ABSTRACT
Attainment of appropriate pharmacokinetic-pharmacodynamic (PK-PD) targets for antimicrobial treatment is challenging in critically ill patients, particularly for cefepime, which exhibits a relative narrow therapeutic-toxic window compared to other beta-lactam antibiotics. Target-controlled infusion (TCI) systems, which deliver drugs to achieve specific target drug concentrations, have successfully been implemented for improved dosing of sedatives and analgesics in anesthesia. We conducted a clinical trial in an intensive care unit (ICU) to investigate the performance of TCI for adequate target attainment of cefepime. Twenty-one patients treated with cefepime according to the standard of care were included. Cefepime was administered through continuous infusion using TCI for a median duration of 4.5 days. TCI was based on a previously developed population PK model incorporating the estimated creatinine clearance based on the Cockcroft-Gault formula as the input variable to calculate cefepime clearance. A cefepime blood concentration of 16 mg/liter was targeted. To evaluate the measured versus predicted plasma concentrations, blood samples were taken (median of 10 samples per patient), and total cefepime concentrations were measured using ultraperformance liquid chromatography-tandem mass spectrometry. The performance of the TCI system was evaluated using Varvel criteria. Half (50.3%) of the measured cefepime concentrations were within ±30% around the target value of 16 mg liter−1. The wobble was 11.4%, the median performance error (MdPE) was 21.1%, the median absolute performance error (MdAPE) was 32.0%, and the divergence was −3.72% h−1. Based on these results, we conclude that TCI is useful for dose optimization of cefepime in ICU patients. (This study has been registered at ClinicalTrials.gov under identifier NCT02688582.)
INTRODUCTION
Inappropriate dosing of antibiotics is a driver for antimicrobial resistance development (1), acute toxicity (2, 3), and poor clinical outcomes (4, 5). This is particularly true for cefepime, a fourth-generation cephalosporin which has been shown to exhibit a narrow therapeutic-toxic window (2, 3, 6). Defining adequate dosing regimens in critically ill patients is challenging, as pharmacokinetics (PK) in these patients are known to vary considerably (4, 7–13), and these patients are more likely to be infected by less susceptible bacteria (12).
Traditionally, dosing of antibiotics is based on nomograms, which define a dosing regimen based on one or a limited set of patient covariates. In the critically ill, these nomogram-based dosing regimens frequently result in a significant proportion of patients not achieving the therapeutic target (4). Hence, treatment should be individualized using therapeutic drug monitoring and/or population PK (PopPK) models. In recent years, several software packages were developed that allow model-based treatment individualization (13). While therapeutic drug monitoring (TDM) linked with Bayesian forecasting provides a powerful opportunity for delivering individualized care for patients (14), several issues in current strategies for dose optimization of antimicrobials have hindered clinical implementation in most intensive care units (ICUs) (15, 16).
Target-controlled infusion (TCI) is a technique of continuously infusing intravenous (i.v.) drugs and is mainly known in the field of anesthetics (17). TCI allows the clinician to target a predefined concentration in a specific body compartment or tissue of interest. The computer then calculates the optimal infusion rate required to achieve this user-defined target concentration as fast as possible without overshooting the target, based on a PopPK model and patient-specific covariates (e.g., age, weight, and serum creatinine level, etc.), which are integrated into the model. An online-coupled infusion pump then delivers this optimal infusion regimen to the patient. In comparison to the above-mentioned manually controlled infusions, TCI systems might provide a more convenient and performant alternative. Treatment individualization is made easy, as the PopPK model and associated covariates are embedded in the TCI devices. Dose adaptations are not limited to practicable changes in infusion rates, dose strengths, and dosing intervals, etc., but TCI continuously calculates and adjusts the infusion rate to exactly match the distribution and elimination kinetics of the drug during treatment.
In this prospective pharmacokinetic study, we evaluated the performance of a cefepime TCI system in a cohort of critically ill patients. Furthermore, the additional PK data were used to update the previously presented PopPK model for cefepime (18).
