Individualised dosing algorithm and personalised treatment of high‐dose rifampicin for tuberculosis

Abstract

Aims

To propose new exposure targets for Bayesian dose optimisation suited for high‐dose rifampicin and to apply them using measured plasma concentrations coupled with a Bayesian forecasting algorithm allowing predictions of future doses, considering rifampicin's auto‐induction, saturable pharmacokinetics and high interoccasion variability.

Methods

Rifampicin exposure targets for Bayesian dose optimisation were defined based on literature data on safety and anti‐mycobacterial activity in relation to rifampicin's pharmacokinetics i.e. highest plasma concentration up to 24 hours and area under the plasma concentration–time curve up to 24 hours (AUC_0–24h). Targets were suggested with and without considering minimum inhibitory concentration (MIC) information. Individual optimal doses were predicted for patients treated with rifampicin (10 mg/kg) using the targets with Bayesian forecasting together with sparse measurements of rifampicin plasma concentrations and baseline rifampicin MIC.

Results

The suggested exposure target for Bayesian dose optimisation was a steady state AUC_0–24h of 181–214 h × mg/L. The observed MICs ranged from 0.016–0.125 mg/L (mode: 0.064 mg/L). The predicted optimal dose in patients using the suggested target ranged from 1200–3000 mg (20–50 mg/kg) with a mode of 1800 mg (30 mg/kg, n = 24). The predicted optimal doses when taking MIC into account were highly dependent on the known technical variability of measured individual MIC and the dose was substantially lower compared to when using the AUC_0–24h‐only target.

Conclusions

A new up‐to‐date exposure target for Bayesian dose optimisation suited for high‐dose rifampicin was derived. Using measured plasma concentrations coupled with Bayesian forecasting allowed prediction of the future dose whilst accounting for the auto‐induction, saturable pharmacokinetics and high between‐occasion variability of rifampicin.

What is already known about this subject

• Rifampicin for tuberculosis treatment is a good candidate for therapeutic drug monitoring (TDM)‐based dose individualisation

• Rifampicin has auto‐induction, saturable pharmacokinetics and high interoccasion variability which complicates TDM

• Tuberculosis treatment has moved towards a high‐dose rifampicin paradigm and rifampicin exposure targets used previously require an update

What this study adds

• We present a complete model‐based TDM approach based on new targets that are up‐to‐date with recent high‐dose rifampicin data allowing dose predictions, even after the first dose

• The developed approach allows for individualised rifampicin treatment to assist physicians with improving the treatment of patients through optimised dose individualisation

1. INTRODUCTION

Rifampicin is crucial for treating drug‐susceptible tuberculosis (TB).1 Low rifampicin exposure has been suggested to increase the risk of treatment failure.2, 3 Therefore, individualised dosing guided by plasma concentrations is becoming increasingly used for rifampicin to identify sub‐therapeutic drug exposure followed by appropriate dose adjustments. An exposure metric of peak‐ or 2‐hour concentration of >8 mg/L represents a common exposure target for rifampicin therapeutic drug monitoring (TDM) during standard TB treatment.4, 5 The current targets reflect normal drug concentration ranges for rifampicin and were derived from pharmacokinetic studies in healthy volunteers and TB patients treated with the currently recommended dose of 10 mg/kg rifampicin.2, 4, 6, 7, 8, 9, 10 The standard dose of rifampicin is now under scrutiny and studies indicate faster and improved clinical outcome with higher doses.11, 12, 13 Thus, updated exposure targets would need to take higher doses and exposure levels into consideration.

A significantly improved time‐to‐culture conversion during the first 12 weeks of treatment was seen using a rifampicin dose of 35 vs 10 mg/kg.11, 12 In a 2‐week rifampicin trial,14 a significant increase in anti‐mycobacterial activity was demonstrated for 35 compared to 10 mg/kg.13 While awaiting confirmatory data on safety and efficacy of 35 mg/kg from larger trials, these initial findings from 79 patients are encouraging.

