Evaluating Bayesian Vancomycin Dosage Adjustment in Neonates: A Simulation-Based Virtual TDM Study

Abstract

Background and Objectives

Therapeutic drug monitoring (TDM) of vancomycin, ideally based on the ratio of the daily area under the curve to the minimal inhibitory concentration at steady state (AUC_24,SS/MIC), is hindered by sampling limitations in neonates, so trough concentration (C_trough) is often used as a surrogate. This virtual TDM study aimed to evaluate the performance of Bayesian model-informed precision dosing (MIPD) for vancomycin in neonates.

Methods

Reference pharmacokinetic (PK) parameters and drug concentrations were simulated in NONMEM for 1000 virtual neonates using a published population PK model. Four TDM strategies were compared based on the percentage of patients achieving an AUC_24,SS/MIC between 360 and 540 h. Strategy 1 maintained the initial regimen unchanged; strategy 2 adjusted doses using a standard rule of three to target a C_trough of 10–15 mg/L; strategies 3 and 4 used the MIPD software Tucuxi with a steady-state C_trough and two concentrations after the first dose, respectively, for Bayesian dosage optimization. Individual dosages were adjusted following each strategy, recalculating AUC_24,SS to determine the percentage of target attainment. Iterative adjustments were performed by resampling C_trough. A sub-study evaluated optimal sampling time by comparing Tucuxi-estimated AUC_24,SS with the reference value.

Results

Less than 49% of patients reached the target with strategy 1, leading to frequent vancomycin overexposure. Strategy 2 enabled over 75% of patients to achieve AUC_24,SS/MIC between 360 and 540 h with every-8-h regimens, but only 31% with longer dosing intervals. Strategies 3 and 4 allowed over 74% of patients to reach the target after one TDM cycle and 100% after four cycles, regardless of the initial dosing regimen. A single sample sufficed for accurate AUC_24,SS prediction.

Conclusion

This virtual TDM study indicates that Bayesian MIPD optimizes vancomycin exposure in neonates, while simplifying the sampling strategy. This approach deserves to be integrated and validated in routine neonatal care.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s40262-026-01642-4.

Key Points

By simulating virtual therapeutic drug monitoring (TDM) scenarios, we demonstrate that Bayesian model-informed precision dosing (MIPD) significantly improves target attainment in vancomycin therapy for neonates.

With MIPD software, accurate dosing can be achieved using just one blood sample taken at least 5 h after the first dose, greatly simplifying the sampling protocol.

These findings support the integration of Bayesian dosing tools into routine neonatal care to enhance treatment safety and efficacy.

Introduction

Vancomycin is a glycopeptide antibiotic with a narrow therapeutic index frequently used for the management of severe Gram-positive infections in neonates [1]. Its disposition is characterized by a large interindividual variability (IIV) throughout all ages. Dozens of population pharmacokinetic (popPK) models have already described developmental changes associated with vancomycin disposition in neonates and children [2–5].

The ratio of the daily area under the curve to the minimal inhibitory concentration at steady state (AUC_24,SS/MIC) is considered the best surrogate marker for vancomycin efficacy, with a target of 400–600 h based on adult data, which is assumed to be suitable also for various pediatric age groups as minimal inhibitory concentration (MIC) values of 1 mg/L are often encountered in children [1]. However, for neonates, retrospective studies have suggested lower AUC_24,SS/MIC cut-offs (240–331 h), based on epidemiological MIC values of treated pathogens [6–8]. Furthermore, the suitability of these daily area under the curve at steady state (AUC_24,SS) targets in neonates is further questioned by the markedly lower and highly variable plasma protein binding in this population, which may lead to higher unbound vancomycin concentrations for a given total AUC_24,SS [1, 9].

Even though therapeutic drug monitoring (TDM) of vancomycin is widely adopted, there is a lack of guidance on the practical implementation, including timing of blood sampling and drug measurements, calculation methods, and dose adjustment strategies [10]. Current guidelines increasingly recommend AUC_24,SS-based dosing, which is gradually being implemented in clinical practice [11–13]. However, monitoring AUC_24,SS remains challenging from a practical standpoint, as it requires the drawing of multiple blood samples. Consequently, many centers still rely solely on trough plasma concentrations (C_trough) for dose adjustment, targeting 10–15 mg/L for non-severe, non–Methicillin-resistant Staphylococcus aureus (non-MRSA) infections and 15–20 mg/L for severe infections [14]. However, this pragmatic approach is problematic because studies have shown that C_trough is an imperfect proxy of AUC_24,SS. Thus, dose adjustment based on C_trough alone, even if C_trough is within the target range, may lead to supratherapeutic exposure and potentially nephrotoxicity in some patients or, conversely, to infra-therapeutic exposure and potentially treatment failure [12, 15]. Therefore, vancomycin TDM guidelines have been recently updated to recommend dose adjustment based on AUC_24,SS, using model-informed precision dosing (MIPD) with Bayesian estimation or first-order PK equations [1].

