used for data management, statistics, of INH and acetyl-INH was used to transform drug con-
and plotting [31]. Model development was performed using centrations to molar concentrations to model the formation
the nonlinear mixed-effects modeling program, NONMEM of acetyl-INH from INH. The bioavailability of INH was
version 7.4 with Pirana as an interface [32, 33]. The first- assumed to be 100% and all estimated PK parameters should
order conditional estimation method with interaction was be interpreted as apparent oral PK parameters. All volume
used for estimation. PsN version 4.7 was used as an aid for and clearance parameters were allometrically scaled with
advanced functionalities [33]. The Xpose4 R package ver- total body weight using an exponent of 1 or 0.75, respec-
sion 4.6.1 was used for graphical visualization of the visual tively [35]. Allometric scaling based on fat-free mass was
predictive checks (VPCs) [33]. not considered because height is not always readily avail-
able in TDM practice. NAT2 genotypic data were not deter-
2.3 Pharmacokinetic Model Development mined in the included studies. A mixture model to distin-
guish between different groups of NAT2 metabolizers in
The PK model was developed with a stepwise approach [34]. the absence of genotypic data was applied [36]. It is known
The INH concentration data were included one study at a that the distribution of NAT2 metabolic function is trimodal
time, starting with the studies that had the most informative [8]. The mixture model was tested with two as well as three
and dense sampling. After finishing the base INH model,

946 S. W. van Beek et al.

metabolizer groups as it can be difficult to distinguish the at 2 and 6 h was not included as it does not offer any benefit
three mixture groups as is seen in other models [4, 17]. In over sampling at 2 and 4 h, which is more feasible to perform
case two, metabolizer groups were used, the fast and inter- in clinical practice. For each LSS, the performance of AUC​24
mediate metabolizers were grouped, as metabolization rates and Cmax prediction was determined using the patients who
of these groups are closer together than those of the slow had a full PK curve including samples at 2, 4, and 6 h after
metabolizers, which have more deviant metabolization rates. dosing (169 patients from five studies) [14, 24–27]. The per-
Inter-individual variability in the pharmacokinetics of formance of the LSSs was also compared to the performance
INH was assumed to be log-normally distributed. Inter- of the model without any sampling to assess the added value
occasion variability was not assessed because the number of sampling in TDM. The most suitable LSS was selected
of individuals who were sampled on multiple occasions was based on predictive performance and clinical practicality.
limited. For residual variability, additive, proportional, and The performance of predicting the AUC​24 was regarded as
combined models were tested for both INH and acetyl-INH. more important than that of the Cmax. In these assessments,
Below the limit of quantification data were handled using the mean error (ME) and root mean square error (RMSE)
the M6 method as described by Beal [37]. The M3 method and their 95% CIs of the prediction of AUC​24 and Cmax were
was tested but did not perform better than the M6 method calculated as measures of bias and precision [43, 44]. The
and increased numerical instability. Using this method, in ME and RMSE were expressed as a percentage of the mean
the elimination phase, the first datapoint BLQ is replaced of the corresponding model-estimated parameter using the
by the lower limit of quantification/2, and subsequent BLQ full PK curve.
points are ignored. In the absorption phase, it is the last The most suitable LSS was used to perform a fit-for-pur-
datapoint BLQ that is replaced and all previous BLQ points pose analysis, meaning that it was investigated how well the
that are ignored. The minimum additive error for data BLQ pharmacokinetics of INH could be predicted from a previ-
was fixed to 50% of the lower limit of quantification values. ous measurement [45]. The patients including data of a sec-
The number of potential covariates tested was kept low ond occasion and the sampling times for the selected LSS
because of the purpose of our model; the covariates available were used for this. Apart from the most suitable LSS, ‘no
in TDM practice are generally quite limited. It was opted to sampling’ was tested as well to evaluate the amount of vari-
include only the allometric scaling with total body weight to ability explained by the most suitable LSS. Suitable data for
not hamper a general implementation of the model in routine this analysis were available from 13 patients from only one
TDM. Although we did test the impact of formulation dif- study [28]. The performance of predicting the AUC​24 for the
ferences, but no effect could be identified. second occasion was determined using the ME and RMSE.
Pharmacokinetic parameter uncertainty was assessed
using sampling importance resampling [38, 39]. Sampling 2.5 Importance of Acetyl‑INH Data and Mixture
importance resampling is a fast method to determine param- Component for the Predictive Performance
eter uncertainty which is free of distributional assumptions,
a good alternative to the bootstrap. The initial estimates The added value of both the acetyl-INH metabolite data and
for the SIR were based on a successful covariance step and inclusion of a mixture model accounting for the polymorphic
inflated by factor 1.5. The samples per iteration were 1000, NAT2 clearance was assessed for the most suitable LSS.
1000, 1000, 2000, and 2000, the re-samples per iteration The selected LSS was used to provide AUC​24 predictions
were 200, 400, 500, 1000, and 1000. Parameter precision using only the INH data. These predictions were compared
was presented using the calculated 95% confidence intervals to model predictions using the full dataset, including the
(CIs). The predictive performance of the final model was acetyl-INH data. To determine the added benefit of a mixture
assessed using VPCs and goodness-of-fit plots. The VPCs model to the performance of predicting the AUC​24, the mix-
were prediction corrected and the PsN mixture option was ture component was removed, and model parameters were
used [40, 41]. re-estimated. The performance was again compared using
the ME and RMSE.
2.4 Comparison of Limited Sampling Strategies