TDM already has a place within TB treatment,4 including examples of dose individualisation using Bayesian forecasting.15, 16, 17 Dose individualisation using Bayesian forecasting improves on traditional TDM since it gives better yield of pharmacokinetic information, especially for a limited set of plasma concentration samples from a patient. A dose‐individualisation strategy using measured plasma concentrations coupled with Bayesian forecasting is a 2‐step process consisting of: (i) an estimation step where individual pharmacokinetic information is obtained from available pharmacokinetic data using a population pharmacokinetic model; followed by (ii) prediction of the future dose using simulations based on an exposure target for Bayesian dose‐optimisation.18 Model‐based TDM examples for rifampicin19, 20, 21 focus only on the first step of the dose individualisation approach with a main goal of developing approaches that can reliably estimate an exposure metric of interest, e.g. the area under the plasma concentration–time curve up to 24 hours (AUC_0–24h) using limited plasma sampling. Most previous rifampicin model‐based TDM examples19, 20 do not specify how to predict an optimal individual dose but rather use exposure target values to confirm if the currently administered dose is appropriate by comparing the observed AUC_0–24h with the exposure target. Also, current rifampicin model‐based TDM examples19, 20, 21 do not describe procedures for handling the high interoccasion variability (IOV) of rifampicin pharmacokinetics. High IOV makes dose individualisations difficult and should be addressed appropriately.22, 23 The most appropriate way is to take IOV into account when estimating individual pharmacokinetic parameters but not when predicting the future dose as described in.24, 25

Most Bayesian dose individualisation methods for rifampicin19, 20 have employed pharmacokinetic models that assume linear pharmacokinetics which carries a risk for over‐simplification, since rifampicin has auto‐induction and saturable pharmacokinetics.26 Linear models have been proven successful in describing the pharmacokinetics of 10 mg/kg rifampicin27, 28 and these simpler approaches for individualised dosing may well predict accurate doses in the dose range of around 10 mg/kg. However, at 35 mg/kg there are large deviations from linearity in observed plasma concentrations14 and if unaccounted for, the dosing recommendations will be too inaccurate, which is further complicated by the auto‐induction that needs to be handled in order to predict an accurate dose at any day.

Finally, current model‐based TDM methods for rifampicin19, 20, 21 are specific for pharmacokinetic data collected after at least week 2 of treatment which is an assumed steady‐state for rifampicin auto‐induction.29 However, using an approach where the auto‐induction is an integrated part of the pharmacokinetic model, it is possible to perform dose individualisation with pharmacokinetic data collected at any day of treatment, including day 1.

Including MIC information in addition to drug concentrations relates the individual dosing decision to the susceptibility of the pathogen. There are suggested exposure targets for rifampicin including MIC.20 Due to the imprecise nature of MIC measurements it is currently debated if it is applicable to base individual dosing decisions on MIC.30, 31

We aimed to propose new exposure targets for Bayesian dose‐optimisation well‐suited for high‐dose rifampicin and to apply them using measured plasma concentrations coupled with a Bayesian forecasting algorithm allowing for dose predictions of the future dose, considering rifampicin's saturable pharmacokinetics, auto‐induction and high IOV.

To meet the aim, we used a workflow consisting of 2 principal steps of first defining new exposure targets for Bayesian dose optimisation using simulations from a previously developed pharmacokinetic model. This was followed by a demonstration of how the targets should be applied by predicting the optimal dose on a retrospective example dataset.

2. METHODS

2.1. New exposure targets for Bayesian dose optimisation

Exposure targets for Bayesian dose optimisation were proposed for different exposure metrics with and without taking literature‐derived MICs into account. The model‐based approach suggested in this work was designed to handle IOV according to24, 25 and to allow prediction of exposures at steady‐state regardless of what day of treatment samples are collected at, including day 1.

The exposure targets were based on model‐predicted exposure metrics including the highest plasma concentration up to 24 hours (C_max) and AUC_0–24h following 35 mg/kg rifampicin. A 35 mg/kg dose has not been shown to be safe in long‐term trials but initial data are promising11, 14 with potential for better efficacy than 10 mg/kg.11, 12, 13 The exposure targets without MIC were derived using the HIGHRIF1 trial as a reference population.14 The predicted exposure metric (C_max and AUC_0–24h) were calculated for the typical subject in HIGHRIF1,14 weighing 53.9 kg with a fat‐free mass of 44.6 kg, using a previously developed rifampicin population pharmacokinetic model.32 A typical prediction from a population model is the predicted AUC_0–24h or C_max generated by model predictions without the influence of within‐ or between subject variability. Briefly, the employed pharmacokinetic model32 was developed on observations from HIGHRIF1 collected for 14 days from 83 TB patients receiving 10–40 mg/kg daily rifampicin.14 The model included capacity‐limited elimination and dose‐dependent bioavailability, both contributing to a nonlinear, more‐than‐proportional increase in exposure at higher rifampicin doses. The model also included time‐dependent elimination since rifampicin increases its own elimination over time, termed auto‐induction.32 The exposure targets were derived as intervals, centred around 35 mg/kg rifampicin. The lower limit of the interval was the mid‐point between the typically predicted C_max or AUC_0–24h for 30 and 35 mg/kg rifampicin. The upper limit was the mid‐point between C_max or AUC_0–24h for 35 and 40 mg/kg rifampicin. The C_max and AUC_0–24h were calculated as total (bound + unbound) concentrations (calculations described in detail below). Note that this way of deriving an exposure target was specific for high‐dose rifampicin where the target was derived based on targeting the exposure of a 35 mg/kg dose and differs from the traditional way of deriving exposure targets or therapeutic windows. A therapeutic window is traditionally defined based on rigorous knowledge on the systemic exposure–response relationship for safety and the clinical endpoint which is currently unavailable for high‐dose rifampicin. Another implication of basing the target on expected exposure given a certain dose was that the targets were dependent on what day of treatment they were derived for, because of time‐varying pharmacokinetics of rifampicin. However, in this work we applied a steady‐state target, assumed to occur on day 56,32 to a patient dataset (see below) but targets for additional days were also derived.

