Development of a predictive platform for target concentration intervention of aminoglycosides

Abstract

Aminoglycosides, including gentamicin, are potent bactericidal antibiotics that are widely utilized for the treatment of severe gram-negative bacterial infections. These drugs demonstrate a concentration-dependent bactericidal effect, whereby higher concentrations result in more rapid bacterial elimination. Despite their efficacy, determining optimal dosing regimens is challenging due to significant inter-individual pharmacokinetic (PK) variability and adverse drug reactions, such as nephrotoxicity and ototoxicity. This study developed a simplified, user-friendly PK-predictive system for target concentration intervention of gentamicin. A semi-automatic platform was developed for the generation of patient-specific NONMEM PK datasets using Shiny and the estimation of individual PK parameters using maximum a posteriori Bayesian estimation. The system provides a dosing assessment based on Bayesian PK predictions, with results presented in detailed reports generated via LaTeX. The predictability of the PK model was confirmed via external validation using clinical PK data from 22 patients who underwent therapeutic drug monitoring consultations at the Asan Medical Center. In the external validation the mean prediction error was −0.19 mg/L (95% confidence interval, −0.88 to 0.41) and the root mean squared prediction error was 2.85 mg/L, and 86.4% of observed concentrations fell within the model-predicted 95% prediction interval on the individual prediction check plots. These results support the applicability of the model to this population, within the limits of precision reported, for dose individualisation. The platform architecture may also be adaptable to other drugs with a narrow therapeutic index.

INTRODUCTION

Aminoglycosides, such as gentamicin, are highly effective bactericidal antibiotics widely used to treat serious gram-negative bacterial infections. Aminoglycoside antibiotics are bactericidal and demonstrate concentration-dependent bacterial killing mechanisms. Such antibiotics typically eliminate bacteria at a faster rate with higher concentrations. Furthermore, aminoglycosides exhibit a concentration-dependent post-antibiotic effect, which involves sustained bacterial killing despite serum concentrations declining below the minimum inhibitory concentration (MIC). Given the concentration-dependent nature of the post-antibiotic effect for aminoglycosides, higher drug concentrations result in a prolonged post-antibiotic effect. The standard approach for administering aminoglycoside antibiotics involves administering multiple doses throughout the day, typically every eight hours, in patients with normal renal function [1].

To exploit the concentration-dependent bactericidal and post-antibiotic effects, an extended-interval aminoglycoside administration regimen (typically the total daily dose administered once per day) represents an alternative dosing option [2]. This dosing method has now become the predominant mode of administration, with the exception of endocarditis treatment, where aminoglycosides are employed for their synergistic effect with penicillin or vancomycin.

Aminoglycoside antibiotics are administered as short-term (0.5–1 hour) infusions. If a 1-hour infusion is employed, the maximum end-of-infusion “peak” concentrations are ascertained upon completion of the infusion. Conversely, when a 0.5-hour infusion is utilized, serum concentrations exhibit a distribution phase, indicating that the drug is present in the blood and tissues, and an equilibrium has not been reached. Consequently, a 0.5-hour waiting period is required for the drug to reach distribution prior to measuring peak concentrations when a 0.5-hour infusion is used.

Therapeutic peak concentrations for gentamicin are generally 5–10 µg/mL for gram-negative infections. Infection sites with more susceptible bacteria, such as intra-abdominal infections, usually can be treated with steady-state peak concentration at the lower end of this range (typically 5–7 µg/mL). Infection sites that are difficult to penetrate and comprise bacteria associated with higher MIC values, such as in cases involving Pseudomonas pneumonia, usually require steady-state peak concentrations at the higher end of the range (typically 8–10 μg/mL). When gentamicin, tobramycin, or netilmicin are used synergistically with penicillins or other antibiotics for the treatment of gram-positive infections, such as infective endocarditis, steady-state peak concentrations of 3–4 μg/mL are often adequate [3].