The performance of three LSSs was tested. These LSSs had 3 Results
sampling times at 2 h, 2 and 4 h, and 2, 4, and 6 h after dos-
ing. These LSSs were chosen considering that sampling at 3.1 Pharmacokinetic Model
2 and 6 h, introduced by Peloquin et al., is common for TB
drugs, and we have added a 4-h sampling point in our TDM The final model is depicted in Fig. 1. It included four transit
service [2, 42]. In addition, samples with these sampling absorption compartments going into a well-stirred liver model
times were available within most of the datasets. Sampling [46]. The pharmacokinetics of INH was best described by
A Model-Informed Method for Dosing of Isoniazid 947

a two-compartment disposition model. Acetyl-INH pharma- from this figure. The model does not describe the median
cokinetics was characterized using a single compartment and observed peak perfectly. Furthermore, the model simulates
first-order elimination. Metabolization in the liver compart- concentrations that are slightly too high on the upper end of
ment could either clear INH or transform it into acetyl-INH. the concentration range for acetyl-INH at timepoints after
It was not possible to estimate INH clearance as a trimodal 8 h. The VPCs of the model were deemed satisfactory for
distribution using a mixture model and thus a bimodal mix- TDM using the AUC​24. The performance of the model for
ture model was included. The volume of the liver compart- TDM using the Cmax should be explored further. Other good-
ment and hepatic plasma flow were fixed to 1 L and 49.5 L/h, ness-of-fit plots are shown in Fig. 3 and do not show any
respectively, and allometrically scaled on total body weight obvious misspecifications. The full code of the final model is
like other volume and clearance parameters. included in the Electronic Supplementary Material (ESM).
The final model parameter estimates and their uncertainty
are shown in Table 2. Uncertainty in the parameter estimates 3.2 Comparison of Limited Sampling Strategies
was generally low (relative standard error, RSE < 12%CV),
except for the IIV of the central volume of acetyl-INH (RSE = The AUC​24 and Cmax predicted by the LSSs were compared
40%CV) and the correlation between the IIV of the clearances to the AUC​24 and Cmax predicted by the model based on full
(RSE = 71%CV). The absorption processes were variable PK curves. Scatterplots of this are shown in ESM 1. The bias
between individuals with an IIV estimated at 83.2%CV (95% and precision of these strategies are shown in Table 3. The
CI 78.1–91.7). The proportion of fast NAT2 acetylators in LSS using only one sample performed substantially worse
the studies was estimated at 43.4% (95% CI 38.4–48.9). Slow in both bias and precision than the strategies with two or
acetylators typically had 13% of the intrinsic NAT2 clearance three sampling points. The strategies employing sampling
of fast acetylators. The second clearance pathway, not through at 2 and 4, and 2, 4, and 6 hours after dosing have a similar
NAT2, typically made up 74% of the total clearance for slow performance. Based on bias, precision, clinical practical-
acetylators and 27% of the clearance for fast acetylators. A ity, and conformity with a previously developed strategy for
correlation was found between the clearance through acetyla- rifampicin [28], the strategy using sampling at 2 and 4 hours
tion and through other pathways, but the uncertainty of this after dosing was selected as the most suitable.
correlation was quite high (0.143, 95% CI − 0.0620, 0.325). To assess if the model and LSS is fit for purpose, the
Additionally, a correlation between the proportional residual AUC​24 of a second sampling occasion was predicted using
errors of INH and acetyl-INH was found. data from a first sampling occasion. The fit-for-purpose
The VPCs, shown in Fig. 2, generally describe the model analysis of the model using the 2- and 4-h sampling strategy
well. Some limitations of the model can be discerned resulted in an ME of 25.7% and an RMSE of 37.2%. Without