MIPD tools combine popPK models with patient-specific data to recommend dosage adjustments that achieve therapeutic targets more accurately than traditional empirical TDM [16]. The use of such approaches is of particular interest in special populations like neonates, as they enable early dose optimization without the need to wait for steady state, as required in traditional TDM [17], thereby reducing delays in achieving effective drug exposure [18]. Moreover, the use of MIPD can optimize sampling design and simplify protocols by reducing the number of blood samplings [19]. In our institution, one peak concentration and one C_trough on the first dose or one C_trough at steady state are sometimes recommended in specific cases to estimate individual pharmacokinetic (PK) parameters, such as in neonates with renal failure, renal replacement therapy, extracorporeal membrane oxygenation, hypothermia, or other conditions that can impact PK. However, a study in critically ill children evaluating the accuracy of a Bayesian approach for estimating AUC_24,SS compared to calculations based on two concentrations found that the bias in the Bayesian AUC_24,SS estimation was minimal with only one plasma sample measured at trough [11]. This suggests that MIPD could potentially reduce the number of blood samples needed, which is particularly relevant in neonates.

The primary aim of this study was to evaluate in silico whether TDM-guided dosage adjustment could control variability in vancomycin exposure, targeting an AUC_24,SS/MIC of 360–540 h, in a virtual neonatal population, while comparing different TDM strategies. The secondary objective was to identify an optimal sampling strategy to perform Bayesian TDM based on in silico simulations in a neonatal population. Overall, this project aimed at establishing the value of virtual Bayesian TDM studies to inform patient management practices, using the case of vancomycin in neonates as an illustrative example.

Methods

This study was based on a virtual population, making ethical approval unnecessary. The MIPD software used to perform Bayesian TDM was Tucuxi, a recent, open-source software application that allows computing a posteriori maximum likelihood predictions of individual PK parameters based on drug concentration measurements and a reference popPK model integrating relevant patient covariates [20, 21]. Details on Tucuxi’s calculation method are provided in the paper by Thoma and Buclin [21]. The calculation engine can be interfaced with Python scripts, thereby facilitating MIPD automation and dosage adjustment. The numerical computations have been extensively validated against NONMEM [22]. The R software (version 4.3.2) was used to generate the virtual population, for statistical analysis and data processing. NONMEM (version 7.4.2) produced reference PK parameters and data for the virtual neonates, and Python scripts (version 3.6.8) served for serially launching Tucuxi and NONMEM runs with these data and collecting their output.

Virtual Population

A virtual population of 1000 patients was created based on the characteristics of the neonates enrolled in the Dao et al. study [3]. A log-normal distribution was applied to generate gestational age (GA) values between 24 and 42 weeks, ensuring a predominance of premature newborns. Postnatal age (PNA) was assigned using a uniform distribution between 0 and 4 weeks and increased over the course of the study. Bodyweight (BW) was determined based on postmenstrual age (PMA) using growth curves to maintain clinical plausibility and remained constant throughout the study, as PMA varied by a maximum of 2 weeks, causing minimal BW changes [23, 24]. Serum creatinine (SCr) values between 5 and 276 μmol/L were generated using a log-normal distribution, limiting extreme SCr values, and did not vary over time.

Initiation of Vancomycin Dosage Regimen

Four different initial neonatal dosing regimens were evaluated: (1) the standard dosage of 15 mg/kg every 8 h recommended by the manufacturer [25]; (2) the regimen proposed by the neonatal formulary (NNF7) with dosage adjustment based on PMA (Table 1) [26]; (3) the recommendation of the Swiss Pediatric Dosage Harmonization platform (SwissPedDose), consisting of a loading dose of 25 mg/kg followed by the NNF7 dosage [27]; and (4) the regimen recently proposed by El Hassani et al., with a loading dose of 15 mg/kg followed by 11 mg/kg every 8 h [28]. These regimens are referred to hereafter as "manufacturer," "NNF7," "SwissPedDose," and "El Hassani et al.," respectively. Infusion duration was fixed at 1 h.

Table 1.