Gentamicin dosing requires careful monitoring of peak and trough levels to ensure efficacy and safety. Based on various references, the target peak concentration for gentamicin is 5.0–10.0 mg/L, whereas the trough should be 1.0–2.0 mg/L for multiple daily dosing (MDD) regimens [4]. Typically, an 80 mg dose is administered thrice daily, with trough levels checked after 24 hours (target < 2 mg/L) and peak levels one hour after dose administration (target 5–10 mg/L) [5]. For infections such as those caused by Pseudomonas spp., higher doses may be needed to achieve optimal peak concentrations. Gentamicin once-daily (OD) dosing regimens suggest 5.1–7 mg/kg q24h with peak levels of 15–25 μg/mL and trough levels below 1 μg/mL [6]. For MDD, 2 mg/kg q8h or 1.7 mg/kg q8h is recommended, with peak levels of 5–10 μg/mL and trough levels of 1–2 μg/mL [6]. Furthermore, for Gram-negative organisms, the targets conventionally used for gentamicin are an area under the curve (AUC) 0–24/MIC of 80 to 100 mg·h/L (assuming MIC ≤ 1 mg/L) and a Cmin (trough) of less than 1 mg/L [7].

However, the therapeutic application of aminoglycosides is complicated by significant inter-individual variability in pharmacokinetic (PK), which presents a challenge in determining the optimal dosing regimen to maximize efficacy while minimizing the risk of nephrotoxicity and ototoxicity. Elevated trough and peak serum gentamicin concentrations consistently result in an increased AUC, thereby emphasizing the necessity for dose optimization. Trough concentrations offer the most reliable indication of equilibrium between blood and tissue distribution. Without serial monitoring of gentamicin serum concentrations at trough times during therapy, identification of patients at risk of developing gentamicin-induced nephrotoxicity is challenging. Exceeding peak steady-state concentrations of 12–14 μg/mL for gentamicin, tobramycin, or netilmicin or 35–40 μg/mL for amikacin when using conventional dosing leads to an increased risk of ototoxicity [7,8].

One key benefit of this approach is that it allows for consistent dosing recommendations to be made when multiple providers are involved in a therapeutic drug monitoring (TDM) program. There is a need for methods to individualize dosing regimens to the specific PK profiles of individual patients. TDM has historically entailed making empirical dose adjustments when a drug concentration is identified as being outside the established “therapeutic range” [9]. In this study, we have employed target concentration intervention (TCI) as a particularly advantageous method when a direct evaluation of the desired therapeutic outcome is difficult. By interpreting both concentration and effect, TCI enables clinicians to select the optimal dose for the desired therapeutic outcome. In essence, TCI entails the active adjustment of treatment based on the concentration of the administered drug and its relationship to the observed effect [10]. TCI maintains the drug concentration at a target level known to be effective, thereby enhancing therapeutic precision even when group-based dosing does not sufficiently reduce inter-individual variability. Unlike TDM, which involves passive monitoring of drug concentrations in correspondence with a predefined therapeutic range, TCI involves making active adjustments to dosing based on individual patient responses to achieve the desired concentration, providing a more dynamic and individualized approach [11].

Given that simpler dosing methods are effective for patients with stable PK parameters and steady-state drug concentrations, many clinicians utilize computer programs for elucidating more complex scenarios. Such situations include non-steady-state serum concentrations, failure to obtain serum concentrations at the required times to use simpler methods, and unstable PK parameters. In the event that only a limited number of aminoglycoside concentrations are available, a Bayesian PK computer program can be used to calculate a complete patient PK profile, including clearance, volume of distribution, and half-life [12].

We employed a nonlinear regression algorithm that incorporates the components of Bayes’ theorem,[12] which is a reliable statistical technique. Bayesian analysis is fundamentally based on the joint posterior distribution of the parameters, p(θ|Y), which is contingent upon the likelihood, p(Y |θ), and the prior distribution, p(θ), as defined by the Bayes' theorem [equation 1] [13].

$$p\left(\theta \right|Y)=\frac{p\left(Y\right|\theta )p\left(\theta \right)}{p\left(Y\right)}\propto p\left(Y\right|\theta )p\left(\theta \right).$$

Nonlinear regression is a statistical technique that employs an iterative process to determine the optimal PK parameters for a concentration/time dataset. Essentially, the medication schedule and serum concentrations of a patient are entered into a computer program. The computer program comprises pre-programmed PK equations for the drug and the method of administration (such as oral, intravenous, and intravenous infusion). Typically, although a single-compartment model is utilized, some softwares allow the user to select from a range of alternative equations. The computer program calculates the estimated serum concentrations for each actual serum concentration using population estimates based on the demographic information (such as age, weight, gender, and renal function) provided by the user. The computer program then makes the necessary adjustments to the kinetic parameters, resulting in a new set of estimated serum concentrations. The computer program retains the PK parameters that produced the estimated serum concentrations most closely aligned with the actual values. This process is repeated until a set of PK parameters is generated that produces estimated serum concentrations with the greatest statistical similarity to the actual serum concentrations. The calculated PK parameters can then be used to create an optimized dosing schedule for the patient. The computer algorithm employs Bayes' theorem to balance the results of the calculation between values based solely on serum drug concentrations and values based solely on population parameters [12,14].