Fig. 1  Schematic overview of the final isoniazid (INH) population two-compartment disposition model and AcINH pharmacokinetic
pharmacokinetic model. The dose enters the well-stirred liver com- by a one-compartment disposition model. CL clearance into acetyl-
partment via a four-compartmental transit model. Absorption con- isoniazid, CL other clearance into other metabolites than acetyl-iso-
stant (KA) is defined as the number of transit compartments plus 1, niazid, CLM clearance of the acetyl-isoniazid metabolite, EH hepatic
divided by the mean transit time (MTT). From the liver compartment, extraction ratio, Q inter-compartmental clearance, QH hepatic plasma
the drug can be distributed to the central INH compartment or be flow, V central volume of isoniazid, VH hepatic volume, VM central
metabolized into acetyl-isoniazid (AcINH) or other metabolites. From volume of the acetyl-isoniazid metabolite, VP peripheral volume of
the central AcINH compartment, the drug is cleared through first- isoniazid
order elimination. Isoniazid pharmacokinetics are described using a

948 S. W. van Beek et al.

Table 2  Final model parameter Parameter Estimate 95% CI RSE (%CV)

estimates
Isoniazid
Central volume, V (L) 57.5 54.9–59.4 2.2
Absorption constant, KA (/h) 5.42 5.40–5.44 0.1
Proportion fast acetylators 0.434 0.384–0.489 6.6
Clearance (L/h)
 Fast acetylators, CLF 32.7 28.6–37.6 7.7
 Slow acetylators, CLS 4.31 3.85–4.75 7.4
 Other clearance pathways, CLO 12.1 11.2–13.1 4.3
Peripheral volume, VP (L) 18.7 16.8–20.5 4.2
Inter-compartmental clearance, Q (L/h) 2.48 2.10–2.87 8.2
IIV V (%CV) 26.4 23.4–29.3 6.5
IIV KA (%CV) 83.2 78.1–91.7 4.9
IIV CLF/CLS (%CV)a 57.5 51.9–63.3 5.0
IIV CLO (%CV) 37.9 34.4–41.0 5.1
Proportional error (%) 37.5 36.4–38.7 3.0
Acetyl-isoniazid
Central volume, VM (L) 39.2 36.1–42.2 4.3
Clearance, CLM (L/h) 6.65 6.10–7.22 4.5
IIV VM (%CV) 10.3 5.98–14.2 37
IIV CLM (%CV) 36.7 34.2–40.4 5.3
Proportional error (%) 23.3 22.8–24.0 2.8
Correlations
Acetylation-other clearance pathways 0.143 −0.0620; 0.325 39
Parent error-metabolite error 0.601 0.556-0.656 2.9

CI confidence interval, CV coefficient of variation, h hour, IIV inter-individual variability, L liter, RSE rela-
tive standard error
a
 The clearances for fast and slow acetylators shared one estimated IIV

sampling in the first occasion, the ME was 82.7% and the The re-estimated model without a mixture component
RMSE 117.9%. using the 2- and 4-hour LSS had an ME of − 1.1% and
RMSE of 24.6% for AUC​24 estimation, and an ME of 7.7%
and RMSE of 23.3% for estimation of the Cmax, meaning that
3.3 Importance of Acetyl‑INH Data and Mixture the mixture component was not essential for the predictive
Component for the Predictive Performance performance of the model. The estimated PK parameters
for the model without a mixture component were similar
The final model using the 2- and 4-h LSS without the acetyl- to those of the final model, except for the clearance repre-
INH data had an ME of − 1.8% and an RMSE of 28.9% senting both metabolization groups and their IIV. The new
for estimating the AUC​24. For estimation of the Cmax, the clearance has a value of 9.12 L/h and lies in between the
ME was 0.1% and the RMSE 23.5%. Comparison to data clearances of the slow and fast metabolizers as estimated
in Table 3 shows that the prediction of the model without by the final model. The estimated IIV of this clearance has
acetyl-INH data is less precise compared to the predictions increased drastically by more than two-fold to 115.8%CV.
including acetyl-INH for the 2- and 4-h sampling strategy. ESM 2b shows the effect of removing the mixture model on
For the 2-, 4-, and 6-h strategy without acetyl-INH data, the the performance of predicting the AUC​24.
estimation of the AUC​24 had an ME of 1.7% and an RMSE
of 20.9%. For estimation of the Cmax, the ME was 1.2% and
the RMSE 23.8%. This is comparable to the performances 4 Discussion
of the 2-, 4-, and 6-h and 2- and 4-h strategies including
acetyl-INH data. ESM 2a shows the effect of removing the In this project, we have developed a model describing the
acetyl-INH data on the performance of predicting the AUC​ pharmacokinetics of INH based on a large and diverse data-
24 for the 2- and 4-h sampling strategy. set. The model should have a wider applicability and be
A Model-Informed Method for Dosing of Isoniazid 949