To derive the MIC‐based targets including C_max/MIC and AUC_0–24h/MIC, a fixed literature‐derived MIC value was used for all individuals in the reference population (however, note that for the example dataset described below, individual observed MIC data existed). This was done since the reference population used to define the exposure targets, the HIGHRIF1 trial, did not include determination of individual MICs.14 A median, or mode MIC of 0.125 mg/L was applied for all individuals in the reference population and used to derive the MIC‐based targets. 0.125 mg/L represents the mode of MGIT‐based rifampicin MIC (MIC_MGIT) in a general TB patient population according to a previous publication.33 The targets were derived as ratios, dividing the corresponding pharmacokinetics‐only exposure metrics by the assumed HIGHRIF1 median MIC of 0.125 mg/L.

Sensitivity analyses were performed for the MIC‐based targets. One sensitivity analysis explored the impact of assuming a lower median MIC in HIGHRIF1 of 0.064 instead of 0.125 mg/L. Assuming that a lower MIC when defining the target represents a stricter target, the C_max/MIC or AUC_0–24h/MIC that needs to be achieved must increase when MIC decreases.

A second sensitivity analysis explored the impact of known method‐based variability30 in the measured MIC_MGIT. The individual MICs within an example dataset (where individual MICs were available, see below) were set to either half the MIC_MGIT or twice the MIC_MGIT. A variability of ±1 dilution step was considered likely, based on assays performed by different technicians and laboratories. The assumed MIC in HIGHRIF1, which we used to define the MIC‐based targets, was 0.125 mg/L for this second sensitivity analysis.

2.2. New exposure targets for Bayesian dose optimisation applied to study cohort

The exposure targets were applied to a subset of patients from a previous study34 conducted at Linköping University Hospital, Linköping, Sweden and Karolinska University Hospital, Solna, Sweden, hereafter referred to as the patient dataset. The study included adult patients (18–65 years) initiated on TB treatment. Only pulmonary TB patients were included in this analysis, treated according to World Health Organisation guidelines35: 5 mg/kg isoniazid, 10 mg/kg rifampicin, 15–20 mg/kg ethambutol and 20–30 mg/kg pyrazinamide for 2 months and then 4 months isoniazid and rifampicin. The study was approved by the regional ethical committee (DNR 2012/197–31) of Linköping University and conducted in accordance with the Declaration of Helsinki (clinicaltrials.gov NCT02042261).

Rifampicin plasma concentrations were measured at 0, 2, 4 and 6 hours after 2 weeks of treatment and at 0 and 2 hours on weeks 4 and 12. The plasma was stored at −70°C until analysed by liquid chromatography coupled to a mass spectrometer (Agilent, Santa Clara, CA, USA). The lower limit of quantification (LLOQ) was 0.05 mg/L.

Baseline morning sputum for MIC determination was collected from all patients. The MIC_MGIT was determined in BACTEC MGIT 960 (Becton Dickinson, Lakes, NJ, USA), a liquid culture‐based system using serial 2‐fold dilutions of rifampicin.36 The MIC was defined as the lowest rifampicin concentration with less growth than 1:100 dilution controls.

2.3. Dose predictions in patient dataset

To predict the dose for each individual, first, the previous rifampicin pharmacokinetic model32 was used without re‐estimation of population parameters to obtain individual pharmacokinetic parameters (i.e. the empirical Bayes estimates) using the population pharmacokinetic model including both IOV and between patient variability. All population pharmacokinetic model parameters were fixed,32 i.e. no model development was performed.