Model-based methods, particularly those that employ nonlinear mixed-effects modeling (NONMEM^®), offer a promising approach for optimizing aminoglycoside therapy. Nevertheless, the complexity of NONMEM^® and the concomitant data management demands may present challenges to its routine implementation in clinical settings. To ensure the clinical utility of these models, rigorous external validation is essential. This involves evaluating the predictive performance of the model using independent datasets, confirming its ability to generalize across different patient populations. Such validation enhances the robustness and reliability of the model, rendering it a valuable tool for personalized therapy [15]. While techniques, such as bootstrapping and posterior predictive checks, are beneficial for internal validation, the most rigorous method for validating a model is through an evaluation of its predictive performance on an external dataset [16].

The purpose of this study was to develop a validated semi-automated, user-friendly platform for the individualization of aminoglycoside therapy through TCI. This study intended to predict patient-specific PK parameters using the Bayesian method and streamline the dose optimization process in the clinic. This approach was designed to optimize individual treatment outcomes and mitigate the risk of toxicities associated with these drugs.

METHODS

We aimed to develop a user-friendly predictive platform for a streamlined approach to Bayesian estimation of individual PK parameter estimation and optimal dosing recommendations of intravenous gentamicin on the request of TCI consultation in the clinical setting.

Construction of PK model and external validation

A PK model for gentamicin was implemented in NONMEM^® based on a previous study [17]. The structural model, the covariate relationships and every parameter value listed in Table 1, including the residual error model, were adopted from that publication; no PK parameter was estimated from the present data. The predictability of the PK model was evaluated by external validation where model-predicted concentrations over time were compared with the PK data of 22 patients who had undergone TCI consultation in the department of clinical pharmacology and therapeutics at Asan Medical Center. Predictive performance was assessed from the prediction error of each observed concentration, defined as the predicted minus the observed concentration. For each patient 1,000 replicates were simulated using only that patient's dosing history and covariates, without any estimation against the patient's own measured concentrations, so that the predictions were a priori. The predicted concentration was taken as the median of the simulated individual model predictions, that is without the residual error term, while the prediction intervals were formed from the full predictive distribution including residual error. Prediction errors were summarised as the mean prediction error (ME) as a measure of bias and the root mean squared prediction error (RMSE) as a measure of imprecision, together with their relative counterparts expressed as percentages (%MPE and %RMSE). The proportion of observed concentrations falling within the model-predicted 80%, 90% and 95% prediction intervals was also determined. Two-sided 95% confidence intervals were obtained by a subject-level bootstrap with 2,000 resamples, which accounts for the repeated sampling of individual patients.

Table 1. Parameters of the pharmacokinetic model for gentamicin adopted in this study, taken from reference 17, in which the model was developed in 939 hospitalized adults receiving once-daily gentamicin.

Parameters	Estimates	RSEs (%)	95% confidence interval
TVCL	0.047	7.5	0.040–0.054
TVVd	0.280	7.6	0.24–0.32
TVK12	0.092	12.1	0.070–0.11
TVK21	0.071	48.6	0.0033–0.14
IIVCL	0.084	9.9	26.6–32.4
IIVVC	0.003	350	-
IIVK12	0.398	43.0	27.6–94.8
IIVK21	0.342	49.0	12.6–89.4
Residual error (proportional; additive)	23.7%; 0.23 mg/L

Units: TVCL, L/h per mL/min of creatinine clearance; TVVd, L/kg; TVK12 and TVK21, h⁻¹. For the four inter-individual variability terms the Estimates column gives the variance (ω²) entered in $OMEGA, derived from the coefficient of variation reported in reference 17 (CL 29.6%, V1 5.8%, K12 69.9%, K21 63.9%) as ω² = ln(CV² + 1); the RSE and 95% confidence interval columns are reproduced from reference 17 and refer to that coefficient of variation, not to ω². The residual error is the combined model reported in reference 17, a proportional component of 23.7% together with an additive component of 0.23 mg/L; the two are entered in $SIGMA as variances of 0.0562 and 0.0529 respectively.