RESULTS
Twenty-one critically ill patients were included in this study. Patients received cefepime for the following indications: suspected or documented respiratory infection (18 of 21; 86%), abdominal infection (1 of 21; 5%), combined respiratory and abdominal infection (1 of 21; 5%), or infection of unknown origin (1 of 21; 5%). Microbiological samples taken before cefepime treatment identified one or more pathogens in 16 of 21 (76%) patients: Klebsiella spp. (n = 8), Escherichia coli (n = 6), Citrobacter spp. (n = 2), Proteus mirabilis (n = 1), Pseudomonas aeruginosa (n = 1), Morganella morganii (n = 1), Enterobacter cloacae (n = 1), Staphylococcus aureus (n = 1), and Haemophilus influenzae (n = 1). MIC values for cefepime ranged from ≤1 mg liter−1 to 4 mg liter−1 (75th percentile, ≤1 mg liter−1). Table 1 shows the clinical characteristics of the study patients.
Clinical characteristics of study patients (n = 21)
The median treatment duration with TCI was 4.0 days (interquartile range [IQR], 2.0 to 5.0 days), and the daily cefepime doses were 1.8 g (IQR, 1.6 to 2.5 g) at day 1, 1.3 g (IQR, 1.1 to 2.2 g) at day 2, 1.3 g (IQR, 1.1 to 2.0 g) at day 3, and 1.3 g (IQR, 1.1 to 1.9 g) at day 4. During treatment, a median of 10 blood samples were taken per patient (IQR, 9 to 11), leading to a total of 201 samples. The median measured cefepime plasma concentration was 19.2 mg liter−1, with an interquartile range of 15.3 to 23.3 mg liter−1 (the mean and standard deviation [SD] were 19.5 and 6.36 mg liter−1, respectively). The percentages of measured concentrations within ±10, 20, 30, 40, and 50% of the 16-mg liter−1 target were 20.7, 36.2, 50.3, 66.3, and 77.7%, respectively. Figure 1 shows the measured cefepime concentrations and the predicted concentrations according to the TCI system. The average performance metrics (Varvel criteria) in this patient cohort were as follows: the median absolute performance error (MdAPE) was 28.7%, the median performance error (MdPE) was 20.3%, the wobble was 12.2%, and the divergence was −0.13% h−1. As shown in Fig. 1, performance varies with MdAPEs on an individual basis, ranging between 4.1% and 64.2%. Similar variability was found for the other performance metrics: the MdPE range was −25.6% to 64.2%, the wobble range was 2.12% to 30.3%, and the divergence range was −4.43% h−1 to 0.68% h−1.
Measured cefepime concentrations (black dots) with a nonparametric smoother (blue line) and a target window of 16 mg/liter for the 21 included patients. The black line represents expected plasma concentrations based on the TCI model. The median absolute performance error (MdAPE) is presented for each patient.
By combining the data from this study with the data from a study previously reported by our group (18), we were able to improve the PopPK model for cefepime. The following modifications led to a significant improvement in the goodness of fit: (i) the implementation of the estimated creatinine clearance (eCrCL) as the time-varying covariate on renal clearance (CLrenal) (ΔOFV [objective function value], −75.3) and (ii) the addition of between-subject variability (BSV) on the nonrenal CL (CLother) (ΔOFV, −54.0). Finally, we made two modifications that slightly worsened the goodness of fit: a power parameterization for the eCrCL effect on CL instead of the original linear relationship (ΔOFV, +2.2) and scaling of all PK parameters with body weight according to allometric theory (ΔOFV, +2.3) (19). The former was added to the model to avoid the prediction of negative CLrenal at very low eCrCL values, whereas the latter was included to ascertain the sensible behavior of a TCI system based on this model when used in patients with a body weight outside the range evaluated in this analysis (50 to 120 kg). None of the other covariates tested in the model (age, plasma albumin levels, and C-reactive protein [CRP] levels) were found to be significant. Parameter estimates and associated relative standard errors for the final model are shown in Table 2. The covariate structure for the final model (for a nondialysis patient) is shown in equations 1 to 4. Goodness-of-fit plots for the final PopPK model are provided in Fig. S1 in the supplemental material.