Secondly, the individual pharmacokinetic parameters were used to generate (i.e. simulate) virtual plasma concentration–time curves after different doses for each individual at steady‐state (day 56)32 using a differential equation solver. The virtual plasma concentration curves were generated only based on interindividual differences and thus the population pharmacokinetic model32 was used omitting IOV when predicting the individual concentration–time curves according to.24, 25 The C_max and AUC_0–24h were calculated from the virtual plasma concentration–time curves. The C_max was calculated using the derivative of the predicted plasma concentration vs time and C_max occurred when the derivative was equal to null. The AUC_0–24h was predicted by calculating the integral beneath the predicted plasma concentration vs time between 0 and 24 hours postdose.

The C_max and AUC_0–24h were predicted for 600, 900, 1200, 1500, 1800, 2100, 2400, 2700 and 3000 mg of daily rifampicin for each individual. The dose resulting in AUC_0–24h, C_max, AUC_0–24h/MIC or C_max/MIC within the developed targets was the optimal dose for each patient. However, given that the allowed doses were 600–3000 mg with dose increments limited to 300 mg it was possible that no dose fulfilled the criteria of falling within the target or that >1 dose fell within the target. The former was handled by choosing the dose that was closest to the target and the latter was handled by choosing the dose that was closest to the mid‐point of the lower and upper range of the target.

In order to verify the pharmacokinetic model, it was validated against the observed rifampicin concentrations in the patient dataset34 by performing a visual predictive check.37, 38

The analysis was conducted in NONMEM 7.3.39 Data processing was performed using R 3.4.3.40

3. RESULTS

3.1. New exposure targets for Bayesian dose optimisation

The exposure targets for Bayesian dose optimisation based on pharmacokinetics only were 33–38 mg/L for C_max and 181–214 h × mg/L for AUC_0–24h at steady state. Targets for days 1, 7, 14, 28 and 56 are shown in Table 1 but the approach can be applied to any day after first dose. The exposure targets for Bayesian dose optimisation were 264–304 for C_max/MIC and 1448–1712 for AUC_0–24h/MIC, both derived by assuming a fixed median and literature‐derived MIC_MGIT of 0.125 mg/L in HIGHRIF1, which was the reference population.

Table 1.

Exposure targets for Bayesian dose optimisation for selected time points after start of treatment

Time after dose (days)	Cmax (mg/L)	Cmax/MIC	AUC0–24h (h × mg/L)	AUC0–24h/MIC
1	37–42	296–336	342–408	2736–3264
7	37–43	296–344	217–259	1736–2072
14	33–38	264–304	189–224	1512–1792
28	33–38	264–304	182–215	1456–1720
56	33–38	264–304	181–214	1448–1712

The exposure targets for Bayesian dose optimisation are model‐based exposure targets where the predicted dose is the dose with an associated individually predicted model‐based C_max, C_max/MIC, AUC_0–24h or AUC_0–24h/MIC that do not include any interoccasion variability in pharmacokinetic parameters. The exposure targets for Bayesian dose optimisation were derived through typical model predictions from a population pharmacokinetic model and do not include any variability between or within individuals. The MIC‐based targets were derived as ratios dividing the pharmacokinetics‐only target by a fixed and literature‐derived MIC of 0.125 mg/L.

C_max; the highest plasma concentration up to 24 hours, MIC; minimum inhibitory concentration, AUC_0–24h; the area under the plasma concentration–time curve up to 24 hours

3.2. Dose predictions in the patient dataset

This analysis included 24 patients with pulmonary TB. Patient characteristics are summarised in Table 2.

Table 2.

Baseline characteristics in patients with pulmonary tuberculosis included in the study

Variable		Median or n (%)	Range
n		24 (100%)
Weight (kg)	Median	60.2	45.0–91.0
Height (m)	Median	1.67	1.55–1.91
Age (y)	Median	32	18–69
Female, n (%)		13 (54.2%)
Patients on 450 mg, n (%)		1 (4.17%)
Patients on 600 mg, n (%)		23 (95.8%)
HIV infection, n (%)		0 (0%)
Cavity/cavities on chest X‐ray		5 (20.8%)
Sputum smear positive		6 (25%)

In total, 179 plasma samples were included in the analysis (33.5% below LLOQ). Samples below LLOQ were included in the Bayesian step using a likelihood based method, as suggested by Beal.41 The observed rifampicin MIC ranged from 0.016 to 0.125 mg/L (mode: 0.064 mg/L) in the 21 patients with available MIC values (Figure S1). A visual predictive check (Figure S2) shows that the model32 described the data well given that no re‐estimation of parameters was performed, suggesting that the model was appropriate given the data.

The individual predicted model‐based exposure metrics at steady state without IOV, for increasing doses are shown in Figure 1.

Figure 1.