Dosage regimens recommended by NNF7 [26]

Post-menstrual age (weeks)	Dosage regimen
≤ 29	20 mg/kg q24h
30–33	20 mg/kg q18h
34–37	20 mg/kg q12h
38–44	15 mg/kg q8h
≥ 45	10 mg/kg q6h

qxh every x hours

Simulation of Individual PK Profiles

Individual PK profiles were simulated using the popPK model developed by Dao et al. [3]. This one-compartment model included 23% IIV on clearance (CL), with all parameters allometrically scaled to BW, and CL further influenced by SCr and a PMA-based maturation function. Final equations for CL and volume of distribution (V) were:

$$\text{CL}=0.273\times {\left(\frac{\text{BW}}{1}\right)}^{0.438}\times {\left(\frac{54}{\text{SCr}}\right)}^{0.473}\times \frac{{\text{PMA}}^{3.54}}{{\text{PMA}}^{3.54}+{46.4}^{3.54}}$$

$$V=0.628\times {\left(\frac{\text{BW}}{1}\right)}^1.$$

The NONMEM software, with the ADVAN6 subroutine, was used to create the “reference” individual PK parameters for virtual patients, while constraining the generation of IIV to avoid implausible CL values. Because NONMEM samples IIV from a normal distribution, extreme values can lead to unrealistic half-lives. Therefore, the simulation was restricted so that individual half-lives remained within a clinically meaningful range, within one-fifth and five times the typical population half-life.

Virtual TDM Procedure

Figure 1 schematizes the virtual TDM procedure. Four different TDM strategies were tested and compared, evaluating the cumulative percentage of patients whose AUC_24,SS/MIC values fell within the therapeutic target range of 360–540 h (with the optimal target set at 441 h, corresponding to the geometric mean of the interval). Both the lower and upper limits of the currently recommended 400- to 600-h target were reduced by 10% to account for higher unbound vancomycin concentrations in neonates, with this threshold empirically defined to consider lower plasma protein binding in the absence of data on therapeutic targets for this population [1, 9]. A sensitivity analysis was performed using the original 400- to 600-h target, with the optimal value set at 500 h.

Fig. 1.

Virtual TDM procedure. (1) The dosage regimens recommended by the manufacturer, NNF7, SwissPedDose, and El Hassani et al. were tested [25–28]. (2) For the first TDM intervention, strategy 3 used a single C_trough measured at steady state to estimate AUC_24,SS, whereas strategy 4 relied on two samples collected after the first dose. AUC_24,SS daily area under the curve at steady state, AUC_24,SS/MIC ratio of the daily area under the curve to the minimal inhibitory concentration at steady state, C_trough trough concentration, PK pharmacokinetic, popPK population PK, TDM therapeutic drug monitoring

Strategy 1 (control: no TDM) involved the absence of dose adjustment regardless of any therapeutic target. Strategy 2 (standard TDM) used the traditional method for dosage adjustment to achieve the C_trough target between 10 and 15 mg/L, the recommended range of non-severe, non-MRSA infections to reflect the use of vancomycin treatment in our institution. Strategies 3 and 4 (Bayesian TDM) used Bayesian estimation via the MIPD software Tucuxi [20] to adjust dosing, based on AUC_24,SS/MIC target, with the difference between these two strategies lying in the initial sampling strategy: strategy 3 used a steady-state C_trough, whereas strategy 4 relied on two concentrations (one at 4 h and one C_trough) after the first dose [3].

The “observed concentrations” required for each TDM strategy were simulated using NONMEM, based on the "reference" individual PK profiles, including residual unexplained variability. The software enabled the prediction of "reference" AUC_24,SS, computed using numerical integration from the differential equations implemented in NONMEM, for each virtual patient for the initial and TDM-based dosage regimens.

Dosage Adjustment

For the first TDM intervention, strategies 2 and 3 used a single C_trough measured at steady state (i.e., just before the fourth dose), while strategy 4 implemented an early dose adjustment strategy, as used in critical neonate patients in our institution, based on two samples taken at 4 h and at trough after the first dose (Fig. 1).

The attainment of the therapeutic target was assessed for each patient individually. In strategies 3 and 4, the dosing adjustment criterion was an AUC_24,SS/MIC of between 360 and 540 h, considering an MIC of 1 mg/L, whereas strategy 2 aimed for a C_trough between 10 and 15 mg/L. Once within the therapeutic target, the patient was excluded from further dosage adjustments (Fig. 1), considering that the current dosing regimen was optimal. If drug exposure was outside the target, the aforementioned strategies were applied to adjust dosage. After each adjustment (see Fig. 1), a control C_trough was simulated following three consecutive doses to verify whether the dose adjustment was adequate. If the target was not reached, another TDM intervention was performed using this control C_trough for strategy 2, whereas strategies 3 and 4 incorporated the control C_trough together with the concentrations obtained during the previous interventions to predict the corresponding AUC_24,SS/MIC. This process was repeated for a maximum of four iterations or until drug exposure reached the strategy-specific target.