Fig. 2  Basic goodness-of-fit plots for the final isoniazid population The solid lines are the y = 0 lines, and the dashed lines define the
pharmacokinetic model. a, b The observed concentrations plotted range between which you want the majority of the observations to lie.
against the model-predicted concentrations or the individual model- d The normalised prediction distribution errors plotted against the
predicted concentrations. The black lines represent the line of unity, time after dose. The solid lines are the y = 0 lines and the blue lines
and the blue lines represent the local polynomial regression fit. c The are the means per time bin
conditionally weighted residuals plotted against the time after dose.

better suited for TDM in various populations compared to that the model performs well for the purpose of TDM. As
previously published models, which were all based on stud- such, we do not see the instability as a sign of an underlying
ies from a single country [4, 5, 17–22]. We also introduced problem. We opted to have a limited inclusion of covariates
and evaluated a model-based TDM approach for personal- in view of the purpose of this model. For other purposes,
ized dosing of INH using sampling at 2 and 4 h or 2, 4, the model could probably be improved by the inclusion of
and 6 h after dosing. These approaches will allow for dose additional covariates. While inter-occasion variability is
adjustments of INH in programmatic TB treatment. known to be present for INH, we were unable to identify it in
The structure of the final PK model is comparable to our model [4, 17]. It has been described that inter-occasion
those previously described [5, 17, 21]. The estimated clear- variability may impact the predictive performance of model-
ances and proportion of fast NAT2 acetylators described in based dosing algorithms [48]. Despite this, we showed in the
this model are very similar to those previously described fit-for-purpose evaluation using multiple occasions that the
[17]. As in other published INH models, our model could model performed well.
only separate two of the three acetylator subgroups [17, 18]. The LSS with sampling at 2 and 4 h after dosing was
We did not differentiate between acetylator proportions for selected as the most suitable strategy. Previously, we
different ethnicities, which are known to vary [47]. The other introduced a model-based TDM approach for rifampicin,
model parameters are also mostly similar to those described just like INH a pillar within the TB treatment [28]. This
previously [5, 17]. The estimates for the peripheral volume approach also uses a LSS with sampling at 2 and 4 h after
and intercompartmental clearance for INH vary between the dosing, which means that the method described here and for
different models. During the model building process, we rifampicin are compatible in clinical practice. In this study,
encountered instability problems, making it difficult to esti- we decided to develop a new PK model rather than evaluate
mate some of the model parameters. Explanations for this existing models like we did for the model-based approach
could be the large number of different data sources included for rifampicin. This decision was based on the added value
in the model building dataset and the model complexity. The of a model built on a large and diverse dataset. A better LSS
VPCs of the model are acceptable and it has been shown could potentially have been found using an optimal design

950 S. W. van Beek et al.

Fig. 3  Visual predictive checks of the final isoniazid population phar- [36]. The solid lines represent the mean of the observed concentra-
macokinetic model for a slow metabolizers and b fast metabolizers. tions and the dashed lines represent the 2.5th and 97.5th percentiles.
The visual predictive checks are based on 1000 simulations and pre- The shaded areas represent the 95% confidence interval of the 2.5th,
diction corrected. The PsN mixture option was used and the plots are 50th, and 97.5th percentiles of the simulated concentrations
based on the individual probability for belonging to a subpopulation

Table 3  Predictive performance comparison of the limited sampling strategies

Data used for prediction Sampling strategy AUC​0–24 Cmax

RMSE, % (SD) ME, % (95% CI) RMSE, % (SD) ME, % (95% CI)