Individual model‐based predictions of C_max (A) and AUC_0–24h (B) at steady state for increasing doses of rifampicin. The dotted horizontal lines indicate the steady state exposure targets for Bayesian dose optimisation. Steady state was assumed to occur 56 days after start of treatment. C_max; the highest plasma concentration up to 24 hours, AUC_0–24h; the area under the plasma concentration–time curve up to 24 hours

For dosing based on AUC_0–24h the optimal dose ranged between 1200–3000 mg (mode: 1800 mg; Figure 2a and Table 3). Since the simulation step for predicting the individual dose was limited to doses between 600–3000 mg (with 300 mg increments) it was possible that no simulated dose gave an AUC_0–24h between 181 and 214 h × mg/L which occurred in 3 patients; the dose with an AUC_0–24h closest to the target was chosen where the deviation from the target was small (<7%) for all 3 patients.

Figure 2.

Comparisons of individual predicted doses according to the exposure targets for Bayesian dose optimisation. Panel (A) compares the predicted doses according to an AUC_0–24h‐based target vs a C_max‐based target, panel (B) compares dosing according to AUC_0–24h vs AUC_0–24h/MIC and panel (C) compares dosing according to C_max vs C_max/MIC. The distributions of the predicted doses are summarised for all patients as boxplots. Open circles show the predicted dose on an individual level where the change in the individual dose is indicated by dashed grey lines. AUC_0–24h; the area under the plasma concentration–time curve up to 24 hours, C_max; the highest plasma concentration up to 24 hours

Table 3.

Distribution of optimal doses according to the different exposure targets for Bayesian dose optimisation

	Pharmacokinetics onlya		Including MICb		Sensitivity analysis
					Cmax/MIC			AUC0–24h/MIC
	Cmax	AUC0–24h	Cmax/MIC	AUC0–24h/MIC	Target MIC = 0.064 mg/L	0.5 × observed MIC	2 × observed MIC	Target MIC = 0.064 mg/L	0.5 × observed MIC	2 × observed MIC
Rifampicin dose (mg)	n (%)	n (%)	n (%)	n (%)	n (%)	n (%)	n (%)	n (%)	n (%)	n (%)
600	0 (0)	0 (0)	8 (38.1)	7 (33.3)	2 (9.5)	18 (85.8)	2 (9.5)	1 (4.8)	15 (71.4)	1 (4.8)
900	0 (0)	0 (0)	6 (28.6)	4 (19.0)	5 (23.8)	1 (4.8)	5 (23.8)	5 (23.8)	3 (14.3)	4 (19.0)
1200	2 (8.3)	2 (8.3)	4 (19.0)	5 (23.8)	1 (4.8)	2 (9.5)	1 (4.8)	1 (4.8)	2 (9.5)	2 (9.5)
1500	8 (33.3)	7 (29.2)	1 (4.8)	2 (9.5)	4 (19.0)	0 (0)	3 (14.3)	4 (19.0)	0 (0)	3 (14.3)
1800	8 (33.3)	8 (33.3)	0 (0)	1 (4.8)	5 (23.8)	0 (0)	5 (23.8)	4 (19.0)	1 (4.8)	5 (23.8)
2100	2 (8.3)	2 (8.3)	0 (0)	1 (4.8)	0 (0)	0 (0)	1 (4.8)	1 (4.8)	0 (0)	1 (4.8)
2400	3 (12.5)	2 (8.3)	2 (9.5)	0 (0)	1 (4.8)	0 (0)	1 (4.8)	2 (9.5)	0 (0)	2 (9.5)
2700	0 (0)	2 (8.3)	0 (0)	1 (4.8)	0 (0)	0 (0)	0 (0)	1 (4.8)	0 (0)	1 (4.8)
3000	1 (4.2)	1 (4.2)	0 (0)	0 (0)	3 (14.3)	0 (0)	3 (14.3)	2 (9.5)	0 (0)	2 (9.5)

All doses were predicted based on the exposure target for Bayesian dose optimisation on day 14. Unless specified, all MIC‐based targets were derived as ratios dividing the pharmacokinetics‐only targets by a MIC of 0.125 mg/L

^aDoses based on pharmacokinetics only were predicted for all 24 patients.

^bDoses involving MIC were predicted for 21 patients with available MIC measurements.

C_max; the highest plasma concentration up to 24 hours, AUC_0–24h; the area under the plasma concentration–time curve up to 24 hours, MIC; minimum inhibitory concentration

For an optimal dose guided by our C_max‐based exposure target the predicted optimal dose ranged between 1200–3000 mg (mode: 1500 and 1800 mg; Figure 2A and Table 3). Six patients had no dose within the target but the deviation from the target was low (<3%).