Model, as reported in reference 17: CL = (THETA(1) * CL_cr)*EXP(ETA(1)), VC = (THETA(2) * WT) * EXP(ETA(2)), K₁₂ = THETA(3) * EXP(ETA(3)), K₂₁ = THETA(4) * EXP(ETA(4))

RSE, relative standard error; TV, typical value; IIV, inter-individual variability; CV, coefficient of variation.

The residual error was implemented as reference 17 specifies it, as a combined model with a proportional component of 23.7% and an additive component of 0.23 mg/L. A combined form is necessary over the concentration range encountered in TDM of gentamicin, which spans roughly 0.3 to 30 mg/L: a purely additive error would be too large relative to the trough concentrations, where it would render the measurement uninformative in the Bayesian step, and too small relative to the peak concentrations.

The study protocol for the use of the TCI PK data for the external validation was also approved by the Asan Medical Center Institutional Review Board (approval number: [2024-1174, 2025-1559]). The requirement for written informed consent was waived by the Board because the study was a retrospective analysis of clinical and laboratory data collected during routine care.

Development of predictive platform

The predictive platform was developed using R with Shiny package, NONMEM^®, and LaTeX. Individual patient dosing and PK sampling histories were designed for user-friendly input and compilation into a NONMEM-compatible data format, facilitated by an automated process involving a Shiny application. Once the data preparation was complete, Bayesian PK predictions were performed to estimate individual patient-specific PK parameters using the individual-specific PK data-sets. The process was further automated by NONMEM^®, summarizing the results using R, and generating comprehensive reports using LaTeX to facilitate clinical decision making (Fig. 1). This prediction method assisted in determining the optimal dosing regimen for each patient.

Figure 1. Overall schematic of the process from data entry to final output.

TCI, target concentration intervention.

Between-subject variability was described by an exponential (log-normal) random-effects model. For clearance, the value in the ith individual was

CL_i = TVCL × exp(η_i), where η_i ~ N(0, ω²)

where TVCL is the typical value of clearance, η_i is the random effect of that individual and ω² is the corresponding diagonal element of the OMEGA matrix, that is OMEGA(1,1). The coefficient of variation implied by this model is

CV = √(exp(ω²) − 1)

Because the published estimates of between-subject variability were reported as coefficients of variation rather than as variances, the OMEGA entries used in the control stream were obtained by inverting this relationship,

ω² = ln(CV² + 1)

and the same relationship was applied to each of the random effects in the model. The resulting variances are those reported in Table 1.

Concretely, reference 17 reports between-subject variability as a coefficient of variation for each parameter (29.6% for CL, 5.8% for V1, 69.9% for K12 and 63.9% for K21). Each of these was converted in the direction CV to variance using the expression above, giving the values entered in $OMEGA and listed in Table 1: 0.0840, 0.0034, 0.3978 and 0.3424 respectively. No other transformation was applied.

The variability on V1 warrants comment. A coefficient of variation of 5.8% is small in absolute terms, and its relative standard error in the source publication is 350% with no confidence interval reported. In maximum a posteriori estimation this tight prior means that an individual’s V1 is held close to the covariate-predicted value of 0.28 x body weight, so measured concentrations inform clearance far more than they inform volume. Individual volume estimates from the platform should therefore be interpreted with that constraint in mind.

For reporting, LaTeX was used to automate the generation of results, with R code handling NONMEM^® runs and data cleanup. The reports included organized tables and figures, summarizing PK parameters.

Importantly, all processes, from data handling to report generation, were coordinated using the master code to enhance consistency and efficiency.

The platform does not apply an automated dose-optimisation algorithm. It predicts the exposure that a candidate regimen would produce and presents that prediction against the target range, leaving the choice of regimen to the prescribing clinician. For each patient the dataset is extended with 24 further doses at each of the 8 hours, 12 hours and 24 hours intervals so that steady state is reached, and the profile is projected to seven days beyond the last recorded dose. For each candidate interval the report gives the individual clearance and volume of distribution together with the steady-state area under the concentration-time curve over 24 hours (AUC24), the peak concentration and the pre-dose trough concentration. Peak and trough are each reported twice: a theoretical pair, being the maximum and the pre-dose concentration of the steady-state interval, and a clinical pair, taken at the times at which samples are drawn in practice, 2 hours after the start of the infusion and 30 minutes before the next dose.