Parameter estimates and associated relative standard errors for the final population PK model derived from simultaneously fitting the data from our previous study and the data from this studya
DISCUSSION
In this study, we describe for the first time the use of TCI for the administration of antibiotics in critically ill patients. PK-pharmacodynamic (PD)-optimized dosing regimens and target attainment are pivotal for effective antimicrobial treatment (4, 5). As a result, different approaches to personalized antibiotic dosing have been attempted (14, 20–23). TCI systems accomplish this individualization via embedded PopPK models and might therefore become a convenient bedside alternative to other approaches. Our prototype TCI system delivers 50.3% of measured cefepime concentrations within ±30% around the target value of 16 mg liter−1. MdPE and MdAPE in this study were 20.3% and 28.7%, respectively. This performance is in line with the performance of current PK models used in TCI pumps in anesthesia (24).
Cefepime was selected as the study drug because it is widely used as a broad-spectrum antibiotic in ICU patients, and individualized TCI dosing has a potential benefit given the relatively small therapeutic-toxic window, compared to other beta-lactam antibiotics. It is important to note that there are no clinically validated target cefepime concentrations for continuous infusion. We chose a target (total) cefepime concentration of 16 mg liter−1 for all patients in our study, which is a compromise between potential toxicity and achieving adequate PK-PD targets. The chosen target concentration is well below the recently advocated threshold for cefepime toxicity of 35 mg liter−1 (6) and is sufficient to achieve free drug above the EUCAST clinical susceptibility breakpoint for the suspected pathogens (e.g., MIC = 1 mg liter−1 for Enterobacterales, and MIC = 8 mg liter−1 for Pseudomonas spp.) (http://www.eucast.org). The target resembles the clinical use of cefepime when microbiology results are absent, e.g., when used empirically or when cultures remain negative throughout the treatment period (25, 26). In these situations, population-level assumptions are made about the most likely organism causing the infection and the distribution of MICs in this population. To achieve true individualization of antibiotic therapy, it might also be necessary to individualize the targeted PK-PD index (i.e., more aggressive PK-PD targets such as time that free drug concentrations are above 2.1 times the MIC [fT>2.1×MIC] [27] or T>4.3×MIC [28]) and to account for the susceptibility of the infecting pathogen (once isolated). TCI systems facilitate the use of a patient-tailored target by reducing the complex dose-concentration relationship via the embedded PopPK models to the selection of an appropriate plasma concentration target. In our opinion, this practicable flexibility could drive the widespread implementation of model-informed precision dosing for antibiotics in the ICU. The use of TCI is not limited to cefepime, but the concept could also be applied to administering any drug that can be given as a continuous infusion.
The additional PK data from this study enabled us to update the PopPK model used in our prototype TCI system. From the pooled data analysis, V1 (volume of distribution of the central compartment) was estimated to be 10.7 liters and not 18.3 liters, as reported previously by our group (18). As a result, loading doses administered by the current version of the TCI system are too high, resulting in an overshoot of the target in the first hour of treatment (as shown in Fig. 1). Furthermore, our analysis indicated that within-individual changes in cefepime clearance are (partly) explained by temporal changes in eCrCL. We hypothesize that an updated version of the TCI system based on the new PopPK model and with eCrCL as a control variable to accommodate within-subject variability in CL will perform better than the system evaluated in this study.
The theoretical lower limit for the performance of this new system depends on the magnitude of the unknown BSV in the PopPK model. When targeting a steady-state plasma concentration and assuming that the PopPK model in the TCI system is unbiased, target attainment is limited by the BSV in CL. In our model, CL consists of CLrenal, with a BSV of 24.6%, and CLother, with a BSV of 69.4%. Consequently, when targeting 16 mg liter−1, 95% of patients are expected to reach a steady-state concentration of between 9.16 and 24.6 mg liter−1 (based on simulations for a population with an average eCrCL rate of 60 ml min−1). This translates to an MdAPE of 21.5%, which is, as expected, lower than the MdAPE reported in this study (28.7%). This shows that it is possible to improve the performance of the current TCI system by updating the embedded PopPK model.