The distributions of optimal doses across the dataset were slightly lower for C_max than AUC_0–24h according to the boxplots in Figure 2A. The dashed lines connecting the open circles show changes in the individually predicted dose visualising how the predicted dose differed between the exposure metrics for 7/24 patients. The difference was mainly small (±300 mg) although 1 patient had a predicted dose 900 mg higher for AUC_0–24h than C_max.

For the AUC_0–24h/MIC‐based target the predicted dose ranged between 600–2700 mg (mode: 600 mg; Table 3). Eleven patients had a dose that fell outside the target where 2 patients with predicted doses of 600 mg had AUC_0–24h/MICs 20% and 79% above the target, respectively. The remaining patients outside the target had small deviations from the target (<10%).

The predicted doses were lower for AUC_0–24h/MIC than AUC_0–24h (Figure 2B). The doses were considerably lower (≥600 mg) in 18/21 patients.

For C_max/MIC the predicted optimal doses ranged between 600–2400 mg (mode: 600 mg; Table 3). Sixteen of 21 patients had a dose resulting in a C_max/MIC outside the target. Seven patients with a predicted dose of 600 mg deviated from the target up to 73%, which was due to a C_max/MIC higher than 304 (the upper limit of the target) at 600 mg (the lowest possible dose). The high C_max/MIC for the lowest dose occurred in individuals with a combination of low MIC and individual pharmacokinetic parameters that resulted in high C_max such as low volume of distribution and/or fast absorption. Of the remaining 9 patients with C_max/MIC outside the target, the deviation from the target was low (<10%).

The predicted doses were lower for dosing based on C_max/MIC compared to C_max only (Figure 2C). All but 3 patients had a considerably lower dose according to C_max/MIC.

The sensitivity analysis for assuming a lower fixed median MIC of 0.064 mg/L when defining the exposure target instead of 0.125 mg/L gave e.g. a stricter steady state AUC/MIC target of 2785–3292 instead of 1448–1712. The predicted doses were substantially higher for the stricter target. This was true for both AUC_0–24h/MIC and C_max/MIC (Table 3 and Figure S3). Note that a lower MIC for defining the target gives a target requiring a higher dose as the individual patient MIC remained unchanged which contrasts a lower MIC in an individual patient which is expected to require a lower dose.

The sensitivity analysis for uncertainty in individually observed MIC showed that the predicted dose varied substantially when the individually reported MICs in the patient dataset were either set to half the reported value or to twice the reported value, which was evident for both exposure metrics. Assuming that the individual MIC was half of the reported MIC value resulted in lower dose predictions (only 3 and 6 patients for C_max/MIC and AUC_0–24h/MIC, respectively, had doses >10 mg/kg, Table 3 and Figure S4). On the contrary, when the individually observed MIC was set to twice the reported value the predicted doses increased considerably. Note that this part of the sensitivity analysis, which concerns individually observed MIC, is expected to give a lower dose for a lower MIC (i.e. a patient with a very sensitive strain requires a lower dose). This contrasts with the first part of the sensitivity analysis where a lower assumed literature‐derived MIC in the reference population, which was used to derive the targets, led to higher doses (i.e. a lower MIC in the reference population gives a stricter target).

4. DISCUSSION

This work demonstrates a new exposure target for high‐dose rifampicin applied using measured plasma concentrations coupled with Bayesian forecasting allowing individualised predictions of the future dose from pharmacokinetic data collected on any day of treatment, including day 1. The approach handles the complex pharmacokinetics and high between‐occasion variability for rifampicin and is therefore warranted.

New optimised exposure targets were suggested based on different exposure metrics including C_max or AUC_0–24h, with and without inclusion of MIC. Both exposure metrics predicted similar doses but since previous studies suggest that AUC_0–24h has better correlation with anti‐mycobacterial killing,12, 42 an AUC_0–24h‐based target is suggested as a priority for future implementation. Targets incorporating rifampicin MIC were explored. The predicted doses were dependent on the assumed median MIC in the reference population, where exposure targets were derived by setting all individual MICs to the mode (0.125 mg/L). Additionally, the simulated technical variability of MIC testing showed a high impact on dose prediction, indicating a need for further studies before MICs can be included as an exposure target in our model.

An optimised exposure target for Bayesian dose optimisation of AUC_0–24h 181–214 h × mg/L at steady‐state is a suggested priority to test clinically. This target is up‐to‐date with recent data on high‐dose rifampicin with potential for more effective treatment than previous exposure targets whilst still being safe. Rifampicin 35 mg/kg was safe in 2‐ and 12‐week studies.11, 14 Despite long‐term safety data not yet being available, these initial safety data are promising. Even though long‐term efficacy data are unavailable for 35 mg/kg, evidence from 2 and 12 weeks are encouraging with increased bacterial killing at high rifampicin exposures.12, 13 Another suggested dose is 40 mg/kg which appeared safe in 15 patients for 2 weeks.32 However, we based our targets on 35 mg/kg due to limited data from 40 mg/kg. This way of deriving an exposure target by targeting the exposure of a certain dose is uniquely tailored to the situation of high‐dose rifampicin due to insufficient data and differs from how exposure targets are traditionally derived. Our suggested exposure target can be refined if needed as data become available but, given current data, the target is relevant.

