Optimizing Vancomycin Dosing and Monitoring in Neonates and Infants Using Population Pharmacokinetic Modeling

ABSTRACT

We determined optimal vancomycin starting dose regimens in infants ≤180 days of age to achieve the highest probability of target attainment with an area under the concentration-time curve for 24 h (AUC₂₄) of ≥400 using population pharmacokinetic (PK) modeling. Secondarily, determination of the relationship between serum creatinine (SCR) and vancomycin clearance in neonates was done. A retrospective population PK study was designed and included pediatric patients ≤180 days old who had received vancomycin and had a serum vancomycin concentration sampled. A population PK model was developed using Pumas (v1.0.5). Simulation was performed with various dosing regimens to evaluate the probability of AUC₂₄ target attainment and probability of trough of ≤20 mg/liter, and comparison to published models was performed. Individual clearance estimates, obtained from the final model, were plotted against SCR and faceted by age quartiles to assess the relationship between SCR and vancomycin clearance. A total of 934 patients were included in the study (58.6% male; median age, 43.6 days [range of 0 to 184]; median number of concentration samples, 1 [range of 1 to 29]). A one-compartment model was developed with body weight (WT), SCR, and postmenstrual age (PMA) identified as significant covariates on clearance. Plotting vancomycin clearance versus SCR demonstrated no clear relationship between the two at <10 days postnatal age (PNA). Dosing regimens to attain AUC₂₄ and trough targets were stratified according to SCR for ≥10 days PNA and PMA for <10 days PNA. A vancomycin population PK model was developed for pediatric patients <180 days of age incorporating WT, SCR, and PMA. The relationship between vancomycin clearance and serum creatinine is not clear at <10 days PNA.

INTRODUCTION

Despite being in clinical use for over 50 years, vancomycin use and monitoring is still imperfect and evolving due to the challenge of balancing the risk of nephrotoxicity and achieving therapeutic target goals. This is especially the case in neonates and infants who are undergoing developmental changes, resulting in variable vancomycin pharmacokinetics. Renal function in neonates is underdeveloped, and the maturation process is continuous during early childhood, making it difficult to monitor renal function and characterize vancomycin clearance (1). Further confounding the ability to evaluate renal function in newborns is the influence of maternal serum creatinine, which is present in newborns (2). As a result of the variability in pharmacokinetics in this pediatric population, there still has not been a consensus established for dosage recommendations in neonates and infants with differing algorithms offered by two major online databases, Lexicomp and Neofax, and by the vancomycin dosing guidelines (3–5).

Vancomycin therapeutic drug monitoring in neonates and infants is another area lacking adequate evidence. The importance of drug monitoring for adults on vancomycin has been established, but there have been no data collected evaluating monitoring for neonates and infants. The prominently used vancomycin trough concentration targets of 10 to 20 μg/ml may not be appropriate for pediatrics, with recent studies indicating that troughs of >15 μg/mL were not associated with improved clinical outcomes but were associated with increased incidence of nephrotoxicity (6). These trough goals were originally implemented to act as a surrogate measure for the established target of area under the concentration-time curve (AUC)/MIC of >400. However, recent studies have demonstrated that trough is a poor marker for AUC/MIC (5, 7). Therefore, it is important to establish what is the best starting dose for neonates and infants to achieve a target AUC/MIC of ≥400 rather than dosing to achieve a trough target.

Population pharmacokinetic (PK) modeling can be used to characterize the intra- and intersubject variability by utilizing patient-specific characteristics, dosing instructions, and observed drug concentrations in patients. It can serve as a valuable tool in selecting appropriate dose recommendations to support personalized therapeutic monitoring. We performed a retrospective population PK modeling and simulation study to determine the optimal starting dose for vancomycin in neonates and infants. We also explored the relationship between serum creatinine and vancomycin clearance in neonates and infants to assess the impact of serum creatinine in dosing decisions.

RESULTS

Vancomycin dosing, concentration, and demographic data of 934 patients (2,471 vancomycin concentrations, median of 1 sample per patient [range, 1 to 29], median sampling time after last dose of 9.3 h [range, 1.1 to 74.3]) were used for analysis in this study (see Fig. S1 in the supplemental material). Of the 934 patients, 455 patients (48.7%) had data obtained when the patient was <28 days postnatal age (PNA) (Table 1).

TABLE 1.