No data No sampling 73.7 (4.0) 55.8 (48.5, 63.1) 44.1 (2.4) 25.3 (19.9, 30.8)
INH + AcINH 2, 4, and 6 h 21.8 (1.2) 1.6 (−1.7, 4.8) 24.2 (1.3) 4.2 (0.6, 7.8)
2 and 4 h 23.4 (1.3) −1.1 (−4.6, 2.4) 23.8 (1.3) 2.7 (−0.9, 6.3)
2h 34.0 (1.8) 10.0 (5.1, 14.9) 26.5 (1.4) 6.0 (2.1, 9.9)
INH 2, 4, and 6 h 20.9 (1.1) 1.7 (−1.5, 4.8) 23.8 (1.3) 1.2 (−2.4, 4.8)
2 and 4 h 28.9 (1.6) −1.8 (−6.2, 2.6) 23.5 (1.3) 0.1 (−3.5, 3.7)
2h 59.1 (3.2) 36.1 (29.1, 43.2) 27.0 (1.5) 11.7 (8.0, 15.4)

Both the ME and RMSE are depicted as a percentage of the mean of the corresponding model-estimated parameter using the full pharmacoki-
netic curve
AcINH acetyl-isoniazid, AUC​0–24 area under the concentration–time curve from 0 to 24 h, CI confidence interval, Cmax peak concentration, INH
isoniazid, ME mean error, RMSE root mean square error, SD standard deviation
A Model-Informed Method for Dosing of Isoniazid 951

experiment. However, by sticking to predefined sampling 5 Conclusions

times, we were able to evaluate the proposed strategy using
the existing time points in the dataset. We developed a model-based LSS using INH and acetyl-
For unbiased predictions using the 2- and 4-h LSS, INH data from sampling at 2 and 4 h after dosing to be used
acetyl-INH concentration data were needed. If using this for individualized dosing in TDM practice. Alternatively, a
LSS without acetyl-INH PK data input, the performance of 2-, 4-, and 6-h LSS can be used if only collecting PK data on
predicting the AUC​24 will be lower. However, if acetyl-INH INH and not on acetyl-INH. Prospective evaluation of this
data are not available, it can be compensated for by adding strategy will show how it performs in a clinical TDM setting.
a 6-h sampling time. Removing the mixture component and
re-estimating the model parameters shows that not account- Supplementary Information  The online version contains supplemen-
ing for the polymorphic clearance of INH does not have a tary material available at https:​ //doi.org/10.1007/s40262​ -020-00971-​ 2.
major impact on performance of the exposure prediction.
After removing the mixture component, the variability Declarations 
caused by the polymorphic NAT2 clearance is described
Funding  No funding was received specifically for this study or for the
by an IIV more than two times the size before removing the preparation of this article. Elin M. Svensson is supported by PanACEA,
mixture component. This increase in variability probably which is part of the European and Developing Countries Clinical Trials
compensates for the lack of mixture model and prevents a Partnership (EDCTP) 2 programme supported by the European Union
significant drop in performance of the LSS. (Grant number TRIA2015-1102-PanACEA).
While we present a sampling strategy using 2- and 4-h
sampling times, deviation from these sampling times does Conflict of interest All authors declare no conflicts of interest that
could have influenced the submitted work.
not mean exposure prediction is not trustworthy anymore.
The presented strategy should be seen as a flexible sam- Ethical approval  The data used in this study was collected in line with
pling strategy. Using sampling times that deviate from the the principles of the Declaration of Helsinki. Approval was granted by
LSS is one of the benefits of using a model-based approach. institutional review boards and independent ethics committees for each
study of which data was used in this work.
However, we did not evaluate the impact of deviating from
the proposed sampling times as such sampling time devia- Consent to participate  Not applicable.
tions were not sufficiently present in the data because of the
regulated nature of the included studies. It would have been Consent for publication  Not applicable.
possible to simulate deviating sampling times as input for Availability of data and material  Not applicable.
the sampling strategies, but this was beyond the purpose of
this study. Code availability  The model code is available within the supplemental
We tested if the LSS was fit for purpose by assessing the materials.
performance of predicting the exposure of a future sampling Author contributions  SvB, RtH and ES performed the analysis, data
occasion. We showed that the strategy is able to explain most interpretation and writing of the manuscript. JA, CME and RE were
of the variability in the pharmacokinetics by comparing its involved in data collection, data interpretation and writing of the manu-
performance to that without sampling. The explained vari- script. MB, EB, RD, AD, SG, CM, NN, NH, WH, MH, GK, KR, IS,
HS and AT were involved in data collection and contributed to the
ability is determined by IIV as inter-occasion variability was writing of the manuscript.
not included in the model. By incorporating inter-occasion
variability in the model, the performance of predicting a
Open Access  This article is licensed under a Creative Commons Attri-
future occasion could be further improved [48]. Prospective bution-NonCommercial 4.0 International License, which permits any
evaluation of this method is needed to show how well it will non-commercial use, sharing, adaptation, distribution and reproduction
perform in a real TDM setting. in any medium or format, as long as you give appropriate credit to the
The model-based TDM approach has advantages, but is original author(s) and the source, provide a link to the Creative Com-
mons licence, and indicate if changes were made. The images or other
also a complex methodology in terms of software usage and third party material in this article are included in the article’s Creative
underlying theory [49]. For this reason, implementation of Commons licence, unless indicated otherwise in a credit line to the
a model-based TDM approach in clinical practice should material. If material is not included in the article’s Creative Commons
be combined with a user-friendly interface to improve the licence and your intended use is not permitted by statutory regula-
tion or exceeds the permitted use, you will need to obtain permission
ease of use. Furthermore, the translation from the model- directly from the copyright holder. To view a copy of this licence, visit
based results to clinical advice is crucial for successful http://creat​iveco​mmons​.org/licen​ses/by-nc/4.0/.
implementation.