Dosage adjustments differed depending on the strategy applied. Strategy 1 served as a reference without any dosing adjustment. In strategy 2, the administered dose was adjusted based on the last measured C_trough using a rule of three according to the following formula:

$$\text{Adjusted dose}\left(\text{mg}\right)=\text{Received dose}\times \frac{\text{Targeted}{C}_{\text{trough}}}{\text{Measured}{C}_{\text{trough}}},$$

With a targeted C_trough of 12.5 mg/L, selected as the midpoint of the C_trough target range. No dosing frequency adjustment was performed in strategy 2. In strategies 3 and 4, the Tucuxi software adjusted both the dose and the administration frequency. The dose could range from 1 to 200 mg, and the frequency from every 6 to every 24 h. Tucuxi estimated each patient’s AUC_24,SS by integrating the concentration–time profile using individual PK parameters derived from Dao et al.’s popPK model, described in Section 2.1.2, and the complete history of doses and concentrations [3]. Tucuxi computed dosage adjustments by modifying the administered dose and/or the dosing frequency.

The four virtual TDM and dosage adjustment procedures were applied to all 1000 virtual neonates after treatment initiation with the four regimens described above, providing 16 strategies for comparison.

Identification of an Optimal Sampling Strategy for Bayesian TDM

We conducted a sub-study to evaluate the impact of sampling times on AUC_24,SS predictions generated by the Tucuxi software. To investigate this question, the SwissPedDose initial dosage regimen (i.e., a loading dose of 25 mg/kg followed by the NNF7 dosage detailed in Table 1) implemented in our institution was used, with two samples taken at predefined time points from 1 to 12 h after the loading dose. Each time point was defined within a window of ± 30 min, except for the 12-h time point, which was restricted to 11.5–12 h.

To compare Tucuxi's AUC_24,SS (${\text{AUC}}_{24,\text{SS}}^{\text{Tucuxi}}$) predictions with the “reference” AUC_24,SS values determined by NONMEM, bias and precision were assessed on the log-transformed scale using the mean prediction error (MPE) and relative root mean squared error (RMSE). The MPE was considered non-significant if its confidence interval included 0. For the relative RMSE, the goal was to obtain the lowest possible value while ensuring it remained below 30%.

$$\begin{array}{c}MPE(\%)=\left(\text{exp},\left(\frac{1}{\text{n}},\sum ,\left(\text{ln},\left({\text{AUC}}_{24,{\text{SS}}_{\text{i}}}^{\text{Tucuxi}}\right),-,\text{ln},\left({\text{AUC}}_{24,{\text{SS}}_{\text{i}}}\right)\right)\right),-,1\right)\times 100\end{array}$$ 1

$$\begin{array}{c}RelativeRMSE(\%)=\left(\text{exp},\left(\sqrt{\frac{1}{n}{\sum \left(\text{ln},\left({\text{AUC}}_{24,{\text{SS}}_i}^{\text{Tucuxi}}\right),-,\text{ln},\left(\text{AU},{{\text{C}}_{24,\text{SS}}}_i\right)\right)}^2}\right),-,1\right)\times 100\end{array}$$ 2

with ${\text{AUC}}_{24,{\text{SS}}_i}^{\text{Tucuxi}}$ and ${\text{AUC}}_{24,{\text{SS}}_i}$ for individual $i$ as previously defined, and n being the number of virtual patients.

This analysis was repeated using a single concentration measurement taken between 1 and 12 h after drug administration, to evaluate whether the sampling protocol could be further simplified.

Results

Virtual Population

The virtual population of 1000 neonates had a median PMA of 31 (range 24–42) weeks, a median GA of 30 (24–41) weeks, and a median PNA of 1 (0–4) week. The virtual neonates had a median BW of 1.40 kg, ranging from 0.40 to 4.30 kg, and an SCr level of 54 (16–217) µmol/L, with an equal sex distribution (50% female and 50% male).

Virtual TDM Procedure

Figures 2, 3, and Supplementary Table 1 (see the electronic supplementary material) summarize the percentages of patients with “reference” AUC_24,SS/MIC between 360 and 540 h for the different initial dosing regimens and TDM strategies tested. Initiating treatment with the dosage regimens recommended by NNF7 or SwissPedDose resulted in a higher proportion of patients reaching the therapeutic target (36% and 33–34%, respectively) compared to those recommended by the manufacturer or El Hassani et al. (3% and 19%, respectively). A slight difference between strategies 3 and 4 at treatment initiation was observed, which may be explained by the subtle variations in PNA that occur when waiting for steady state in strategy 3 but are not accounted for when steady state is estimated (strategy 4).