The targets against which these predictions are read are those in routine use for gentamicin in Gram-negative infection. For extended-interval dosing the platform targets a peak concentration of 15–20 mg/L, a trough below 1 mg/L and an AUC24 of 70–120 mg·h/L; for multiple-daily dosing, a peak of 5–10 mg/L and a trough below 2 mg/L [3,6,7,18]. An organism-specific MIC is not measured by the platform, and an MIC of 1 mg/L is assumed throughout, so that the conventional targets of a Cmax/MIC ratio of 8–10 and an AUC24/MIC of 80–100 correspond to the absolute concentrations above [7,18]. The EUCAST clinical breakpoint for gentamicin against Enterobacterales is 2 mg/L, so for an organism at the breakpoint the same Cmax/MIC ratio would require approximately twice the peak concentration; where the MIC of the isolate is known it should be used in place of the assumed value [18].

Software availability

The predictive platform developed in this study is based on R Shiny, NONMEM, and LaTeX. To improve transparency and reproducibility, the corresponding source code and related materials are available at https://kyun-gran.github.io/tci-web/. The demonstration site is statically hosted and therefore covers only the data-preparation step: a user can enter a dosing and sampling history and download the resulting NONMEM-ready dataset. The buttons that invoke NONMEM and that generate the LaTeX report are present but inoperative on the demonstration site, because both require a local NONMEM installation and a LaTeX distribution; those steps must be run in the user’s own environment using the source code provided. A worked example of the generated report is provided in Supplementary Data 1, 2, 3. The analyses reported here used NONMEM 7.5.0 for individual parameter estimation and NONMEM 7.6.0 for the simulations, with R 4.5.3 and the shiny, shinydashboard, dplyr, DT and zoo packages.

RESULTS

Construction of PK model and external validation

The final PK model for gentamicin was implemented in NONMEM [17]. PK parameters used for the construction model presented in Table 1. The inter-individual variability values were calculated in accordance with the methodology described in the Methods section. A total of 22 adult patients received intravenous gentamicin during the course of their hospitalization. The characteristics of the patients are presented in Table 2. A total of 170 gentamicin plasma concentrations were measured, of which 169 from 22 patients were above the limit of quantification and were used for the validation; the single value reported as below the limit of quantification was excluded. The observed concentrations had a median of 2.2 mg/L (range 0.2 to 34.0 mg/L). The individual prediction check showed that the 95% prediction interval generated from each patient’s own dosing history and covariates contained the majority of that patient’s observed concentrations (Fig. 2). The ME was −0.19 mg/L (95% CI, −0.88 to 0.41) and the mean percentage prediction error was +14.8% (95% CI, −2.6 to 35.4); both confidence intervals include zero, indicating no appreciable systematic over- or under-prediction on either the absolute or the relative scale. The RMSE was 2.85 mg/L (95% CI, 1.82 to 4.13), equivalent to a relative RMSE of 81.2% of the mean observed concentration. Imprecision was larger in absolute terms for peak samples (RMSE 3.86 mg/L, n = 87) and larger in relative terms for trough samples (relative RMSE 85.8%, n = 81), as expected from the combined residual error model. Observed concentrations fell within the model-predicted 80%, 90% and 95% prediction intervals in 73.4%, 81.7% and 86.4% of cases respectively, so that 22 of the 169 observations (13.6%) lay outside the 95% prediction interval. The model therefore described the central tendency of the external data without appreciable bias, while the prediction intervals were consistently narrower than the observed scatter, indicating that the adopted model somewhat under-estimates the total variability of this population. Most of the discrepancy arose at the peak samples, of which 18 of 87 (20.7%) lay outside the 95% interval, against 5 of 81 (6.2%) of the trough samples; this residual imprecision, which is expected with the sparse and irregular sampling of routine clinical care and with a prior developed under a different dosing schedule, should be taken into account when individual predicted concentrations are interpreted. Based on this validation, a Bayesian approach employing the automated Shiny application was implemented, exhibiting a strong data fit.

Table 2. Demographic information of all subjects (n = 22) included in the external validation of gentamicin pharmacokinetic model.

Characteristics	Values
No. of patients	22
Age (yr)	61.00 ± 13.24
Height (cm)	162.00 ± 9.12
Body weight (kg)	62.00 ± 17.31
Sex (male/female)	20/3
Trough (mg/L)	1.00 ± 0.61 (n = 22)
Peak (mg/L)	6.00 ± 4.22 (n = 22)
Creatinine (mg/dL)	0.77 ± 0.25
Amount (mg)	127.00 ± 72.99

Values are presented as number or mean ± standard deviation.