952 S. W. van Beek et al.

References 19. Lalande L, Bourguignon L, Bihari S, et al. Population modeling

and simulation study of the pharmacokinetics and antituberculo-
sis pharmacodynamics of isoniazid in lungs. Antimicrob Agents
1. World Health Organization. Global tuberculosis report 2019.
Chemother. 2015;59(9):5181–9.
http://www.who.int/tb/publi​catio​ns/globa​l_repor​t/en/. Accessed
20. Peloquin CA, Jaresko GS, Yong AC, et al. Population pharma-
7 Dec 2020.
cokinetic modeling of isoniazid, rifampin, and pyrazinamide.
2. Alsultan A, Peloquin CA. Therapeutic drug monitoring in the
Antimicrob Agents Chemother. 1997;41(12):2670–9.
treatment of tuberculosis: an update. Drugs. 2014;74(8):839–54.
21. Seng KY, Hee KH, Soon GH, et al. Population pharmacoki-
3. World Health Organization. Essential medicines list. 20th ed.
netic analysis of isoniazid, acetylisoniazid, and isonicotinic
2017. http://www.who.int/medici​ nes/public​ ation​ s/essent​ ialme​ dici​
acid in healthy volunteers. Antimicrob Agents Chemother.
nes/en/. Accessed 7 Dec 2020.
2015;59(11):6791–9.
4. Wilkins JJ, Landon G, McIlleron H, et al. Variability in the popu-
22. Zvada SP, Denti P, Donald PR, et al. Population pharmacokinet-
lation pharmacokinetics of isoniazid in South African tuberculosis
ics of rifampicin, pyrazinamide and isoniazid in children with
patients. Br J Clin Pharmacol. 2011;72(1):51–62.
tuberculosis: in silico evaluation of currently recommended
5. Sundell J, Bienvenu E, Janzén D, et al. Model-based assessment
doses. J Antimicrob Chemother. 2014;69(5):1339–49.
of variability in isoniazid pharmacokinetics and metabolism in
23. Semvua HH, Mtabho CM, Fillekes Q, et al. Efavirenz, tenofovir
patients co-infected with tuberculosis and HIV: implications for
and emtricitabine combined with first-line tuberculosis treat-
a novel dosing strategy. Clin Pharmacol Ther. 2020;108(1):73–80.
ment in tuberculosis-HIV-coinfected Tanzanian patients: a phar-
6. Pasipanodya JG, Srivastava S, Gumbo T. Meta-analysis of clinical
macokinetic and safety study. Antivir Ther. 2013;18(1):105–13.
studies supports the pharmacokinetic variability hypothesis for
24. Tostmann A, Mtabho CM, Semvua HH, et al. Pharmacokinetics
acquired drug resistance and failure of antituberculosis therapy.
of first-line tuberculosis drugs in Tanzanian patients. Antimi-
Clin Infect Dis. 2012;5(2):169–77.
crob Agents Chemother. 2013;57(7):3208–13.
7. Sturkenboom MG, Akkerman OW, van Altena R, et al. Dosage
25. Burhan E, Ruesen C, Ruslami R, et al. Isoniazid, rifampin, and
of isoniazid and rifampicin poorly predicts drug exposure in
pyrazinamide plasma concentrations in relation to treatment
tuberculosis patients. Eur Respir J. 2016;48(4):1237–9.
response in Indonesian pulmonary tuberculosis patients. Anti-
8. Parkin DP, Vandenplas S, Botha FJ, et al. Trimodality of iso-
microb Agents Chemother. 2013;57(8):3614–9.
niazid elimination: phenotype and genotype in patients with
26. Aarnoutse RE, Kibiki GS, Reither K, et al. Pharmacokinetics,
tuberculosis. Am J Respir Crit Care Med. 1997;155(5):1717–22.
tolerability, and bacteriological response of rifampin admin-
9. Jayaram R, Shandil RK, Gaonkar S, et  al. Isoniazid phar-
istered at 600, 900, and 1,200 milligrams daily in patients
macokinetics-pharmacodynamics in an aerosol infection
with pulmonary tuberculosis. Antimicrob Agents Chemother.
model of tuberculosis. Antimicrob Agents Chemother.
2017;61(11):e01054-17.
2004;48(8):2951–7.
27. Boeree MJ, Heinrich N, Aarnoutse RE, et  al. High-dose
10. Jayaram R, Gaonkar S, Kaur P, et al. Pharmacokinetics-pharma-