Fig. 2.

Boxplot representing AUC_24,SS/MIC according to the different strategies studied and the initial dosages tested. The lines represent the required targets, with the semi-solid line representing the best at 441 h, and the solid lines at 360 and 540 h, the extremities of the target interval. The initial dosages tested were those recommended by the manufacturer, NNF7, SwissPedDose, and El Hassani et al. [25–28]. AUC_24,SS/MIC ratio of the daily area under the curve to the minimal inhibitory concentration at steady state. TDM therapeutic drug monitoring

Fig. 3.

Probability of target attainment, defined as the percentage of AUC_24,SS/MIC values falling within the 360–540 h range (black areas), for the different tested dosages and the four strategies after the first (above) and the fourth TDM (below) intervention. White and grey areas indicate AUC_24,SS/MIC values predicted below or above this target range, respectively. White numbers represent the percentage of patients within the target. AUC_24,SS/MIC ratio of the daily area under the curve to the minimal inhibitory concentration at steady state, TDM therapeutic drug monitoring

Strategy 1 (no TDM, no dosage adjustment) led to a low proportion of patients achieving the AUC_24,SS/MIC therapeutic target, ranging from 4% with the manufacturer-recommended regimen to 49% with the SwissPedDose regimen. With strategy 2, 100% of patients achieved a C_trough within the target range of 10–15 mg/L after at most two dose adjustments (data not shown). This trough-guided approach resulted in 92% and 75% of patients achieving a “reference” AUC_24,SS/MIC between 360 and 540 h after two TDM interventions when using the every-8-h dosing regimens recommended by the manufacturer and El Hassani et al. However, for the NNF7 and SwissPedDose dosing regimens defining intervals between 8 and 24 h, this strategy was ineffective, as only 31–32% of patients reached the AUC_24,SS/MIC therapeutic target despite appropriate C_trough adjustment. Lower C_trough values, in the range of 5–15 mg/L, would need to be targeted in order to achieve AUC_24,SS/MIC values between 360 and 540 h, as shown in Supplementary Figure 1. Strategy 3 (Bayesian TDM based on one C_trough sample, using Tucuxi) allowed over 74% of patients to reach the AUC_24,SS/MIC therapeutic target after a single TDM intervention, and 100% after four TDM interventions, regardless of the initial dosing regimen used. Comparable results were obtained with strategy 4, which used two samples after the first dose to predict and adjust the steady-state concentration, demonstrating the accuracy of Tucuxi's steady-state predictions. Of note, similar findings were obtained for all tested strategies using the adult AUC_24,SS/MIC therapeutic target (400–600 h, data not shown).

Optimal Sampling Strategy for Bayesian TDM

In this study, the AUC_24,SS predictions made by Tucuxi were highly comparable to the “reference” patient values computed by NONMEM based on the simulated PK parameters, regardless of sampling times for the two concentrations. As shown in Fig. 4, the median (min, max) MPE was − 1 (− 3, 2)% and the median relative RMSE was 16 (12, 23)%. Under these conditions, accuracy was further improved when both samples were obtained later in the dosing interval (approximately ≥ 8 h post-dose).

Fig. 4.

Heatmaps showing the impact of different two-point sampling strategies on the accuracy of Bayesian AUC_24,SS estimation using Tucuxi. Each square represents a combination of first and second sampling times after the initial dose (x-axis: first sample; y-axis: second sample). The left panel displays the median MPE (%), indicating bias, while the right panel shows the median RMSE (%), indicating precision. Green areas correspond to values close to zero (better performance), yellow areas indicate positive deviation (overestimation), and orange areas indicate negative deviation (underestimation). AUC_24,SS daily area under the curve at steady state, MPE mean prediction error, RMSE root mean squared error

Similar results were obtained with a single sample, as shown in Figure 5, with a median MPE of 0.7 (− 1, 3)% and a relative RMSE of 20 (15, 23)%. Although samples collected very early after dosing (< 4 h post-dose) were somewhat less reliable, the prediction bias remained low (MPE < 5%). Moreover, beyond approximately 5 h post-dose, the 95% confidence interval of the MPE includes zero, indicating no significant bias.

Fig. 5.

Comparison of Tucuxi-predicted AUC_24,SS with reference AUC_24,SS for a single concentration collected between 1 and 12 h post-dose. The vertical solid line represents the desired value (i.e., 0 for MPE). The black points and horizontal lines correspond to the median and the 95% confidence interval for MPE. AUC_24,SS daily area under the curve at steady state, MPE mean prediction error

Discussion

This virtual TDM study in neonates receiving intermittent vancomycin compared four strategies, including conventional trough-based monitoring and computer-assisted Bayesian MIPD.