Figure 2. (A) External validation, gentamicin PK model. (B) Individual prediction check -full time (left). (B) Individual prediction check -~168 hours (right). This figure illustrates the Individual prediction results for the gentamicin results for the gentamicin population pharmacokinetic model, validated with observed PK data from 22 patients over 168 hours. Observed plasma gentamicin concentrations (depicted as circles) were compared with the concentration percentiles predicted by the model. The solid line represents the model-predicted median, whereas the simulated 95% prediction interval is shaded. The Individual prediction demonstrates the ability of the model to capture the central tendency and variability of the observed data, supporting its reliability for clinical applications.

PK, pharmacokinetic.

With external validation data providing confidence in the model developed based on the literature, we proceeded further by employing an automated Shiny application to create NONMEM^®-compatible datasets. This process enabled the straightforward formatting of individual patient-specific dosing and PK histories for Bayesian PK parameter estimation. NONMEM^® was then employed to generate predictions, which were summarized in a concise report using LaTeX. This system supports clinical decision-making by facilitating optimal dosing recommendations for intravenous gentamicin in patients requiring TCI consultation at Asan Medical Center. The predictive performance reported above supports the use of the model for this purpose, within the limits of precision quantified in the external validation.

Predictive platform

To execute the Shiny application, the following packages were utilized: shiny, shinydashboard, dplyr, DT, and zoo. The code utilized in this study is presented in the Appendix for reference. To execute the primary code, the user was first required to input the code, which then generated a Shiny window (Fig. 3). A NONMEM^® dataset was generated automatically for each patient. This dataset comprised information specific to the patient, including dosing history, time of gentamicin concentration measurements, serum creatinine (SCr) level, age, weight, and gender. This information was entered via the input window using the Shiny application. The left side of the window was used to input data, including the following variables: ID, TIME, DV, dosing time, amount, duration, SCr Time, and demographics (age, sex, weight, height). For user inputs, the data are displayed in Fig. 4. Notably, the SCr column was based on the last observation carried forward (LOCF). Therefore, if the user inputs data, it will be processed in accordance with the LOCF rule.

Figure 3. View of the initial window as observed upon launching the application.

TCI, target concentration intervention.

Figure 4. Upon data input, the output was displayed as shown. In order to enhance the visual representation of the LaTeX results, an additional row was introduced at the specified time interval, divided by 0.5.

TCI, target concentration intervention.

Typically, gentamicin is administered in three divided doses (8H, 12H, or 24H), with each dose given at see Fig. 5. To facilitate convenient dosing, we created tabs indicating the appropriate dosage for each interval. From the last dose in the entered data, 24 additional doses were automatically added to the dataset in each of the 8H, 12H, or 24H tabs in Shiny in order to predict the steady-state PK of gentamicin. In some cases, consultations for individualized optimal dosing of gentamicin were requested based on gentamicin concentrations measured early after the onset of drug administration, before the PK had reached a steady state. For instance, patients may be referred for peak/trough concentration measurement only around the second medication. Given that the half-life of gentamicin is 2–3 hours in patients with normal renal function and the potential for an elongated half-life is reduced in patients with reduced renal function, these additional doses enable more accurate predictions. Once the data were entered, the .csv file could be downloaded using the Download button located at the bottom of the window. A new folder was generated in the directory specified at the time the code was initiated, and the file was downloaded to that directory with the DM+ID filename.

Figure 5. Automatically added dosing adjustments at 8-hour, 12-hour, and 24-hour intervals.

At this point, the second line of code in the master code may be executed, which initiated NONMEM^®. Once the NONMEM^® completed its processing, the results were documented using LaTex, which represents the final line of code in the master code. It is imperative to utilize the NONMEM^® result file with the extension of IPK, which contains the individual PK parameters, to generate the report.

DISCUSSION

The objective of this study was to develop a user-friendly predictive platform for the TCI of gentamicin. The study involved the creation of a semi-automatic system that facilitates Bayesian estimation of individual PK parameters using the sparse concentration data of gentamicin of a referred patient, thereby enabling TCI through a streamlined process.