rifampicin, moxifloxacin, and SQ109 for treating tuberculosis:
codynamics of rifampin in an aerosol infection model of tuber-
a multi-arm, multi-stage randomised controlled trial. Lancet
culosis. Antimicrob Agents Chemother. 2003;47(7):2118–24.
Infect Dis. 2017;17(1):39–49.
11. Gumbo T, Louie A, Liu W, et al. Isoniazid bactericidal activity
28. van Beek SW, ter Heine R, Keizer RJ, et al. Personalized tuber-
and resistance emergence: integrating pharmacodynamics and
culosis treatment through model-informed dosing of rifampicin.
pharmacogenomics to predict efficacy in different ethnic popu-
Clin Pharmacokinet. 2019;58(6):815–26.
lations. Antimicrob Agents Chemother. 2007;51(7):2329–36.
29. Sturkenboom MG, van der Lijke H, Jongedijk EM, et  al.
12. Pasipanodya JG, McIlleron H, Burger A, et al. Serum drug con-
Quantification of isoniazid, pyrazinamide and ethambutol in
centrations predictive of pulmonary tuberculosis outcomes. J
serum using liquid chromatography-tandem. J Appl Bioanal.
Infect Dis. 2013;208(9):1464–73.
2015;1(3):89–98.
13. Rockwood N, Pasipanodya JG, Denti P, et al. Concentration-
30. Aarnoutse RE, Sturkenboom MGG, Robijns K, et al. An inter-
dependent antagonism and culture conversion in pulmonary
laboratory quality control programme for the measurement of
tuberculosis. Clin Infect Dis. 2017;64(10):1350–9.
tuberculosis drugs. Eur Respir J. 2015;46(1):268–71.
14. Magis-Escurra C, Later-Nijland HMJ, Alffenaar JWC, et al.
31. R-Core-Team. R: a language and environment for statistical
Population pharmacokinetics and limited sampling strategy for
computing. 2016. https​://www.R-proje​ct.org/. Accessed 7 Dec
first-line tuberculosis drugs and moxifloxacin. Int J Antimicrob
2020.
Agents. 2014;44(3):229–34.
32. Beal S, Sheiner LB, Boeckmann A, Bauer RJ. NONMEM user’s
15. Maze MJ, Paynter J, Chiu W, et al. Therapeutic drug moni-
guides (1989–2009). Ellicott City: Icon Development Solutions;
toring of isoniazid and rifampicin during anti-tuberculosis
2009.
treatment in Auckland, New Zealand. Int J Tuberc Lung Dis.
33. Keizer RJ, Karlsson MO, Hooker A. Modeling and simulation
2016;20(7):955–60.
workbench for NONMEM: tutorial on Pirana, PsN, and Xpose.
16. Zuur MA, Pasipanodya JG, van Soolingen D, et al. Interme-
CPT Pharmacometrics Syst Pharmacol. 2013;2(6):e50.
diate susceptibility dose-dependent breakpoints for high-
34. Svensson E, van der Walt JS, Barnes KI, et al. Integration of
dose rifampin, isoniazid, and pyrazinamide treatment in
data from multiple sources for simultaneous modelling analysis:
multidrug-resistant tuberculosis programs. Clin Infect Dis.
experience from nevirapine population pharmacokinetics. Br J
2018;67(11):1743–9.
Clin Pharmacol. 2012;74(3):465–76.
17. Chirehwa MT, McIlleron H, Wiesner L, et al. Effect of efa-
35. Anderson BJ, Holford NH. Mechanism-based concepts of size
virenz-based antiretroviral therapy and high-dose rifampicin
and maturity in pharmacokinetics. Annu Rev Pharmacol Toxi-
on the pharmacokinetics of isoniazid and acetyl-isoniazid. J
col. 2008;48:303–32.
Antimicrob Chemother. 2019;74(1):139–48.
36. Carlsson KC, Savić RM, Hooker AC, et al. Modeling subpopula-
18. Denti P, Jeremiah K, Chigutsa E, et  al. Pharmacokinetics
tions with the $MIXTURE subroutine in NONMEM: finding the
of isoniazid, pyrazinamide, and ethambutol in newly diag-
individual probability of belonging to a subpopulation for the
nosed pulmonary TB patients in Tanzania. PLoS ONE.
use in model analysis and improved decision making. AAPS J.
2015;10(10):e0141002.
2009;11(1):148–54.
A Model-Informed Method for Dosing of Isoniazid 953