An overview of how to practically implement the suggested approach which also highlights the ability of the approach to predict the optimal dose on day 1 is shown in Figure 3. We suggest inclusion of pharmacokinetic sampling on at least 2 occasions due to the IOV of rifampicin pharmacokinetics. There should be sufficient time between visits to analyse pharmacokinetic samples (using e.g. liquid chromatography–tandem mass spectrometry) and to apply the Bayesian forecasting procedure on gathered plasma concentration data. The forecasting procedure is quick (minutes) without any need for model development. It is preferable to collect blood samples early, even after the first dose, to allow for a timely initial dose prediction. When data are collected at later time‐points, the forecasting procedure can be updated with new plasma concentration data, thus increasing the precision of the dose prediction as indicated by the lower panel in Figure 3 (the graph is arbitrary and illustrates that more data give higher precision). The exact timing and number of pharmacokinetic samples and occasions and whether to include MIC should be tailored to the individual patient and the clinical setting. The employed rifampicin pharmacokinetic model32 was shown to reliably predict rifampicin AUC_0–24h using 2‐ and 4‐hour samples.21

Figure 3.

Suggested practical implementation of the advocated approach using measured plasma concentrations coupled with Bayesian forecasting. The top panel shows a suggested time‐line and sampling scheme for minimum inhibitory concentration (MIC) and pharmacokinetics (PK) where dots (···) indicate waiting time for MIC readout. Middle panel shows the Bayesian forecasting step where exposures for increasing rifampicin are simulated to predict the dose which leads to a dose decision. Exemplified here is a patient with dose increase, as seen on the dosing time‐line, based on the dose prediction for visit 1. The updated dose prediction, based on data for later visits, confirmed the initially predicted dose. Some time (e.g. a few days up to a week) will be required in between collection of PK data and the dose decision in order to analyse the samples and to conduct the Bayesian forecasting step. The lower panel shows the arbitrary increase in precision as more data are collected. The precision has no unit and only serves to illustrate the fundamental principle that more data leads to a more precise dose prediction. The third visit is optional and is suggested in order to increase the precision of the dose and the exact time‐point for visit 3 may vary, e.g. from day 14–28, depending on the clinical setting. At least 2 visits with collection of PK data are strongly recommended due to the high interoccasion variability of rifampicin PK

We consider this model‐based approach to be optimal since firstly, model‐based approaches give the most information on individual pharmacokinetic parameters, especially when limited plasma samples are available. Furthermore, the nonlinear pharmacokinetics of rifampicin was accounted for. The model32 predicted 5‐fold increased AUC_0–24h at 2100 vs 600 mg. This more‐than‐linear increased exposure will not be predicted by models where exposure increases linearly with dose. Thus, linear models under‐predict exposure metrics at higher doses, resulting in unnecessary high dose recommendations.

The suggested dose individualisation approach addresses the problem of using measured plasma concentrations coupled with Bayesian forecasting for drugs with high IOV which normally causes too uncertain predictions when including data from multiple occasions.22 Variability between occasions is important for dose individualisation23 and were handled appropriately in our approach.24, 25 To ensure valid dosing predictions with IOV present, we recommended pharmacokinetic sampling on at least 2 separate occasions.

Our approach was designed to give procedures for specific individual dose recommendations which contrasts previous model‐based TDM examples for rifampicin19, 20, 21 which are more focused towards evaluating if the currently administered dose is appropriate or not. Since our approach involves a simulation step where AUC_0–24h for a wide range of potential doses are predicted, it results in specific dose recommendations. The most important improvement with our work compared to previous examples19, 20, 21 is that we suggest new high‐dose rifampicin exposure targets. Furthermore, our approach specifies how to predict doses not dependent on pharmacokinetic data collected at day 14 or later; instead, the dose can be predicted based on data collected on any day of treatment including day 1.