In this Bayesian prediction, “maxeval = 0” option in NONMEM allows the estimation of individual PK parameters relying on a prior PK model for gentamicin and sparse gentamicin concentration data of the patients without iterative re-evaluation of the population PK parameters of gentamicin. This is a departure from the conventional $PRIOR approach, which involves iterative re-estimation of population PK parameters based on prior PK models. By setting maxeval = 0 in the estimation step in NONMEM, the predictions are derived directly from the prior information, thereby improving the computation speed while maintaining accuracy. This approach offers clinical efficiency, which is especially crucial in settings where timely PK adjustments are necessary, as it circumvents the necessity for full parameter re-optimization. In contrast, the standard $PRIOR method may entail a greater computational demand due to the necessity of multiple evaluations, which could potentially limit its feasibility for rapid clinical decisions [19,20].

PK-PD models for anesthetic drugs are typically created using carefully selected patient populations, often omitting groups like children, the obese, or the elderly, whose body composition or elimination mechanisms can significantly differ from those of lean adults. While separate PK-PD models have been developed for some of these subpopulations, the availability of multiple models for the same drug increases the likelihood of incorrect model selection, which may result in potential inaccuracies in clinical application. One potential solution to these limitations is the integration of historical data from disparate studies into a unified, globally collected database. This database would contain PK-PD information for commonly utilized anesthetic drugs across a diverse range of patient groups. The open TCI initiative exemplifies this approach. By consolidating such datasets, the development of “general purpose” PK-PD models that could be applied more broadly across different patient populations is possible. This could help mitigate issues related to model selection and enhance accuracy in clinical practice [21].

At present, the shiny application is only able to generate NONMEM datasets. However, incorporating the capability to run NONMEM^® simulations and generate LaTeX reports within the same interface would enhance user convenience. Consequently, further development of the Shiny code is necessary to streamline this process.

The semi-automatic nature of this system demonstrates its practical utility in clinical settings. The implementation of Bayesian PK models is frequently challenging due to the complexity of the equations and the computational demands involved. However, this platform provides clinicians with accessible PK predictions, which may increase the use of Bayesian models in routine practice.

Moreover, the incorporation of LaTeX for report generation provides substantial benefits, including professional formatting, automated summaries, and high readability, which facilitate clinical decision-making. By integrating R, NONMEM, and LaTeX, the platform offers a comprehensive solution that enhances the user experience and supports timely, personalized patient care.

The prior used here was developed in a different setting from the one in which the platform is applied. Reference 17 studied 939 hospitalized adults receiving OD gentamicin, with a mean age of 55 years, a mean weight of 70 kg and a mean creatinine clearance of 92 mL/min. The present cohort was older, lighter and had poorer renal function, and the platform offers 8 hours, 12 hours and 24 hours dosing intervals to match local prescribing. A prior estimated under OD dosing is not guaranteed to transport to shorter intervals: the sampling in the source study was concentrated in the OD profile, so the peripheral compartment and the distribution rate constants are less well identified over the time course relevant to q8h or q12h regimens. Predictions under those regimens should therefore be treated as provisional and confirmed with measured concentrations, and prospective evaluation across dosing schedules is needed.

Peak concentration remains a determinant of outcome for the aminoglycosides. In an earlier series the mortality rate fell from 20.9% among patients whose peak concentrations were subtherapeutic to 2.4% among those in the therapeutic range, and although gentamicin and tobramycin were not analysed separately, both showed an association between peak level and survival. This is the clinical rationale for individualising the dose rather than an outcome of the present work.

This study has several limitations. The external validation cohort was modest, comprising 22 patients and 169 quantifiable concentrations from a single centre. The samples were drawn opportunistically during routine care rather than by a fixed schedule, so the sampling is sparse and irregular and the peak and trough windows vary between patients. The prior was developed in a different population and under a different dosing schedule, as discussed above. Dose recommendations rest on an assumed MIC rather than an organism-specific measurement, so the exposure targets are indicative rather than patient-specific. No clinical outcome was measured: neither efficacy, nephrotoxicity nor ototoxicity was assessed, so the platform's effect on patient outcome is untested. Finally, the evaluation was retrospective; a prospective study comparing platform-guided dosing with current practice is the necessary next step.

The present work is therefore best understood as a feasibility study and an external evaluation of an adopted model, rather than as a demonstration of clinical benefit.

SUPPLEMENTARY MATERIALS

Supplementary Data 1

Gentamicin_PK_model

Supplementary Data 2

Individual prediction check (all times)

Supplementary Data 3

Individual Prediction check (~168 hours)