Strategy Comparison

As anticipated, the absence of TDM (strategy 1) keeps the wide variability in exposure unchanged, leading most virtual patients outside the defined therapeutic range with a significant risk of toxicity, while only a minority of patients reach the therapeutic target. The NNF7 and SwissPedDose dosing regimens propose age-dependent dosing intervals ranging from 6 to 24 h, which may be more physiologically adapted to neonatal maturation, whereas the El Hassani et al. and manufacturer recommendations rely on a fixed 8-h interval, resulting in more uniform dosing schedules across ages. Notably, the previous assessment of the NNF7 and SwissPedDose regimens in the Dao et al. study did not reveal the observed overexposure, as a broader AUC_24,SS/MIC therapeutic range of 400–700 h was used, at a time when the upper limit of this range was less well defined [3, 26, 27].

Interestingly, TDM based on C_trough (strategy 2) appears suboptimal for long dosing intervals (NNF7 and SwissPedDose), leading most patients in the supratherapeutic range, and only 31% in the desired AUC_24,SS/MIC range. This strategy appears, however, satisfactory for regimens with shorter dosing intervals (manufacturer and El Hassani et al.) and leads, with two cycles of TDM, up to 75% of the population into the therapeutic range. This shows that simple linear adaptation based on C_trough targeting the 10–15 mg/L range is less predictive of an adequate exposure in terms of AUC_24,SS, and depends mostly on the regularity of dosing intervals. Wicha et al. indeed illustrated well how C_trough is a suboptimal surrogate of AUC_24,SS and can be misleading, especially for wide dosing intervals, as promoted in several neonatal dosing regimens, including the NNF7 [18]. These results are consistent with other studies demonstrating a weak correlation between C_trough and AUC_24,SS for vancomycin [12, 15]. In addition, the very high AUC_24,SS values observed for the SwissPedDose and NNF7 regimens under strategy 2 arise from the fact that dose adjustment was performed exclusively on the total dose while keeping the original long dosing intervals (e.g., every 18 h or 24 h) unchanged. Achieving a C_trough within 10–15 mg/L under such extended intervals requires substantial, and in some cases clinically unrealistic, dose increases, which automatically drives AUC_24,SS to markedly elevated levels. However, these extreme scenarios concern only 3% of the simulated population and therefore do not influence the overall conclusions of the study.

Across all initial dosing regimens, one cycle of Bayesian TDM based on a single C_trough sample at steady state (strategy 3) optimizes exposure in terms of AUC_24,SS in at least 74% of our virtual population of neonates, and 100% after four TDM cycles. This approach has the potential to limit supratherapeutic exposure and potential toxicity. A similar difference between strategy 2 and strategy 3 was also observed in adults in a prospective study, where AUC_24,SS-based TDM proved preferable, as C_trough-based TDM led to overexposure and increased the risk of adverse effects [12].

Strategy 4, corresponding to Bayesian TDM based on two samples collected from the first dose, showed results similar to strategy 3. These findings highlight the good predictive accuracy of Tucuxi, which allows for precise dose adjustments as early as on the second dose, without waiting for steady state. This early adaptation is clinically crucial, as it enables early optimization of drug exposure. However, obtaining two early samples in real-life neonatal care can be challenging due to limited blood volume, clinical instability, and workload constraints. These practical considerations may therefore encourage clinicians to favor strategy 3 in routine practice.

Sampling Time Optimization and Literature Comparison

The study also addressed the question of optimal sampling timing for Bayesian TDM of vancomycin in neonates. Our results show good precision with small values of MPE and RMSE for both the single-sample and the two-sample strategies. Precision was slightly better in the two-sample strategy when both concentrations were measured at least 8 h post-dose. These results support drawing one sample at least 5 h after the first drug administration to reliably inform the Tucuxi software tool and guide further dose optimization. This is a consequence of the remarkably powerful Bayesian MIPD approach. It represents a real advantage over C_trough adaptation, which relies on a sample being taken precisely on trough time at steady state, the measurement being uninterpretable otherwise. However, in clinical practice, timing delays can occur due to organizational constraints, compromising the reliability of dosage adjustments based solely on C_trough [12]. A prospective study conducted in children aged 1–17 years showed that the AUC_24,SS estimated using MIPD on a single sample was comparable to the AUC_24,SS estimated from two samples [11]. MIPD simplifies patient management by reducing the number of required samples, which is particularly important in neonatology, where blood sampling should be minimized.