Targets incorporating MIC were explored but due to the lack of individual MIC data for defining the exposure targets, our results should be interpreted with caution and further studies are needed. A problematic issue for implementing MICs in the target is the technical variability of the observed individual MIC values in our patient dataset (±1 dilution step). Since no MICs were determined in HIGHRIF1, a literature‐derived MIC value of 0.125 mg/L was assumed, which is a limitation with this analysis. If MICs would have been determined in HIGHRIF1 it may have been possible to derive a more accurate target. Rifampicin MIC has been shown to be a predictor for long‐term outcome in a clinical trial43 but the trial did not include pharmacokinetic observations. We inquire for future studies with joint observations of MIC, pharmacokinetics and pharmacodynamics such as,44 in order to develop a better MIC‐based target.

The simulated doses within each individual were defined as precise mg doses (i.e. flat dosing), as opposed to mg/kg dosing (i.e. weight‐based dosing), which is standard dosing for rifampicin. The reason for using weight‐based dosing is to achieve similar exposures despite different body weights. This benefit is not relevant for our suggested dose individualisation approach since the method itself aims to control the exposure in each individual patient.

Since no dose adjustments were performed according to the suggested targets in the underlying study,34 the targets require confirmation in future clinical studies. Nevertheless, this work exemplifies how dose individualisation of rifampicin ideally should be performed.

A limitation is the limited number of patients in the study (n = 24), thereby reducing precision. Since all patients were included in Sweden, the generalisability of the distributions of optimal doses (Table 3) to other datasets is unclear.

A large proportion of the observed drug concentration data were below LLOQ (33.5%), which may lead to imprecise individual pharmacokinetic parameters. However, this was given that the patients in this study only received 10 mg/kg rifampicin. If our approach is implemented, patients will be likely to receive higher doses than 10 mg/kg, leading to lower proportion of data below LLOQ, which may increase precision of individual pharmacokinetic parameters.

In conclusion, we present a new dose individualisation approach using measured plasma concentrations coupled to Bayesian forecasting including dose predictions for high‐dose rifampicin TB treatment, warranted due to complex dose‐ and time‐dependent pharmacokinetics and high between‐occasion variability of rifampicin.

COMPETING INTERESTS

There are no competing interests to declare.

CONTRIBUTORS

R.J.S. and U.S.H.S. drafted the manuscript and analysed the data. K.N., L.D.F., J.B., J.P., E.E. and T.S. revised the manuscript critically for intellectual content and contributed to acquisition of data. All authors gave final approval of the submitted version and agreed to be accountable for all aspects of the work.

Supporting information

Figure S1 Observed distribution of minimum inhibitory concentration within the study population

Figure S2 Visual predictive check of the observed data. The observations are represented by open circles, shown as actual values in the upper panel and the proportion of samples below the lower limit of quantification (BLOQ) in the lower limit of quantification (LLOQ) panel. The lower dashed, solid and upper dashed lines in the upper panel are the 2.5^th, 50^th and 97.5^th percentiles of the observed data, respectively and the lower blue, middle red and upper blue shaded areas are the 95% confidence interval of the predicted 2.5^th, 50^th and 97.5^th percentiles, respectively. In the lower panel, the grey shaded area is the 95% confidence interval of the predicted proportion of samples BLOQ.

Figure S3 Comparisons of individual predicted doses according to the minimum inhibitory concentration (MIC)‐based exposure targets for Bayesian dose optimisation assuming a fixed individual MIC at either 0.125 or 0.064 mg/L as the assumed median MIC when defining the target. Panel (a) compares the predicted doses according to an AUC_0–24h/MIC‐based target and panel (b) compares dosing according to a C_max/MIC‐based target. The distributions of the predicted doses are summarised for all patients as boxplots. Open circles show the predicted dose on and individual level where the change in the individual dose is indicated by dashed grey lines. AUC_0–24h; the area under the plasma concentration–time curve up to 24 hours, C_max; the highest plasma concentration up to 24 hours

Figure S4 Comparisons of individual predicted doses according to the evaluation of uncertainty in measured minimum inhibitory concentration (MIC). From left to right the figure compares the predicted dose assuming half the reported MIC, the actual reported MIC and twice the MIC for an AUC_0–24h/MIC‐based target in panel (a) and for a C_max/MIC‐based target in panel (b). The distributions of the predicted doses are summarised for all patients as boxplots. Open circles show the predicted dose on and individual level where the change in the individual dose is indicated by dashed grey lines. TDM; therapeutic drug monitoring, AUC_0–24h; the area under the plasma concentration–time curve up to 24 hours, C_max; the highest plasma concentration up to 24 hours

Svensson RJ, Niward K, Davies Forsman L, et al. Individualised dosing algorithm and personalised treatment of high‐dose rifampicin for tuberculosis. Br J Clin Pharmacol. 2019;85:2341–2350. 10.1111/bcp.14048

DATA AVAILABILITY STATEMENT

Research data are not shared.