37. Beal SL. Ways to fit a PK model with some data below of an adequate test set in multivariate calibration. Chemometric
the quantification limit. J Pharmacokinet Pharmacodyn. Intell Lab Syst. 1999;49(1):79–89.
2001;28(5):481–504. 45. Keizer RJ, ter Heine R, Frymoyer A, et al. Model-informed pre-
38. Dosne AG, Bergstrand M, Harling K, et  al. Improving the cision dosing at the bedside: scientific challenges and opportuni-
estimation of parameter uncertainty distributions in nonlinear ties. CPT Pharmacometrics Syst Pharmacol. 2018;7(12):785–7.
mixed effects models using sampling importance resampling. J 46. Gordi T, Xie R, Huong NV, et al. A semiphysiological pharma-
Pharmacokinet Pharmacodyn. 2016;43(6):583–96. cokinetic model for artemisinin in healthy subjects incorporat-
39. Dosne AG, Bergstrand M, Karlsson MO. An automated sam- ing autoinduction of metabolism and saturable first-pass hepatic
pling importance resampling procedure for estimating parameter extraction. Br J Clin Pharmacol. 2005;59(2):189–98.
uncertainty. J Pharmacokinet Pharmacodyn. 2017;44(6):509–20. 47. Aarnoutse R. Pharmacogenetics of antituberculosis drugs
40. Bergstrand M, Hooker AC, Wallin JE, et al. Prediction-corrected Antituberculosis Chemother. 2011;40:176–90.
visual predictive checks for diagnosing nonlinear mixed-effects 48. Abrantes JA, Jönsson S, Karlsson MO, et al. Handling interoc-
models. AAPS J. 2011;13(2):143–51. casion variability in model-based dose individualization
41. Arshad U, Chasseloup E, Nordgren R, et al. Development of using therapeutic drug monitoring data. Br J Clin Pharmacol.
visual predictive checks accounting for multimodal parameter 2019;85(6):1326–36.
distributions in mixture models. J Pharmacokinet Pharmacodyn. 49. Donagher J, Martin JH, Barras MA. Individualised medicine: why
2019;46(3):241–50. we need Bayesian dosing. Intern Med J. 2017;47(5):593–600.
42. Magis-Escurra C, van den Boogaard J, Ijdema D, et al. Thera-
peutic drug monitoring in the treatment of tuberculosis patients.
Pulmon Pharmacol Ther. 2012;25(1):83–6.
43. Sheiner LB, Beal SL. Some suggestions for measuring predictive
performance. J Pharmacokinet Biopharm. 1981;9(4):503–12.
44. Faber NM. Estimating the uncertainty in estimates of root mean
square error of prediction: application to determining the size