Our results contrast to some extent with adult studies. One virtual study evaluated the optimal sampling strategy in adults with hematological malignancies and various degrees of renal failure. It compared vancomycin AUC_24,SS obtained with a full PK profile with estimates derived from single- or two-sample strategies [19]. A single sample yielded better predictions when drawn later than 4 h after the third maintenance dose, compared to samples taken after the first maintenance dose [19]. Similarly, other studies found that in adults, the ideal single-sample timing to estimate vancomycin AUC_24,SS was around 6 h post-dose after the fifth dose, while C_trough led to larger imprecision [29, 30]. On the other hand, Oda et al. found that single-sample strategies on day 2 were less accurate than two-sample strategies for predicting AUC_24,SS [31].

The choice of the popPK model used to perform Bayesian TDM may explain part of these differences. The popPK model was selected primarily for its accurate representation of our local population, for which the Bayesian TDM implementation is being considered [3]. It was developed based on a majority of C_trough and some peak concentrations, and is characterized by IIV on CL, but not on V (probably due to the inclusion of BW in the model explaining most variability on V in this population), with a mixed error model for the residual variability [3]. The predictive performance of our model was evaluated as suitable for MIPD in infants and neonates according to one recent study [32]. The absence of variability in our model on V may explain the good precision of AUC_24,SS estimates based on a single sample, as it is sufficient to inform CL. In contrast, studies in adults have used two-compartment models incorporating variability on both CL and V, making parameter estimation more challenging with a single sample and increasing the risk of bias depending on sampling times [19, 29, 31]. Such complex popPK models require more sampling points to adequately capture all PK phases, and optimal sampling times may thus vary depending on the model used, as shown in the study by Shingde et al. [30]. Furthermore, Oda et al. hypothesized that the residual error model used could influence estimation accuracy, with an additive error model favoring earlier samples, while a proportional error model rather privileges C_trough samples [31]. In our case, a mixed error model appears to be a good compromise to ensure accurate predictions regardless of sampling time. Therefore, the choice of the popPK model appears to play a key role in defining optimal sampling times, highlighting the importance of tailoring Bayesian TDM strategies to the underlying popPK model.

Clinical Value of MIPD

Although the advantages of MIPD are well described in the literature, their implementation in clinical practice remains limited, requiring an interface with clinical software, clinical validation, and trained personnel [12, 33]. The advantages of MIPD include (1) reduced costs related to uninterpretable data from samples drawn at the wrong time [34], (2) a smaller risk of bias than linear PK calculations on two distinct samples [18], and (3) ease of use for clinicians [35]. Our findings support further implementation of MIPD for the TDM of vancomycin in neonates. This approach has already been described as an effective strategy for improving target attainment in neonates [36]. Notably, a prospective study by Kalamees et al. conducted in 48 neonates demonstrated that the use of MIPD increased the proportion of patients reaching the therapeutic target from 37% to 60%, compared to standard dosing without MIPD [37]. Similarly, a study conducted in children over 1 year showed that MIPD was more effective than simple clinician judgment in recommending vancomycin dosing regimens, achieving target concentrations in 93% compared to 72% of patients [38]. A multicenter, randomized, controlled clinical trial, named BENEFICIAL, is currently underway to study the impact of vancomycin precision dosing based on a bedside model in critically ill neonates and children [39].

The AUC_24,SS/MIC target range for vancomycin is not clearly defined in neonates, but is derived from adult recommendations [1]. To account for the higher unbound fraction of vancomycin in neonates, we empirically set a lower limit of 360 h and a upper limit of 540 h as a conservative approach, considering the lack of data on therapeutic targets in this age group [9]. This threshold aims to ensure efficacy comparable to adults despite lower total concentrations. In practice, AUC_24,SS/MIC values between 360 and 400 h would generally be considered clinically acceptable without a need for dose adjustment, especially if treating non-MRSA infections. Importantly, the use of the adult AUC_24,SS/MIC target range consistently supported the superiority of Bayesian TDM over C_trough-based approaches for dose adjustment. Therefore, despite the remaining uncertainty regarding which AUC_24,SS/MIC target is best in pediatric patients, our main conclusion remains unchanged: Bayesian TDM provides more reliable dose refinement than reliance on C_trough alone. Prospective pharmacodynamic data regarding efficacy and toxicity in neonates are scarce. The NeoVANC study reported increased ototoxicity in neonates receiving a more intensive regimen of vancomycin based on simulations targeting an AUC_24,SS/MIC > 400 h without an upper limit, versus a standard regimen [40]. However, PK data are not yet available to correlate these results with differences in exposure. Regarding cost-effectiveness, clinical studies remain to be conducted in neonates to confirm that avoiding vancomycin-related ototoxicity and/or nephrotoxicity is cost-effective. Nevertheless, monitoring AUC_24,SS using MIPD software rather than C_trough is expected to be beneficial, in particular because it may reduce the risk of acute kidney injury, thereby decreasing hospitalization duration and associated costs [12, 41].