Another useful approach for further refining the accuracy of the system is to use model-based feedback control based on Bayesian forecasting of PK parameters. Open-loop TCI systems (or adaptive TCI systems) (29) where feedback from TDM is used as a control variable in the TCI system are interesting in that respect. Neely et al. (20), Matthews et al. (22), and Pea et al. (23) have shown for aminoglycosides and vancomycin that TDM and Bayesian forecasting of PK parameters result in improved dosing accuracy over conventional dosing strategies. Hence, a TCI system based on the same principles might be advantageous when higher accuracy is needed. The lower limit for the performance of such a system is not dependent on the BSV in the PK but is governed by the residual variability of the PopPK model, which incorporates both inaccuracy in the drug assay and model misspecification. For the updated model, this would result in an MdAPE of 12.8%. Nevertheless, timely availability of appropriate antimicrobial assays could be problematic, as TDM programs for cefepime or other beta-lactam antibiotics are not yet widespread. To this end, biosensor technology could offer an alternative by providing real-time monitoring of antimicrobials in a minimally invasive fashion (30).
There are some limitations to the research presented here. The first limitation is the small number of patients examined and the fact that all patients originated from only one ICU site. Although patient inclusion was not restricted to any medical condition and all patients receiving cefepime with an estimated glomerular filtration rate (eGFR) of >15 ml/min were eligible, extrapolation of the results to a specific subgroup of patients may not be appropriate. For instance, only a few patients with augmented renal clearance were included. Second, the model by Jonckheere et al. (18) uses only eCrCL to individualize cefepime dosing. A more sophisticated PopPK model, also including patient covariates on the volume of distribution, would have likely resulted in better treatment individualization and potentially better performance. Finally, the TCI performance might be overestimated because the PopPK model that was integrated into the TCI was developed in the same ICU.
In conclusion, novel systems are urgently required to individualize antimicrobial therapy, to address the wide variations in PK currently observed across a range of patient populations, and to minimize the occurrence of suboptimal dosing. We demonstrate that cefepime TCI is able to deliver antibiotic concentrations within the expected range around the targeted plasma concentrations in a cohort of critically ill ICU patients. In our opinion, TCI offers exciting possibilities for the individualization of antibiotic treatment in ICU patients and could drive the widespread implementation of model-informed precision dosing in this vulnerable patient population. Further research is needed to confirm that target attainment is superior and to demonstrate increased clinical efficacy in terms of clinical outcomes. The role of TDM in an adaptive TCI approach also requires further investigation.
MATERIALS AND METHODS
Patient inclusion and research ethics.Patients requiring cefepime according to local treatment protocols were included between May 2016 and August 2017. Patients with an eGFR (according to the chronic kidney disease-epidemiology collaboration [CKD-EPI] formula) of less than 15 ml/min and patients who were on hemodialysis were excluded. This trial was conducted at the intensive care department of the OLV Hospital, Aalst, Belgium, in accordance with the Declaration of Helsinki and in compliance with good clinical practice and applicable regulatory requirements. Ethical approval was obtained from the institutional review board of the hospital (Belgium registration number B126201626975). The study was registered in the ClinicalTrials.gov database (identifier NCT02688582) and was monitored by an independent quality specialist.
Drug administration.Patients received cefepime i.v. using a TCI system based on a PopPK model developed previously by Jonckheere et al. (18). In this model, the estimated creatinine clearance (eCrCL) based on the Cockcroft-Gault formula measured on the day of inclusion was used as the only input variable, and a cefepime blood concentration of 16 mg/liter was targeted. There were no adaptations based on changes in eCrCL or measured cefepime concentrations during treatment. Cefepime (20 mg/ml; Fresenius Kabi, USA) was administered by a syringe pump (Orchestra module DPS; Fresenius Kabi, USA) controlled by RUGLOOPII software (Demed, Temse, Belgium) on a personal computer. The maximum infusion rate was set to 4 g of cefepime per h.