Study Limitations

The main limitation of our study is the lack of prospective validation in a real neonatal population. The virtual population was designed based on characteristics observed among neonates treated in the Lausanne University Hospital, ensuring a realistic representation of vancomycin PK in this group. However, this population mainly consisted of extremely and very preterm neonates, which limits the generalizability of our findings to the entire neonatal population. As such, our conclusions regarding the relative performances of initial dosing regimens should be interpreted primarily for these populations of neonates and may not directly extrapolate to infants with more mature physiology. Nonetheless, based on similar studies, we believe this group is representative of the neonatal population most likely to receive vancomycin in real-life clinical practice [5, 42]. Due to the complexity of accurately modeling changes in SCr and BW over time, we opted to keep these parameters constant throughout the study. Although this simplification may underestimate variability, simulating realistic individual trajectories for neonates remains highly challenging due to their heterogeneous and clinically dependent evolution. Given the short observation period of our simulations, this approach has probably a limited impact on our findings. Nonetheless, a prospective validation incorporating dynamic changes in these parameters remains warranted to bring further confirmation of our findings in real-world conditions.

In addition, our study was based on a single popPK model, developed in our institution, used to simulate the data and to perform Bayesian TDM. While we cannot exclude a slight positive bias in performance estimates, this is not expected to impact our comparative evaluation of TDM strategies. Theoretically, such a comparison could be performed using any popPK model adequately describing vancomycin disposition. However, using the same model that generated the simulations eliminates a potential issue of model misspecification that could distort the results. Our findings indeed agree with existing studies, which used different popPK models and demonstrated a benefit of Bayesian TDM in improving target concentration attainment [37, 38].

Conclusion

In silico virtual TDM studies like the present one are ideal to compare various strategies and find the best clinical algorithm expected to optimize patient management while considering convenience and sparing resources [43, 44]. This approach is fostered by the availability of open-source software tools like Tucuxi, which can be automated to run with large amounts of simulated clinical data. Only the winning strategy needs to be eventually validated in real patients, available in limited number. A recent neonatal study has confirmed the similarity between results obtained through popPK simulations and real-world clinical data, reinforcing the relevance of these methods for guiding dose optimization in clinical practice [45].

In conclusion, TDM is an important tool to optimize vancomycin exposure in neonates and to increase the proportion of patients in the AUC_24,SS/MIC therapeutic target. C_trough-based adjustments have limitations and do not correlate well with an AUC_24,SS/MIC within the desired target. In contrast, our results demonstrate that MIPD software, such as Tucuxi, enables optimization of vancomycin exposure regardless of the initial dosing regimen in almost all patients, helping them reach the defined therapeutic target. Furthermore, Bayesian TDM simplifies sampling strategies, requiring only a single sample after the first administration, drawn at least 5 h after the end of infusion. These findings reinforce the importance of integrating Bayesian-guided TDM into routine neonatal care to improve precision dosing and therapeutic outcomes.

Supplementary Information

Below is the link to the electronic supplementary material.

Funding

Open access funding provided by University of Lausanne. No dedicated funding has been allocated for this project, which was supported by usual research funds of the Centre for Research and Innovation in Clinical Pharmaceutical Sciences, Lausanne University Hospital, and University of Lausanne (Switzerland).

Declarations

Conflict of interest/competing interests

Thierry Buclin is an Editorial Board member of Clinical Pharmacokinetics. Thierry Buclin was not involved in the selection of peer reviewers for the manuscript nor any of the subsequent editorial decisions. The other authors have nothing to declare.

Ethics approval

Not applicable.

Consent to participate

Not applicable.

Consent to publication

Not applicable.

Availability of data and material

No data were used in this study.

Code availability

The MIPD software used for this project is openly available on GitHub at https://github.com/sotalya/tucuxi-core. The Python code required to run the simulations and analyses is available upon request from the corresponding authors.

Authors’ contributions

Conceptualization: AR, YT, AC, PT, CM, EG, TB, CC, KD, and MG; formal analysis: AR, YT, KD, and MG; methodology: AR, YT, AC, TB, CC, KD, and MG; supervision: KD and MG; writing—original draft: AR and KD; writing—review and editing: AR, YT, AC, PT, CM, EG, TB, CC, KD, and MG. All authors have read and agreed to the published version of the manuscript.