Descriptive statistics.The administered daily cefepime dose was extracted from case report forms or RUGLOOPII files. CRP measurements were summarized according to 24-h intervals. Measurements up to 24 h before inclusion in the study were grouped as baseline measurements. Daily doses of cefepime and CRP levels were analyzed for the first 4 days of therapy only; afterward, the number of patients treated was too low to calculate meaningful summary measures. Length of stay in the ICU/hospital and mortality are competing risks (i.e., very sick patients who die would have likely had a very long stay in the ICU/hospital); hence, the length of stay was calculated by replacing the length of stay for patients who died by the maximum length of stay in that patient cohort (31). The presence of neurotoxicity was based on clinical assessment.
Arterial blood and urine sampling and laboratory procedures.Arterial blood was sampled 0.5, 1, 3, 6, 12, 24, 36, 48, 72, 96, and 120 h after the start of the infusion. The exact timing of obtaining blood samples was recorded in the case report form. Samples were collected in lithium heparin tubes, transported immediately to the laboratory, and centrifuged at 1,000 × g for 5 min at 4°C. Subsequently, plasma samples were stored at below −70°C until analysis. Urine was collected daily from a urinary catheter over a 12-h interval. The quantification of cefepime levels was based on a validated solid-phase extraction–liquid chromatography–electrospray–tandem mass spectrometry method (32) using a 13C12-2H3-labeled cefepime isotope as the internal standard (AlsaChim, Illkirch, France). The range of the analytical method was 0.15 mg liter−1 to 15 mg liter−1, with an average bias and imprecision of +5.9% and 8.6 CV% (percent coefficient of variation). Plasma samples were diluted 1/5 in blank human plasma, whereas urine samples were diluted 1/50 in blank human plasma prior to analysis. All samples were measured in duplicate. Microbiological samples were taken according to the standard of care and analyzed using standard culture procedures. Identification was performed using matrix-assisted laser desorption ionization–time of flight mass spectrometry (MALDI-TOF MS) (Bruker Daltonik GmbH, Germany), and antimicrobial susceptibility testing was performed using the Phoenix system (Becton, Dickinson, USA) according to the manufacturer’s instructions.
Calculation of predictive performance.In line with studies on the performance of TCI systems in anesthesia, we used the “Varvel criteria” to evaluate the performance of our TCI system (33). For this, the performance error (PE) is calculated for all samples (j) for the different patients (i) according to the following equation:
Update of the previously reported population pharmacokinetic model.The plasma and urine cefepime concentration-versus-time data were fitted using the first-order conditional estimation with interaction (FOCE-I) algorithm in NONMEM (version 7.3; GloboMax LLC, Hanover, MD, USA). The “tidyverse” package (version 1.1.1.; H. Wickham, 2017) in R (R Foundation for Statistical Computing, Vienna, Austria) was used to graphically assess the goodness of fit. As a starting point, the model previously reported by our group (18), which was used as the PopPK model in the presented TCI system, was fitted to the combined data set (PK data from the pilot study [18] and additional PK data from this TCI study). Modifications to the model were accepted if they resulted in a decrease in the OFV. A decrease in the OFV was judged statistically significant if the inclusion of an additional parameter decreased the OFV by more than 3.84 points.
ACKNOWLEDGMENTS
Michel M. R. F. Struys and Ghent University have a financial interest in RUGLOOPII, a software program for target-controlled infusion. His research group/department received grants and funding from The Medicines Company (Parsippany, NJ, USA), Masimo (Irvine, CA, USA), Fresenius (Bad Homburg, Germany), Acacia Design (Maastricht, The Netherlands), Medtronic (Dublin, Ireland), Paion (Aachen, Germany), and PRA (Groningen, The Netherlands) and honoraria from The Medicines Company (Parsippany, NJ, USA), Masimo (Irvine, CA, USA), Fresenius (Bad Homburg, Germany), Baxter (Deerfield, IL, USA), Medtronic (Dublin, Ireland), Becton, Dickinson (San Diego, CA, USA), and Demed Medical (Temse, Belgium). All other authors have no conflicts of interest to declare.
This study was supported by internal funding.
FOOTNOTES
- Received 3 August 2019.
- Returned for modification 3 October 2019.
- Accepted 28 October 2019.
- Accepted manuscript posted online 4 November 2019.
Supplemental material is available online only.
- Copyright © 2019 American Society for Microbiology.