Baseline characteristics^a

Parameter	Mean	Median	IQR	Min	Max
SCR (mg/dL)	0.46	0.33	0.25, 0.51	0.1	4.54
WT (kg)	3.57	3.58	1.97, 4.91	0.37	11.88
ht (cm)	49.27	51	41.76, 57	17.5	73.5
PMA (wk)	42.41	41.99	34.85, 48.58	20.89	66.68
PNA (days)	57.89	43.6	18.64, 86.18	0	184
Follow time (h)	409.7	64	25.01, 422.5	4.68	4,165.97
No. of doses	12.6	6	4, 14	1	139
No. of samples	2.65	1	1, 3	1	29
Dose (mg)	50.4	48	27, 70	5	175
Dose (mg/kg)	14.72	14.75	13.79, 15.28	2.45	28.38
Vancomycin serum concn (mg/liter)	13.01	11.7	8.5, 15.8	5	49.5
Sampling time after dose (h)	12.37	9.3	7.06, 16.4	1.09	74.33

^aFollow time, amount of time the dosing and concentration history was recorded for each patient; number of doses, the number of doses each patient received; number of samples, the number of concentrations collected for each patient; vancomycin serum concentration, summary of all values of concentrations collected in the study; sampling time after dose, when the concentrations were collected with regard to dose for all concentrations.

A one-compartment model with proportional and additive error demonstrated the best overall fit. Allometrically scaled weight, postmenstrual age, and serum creatinine were identified as significant covariates on clearance, and weight was identified as a significant covariate on volume of distribution. The inclusion of weight as a covariate resulted in a 1,281.6-point drop in objective function value (OFV), the inclusion of serum creatinine resulted in a 2,579.7-point drop in OFV, and the inclusion of PMA as a covariate resulted in a 330.6-point drop in OFV. The parameter estimates and their relative standard errors (RSE) for the final covariate model are provided in Table 2. The eta shrinkage was 19.4% for clearance (CL) and 57.9% for volume (V). The goodness-of-fit diagnostic plots and prediction-corrected visual predictive check plot (Fig. 1 and 2) demonstrated good model fit and no trends of bias. The model was validated internally by simulating the dosing profile of each patient using the patient-specific covariates and then comparing the resulting predicted concentrations to the observed concentrations (Table 3).

TABLE 2.

Final model parameter estimates

Population PK parameter	Estimate	RSE (%)
CLa (liters/h)	0.237	1.01
Vb (liters)	2.98	2.94
Effect of SCR on CL	0.87	6.35
Effect of PMA on CL	0.81	2.71
Between-subject variability CL (%)	23	4.74
Between-subject variability V (%)	25	10.4
Proportional error	0.19	6.39
Additive error	1.66	36.5

^a$\mathrm{CL}\hspace{0.17em}=\hspace{0.17em}\mathrm{0.237\times }{\left(\frac{\text{WT}}{3.5}\right)}^{0.75}\times {\left(\frac{0.45}{\text{SCR}}\right)}^{0.87}\times {\left(\frac{\text{PMA}}{42}\right)}^{0.81}\times e^{\eta \text{CL}}$.

^b$V=2.98\times \left(\frac{\text{WT}}{3.5}\right)\times e^{\eta V}$.

FIG 1.

Goodness-of-fit diagnostic plots. (A) Observed versus individual predicted vancomycin concentration plot. (B) CWRES versus time after dose plot. (C) CWRES versus predicted vancomycin concentration plot.

FIG 2.

Prediction-corrected visual predictive check by time after dose for first 24 h.

TABLE 3.

Comparison of final model to models previously found in literature

Source	Model	Mean CLi (liters/h)	Mean Vi (liters)	Median prediction error (%)	Median absolute prediction error (%)
Current study	CL = TVCL × (WT/3.5)0.75 × (0.45/SCR)0.87 × (PMA/42)0.81, V = TVV × (WT/3.5)	0.398	3.09	−13.9	30.9
Frymoyer et al. (8)	CL = 0.345 (WT/2.9)0.75 × Fmat × (1/SCR)0.267, Fmat = 1/(1 + [PMA/TM50]−Hill), V = 1.75 (WT/2.9), TM50 = 34.8, Hill = 4.53	0.398	2.17	−32.9	44.1
Lo et al. (9)	CL = 1.0 × (WT/70)0.75 × (PMA/30)3.16 × [0.83 × SGA + 1.03 × (1 − SGA)], V = 36.6 × WT/70, SGA = 1 for small for gestational age infants, 0 for appropriate for gestational age infants	0.508	1.88	−38.5	61
Zhou et al. (10)	CL = 0.0571 × (WT/1.416)0.513 × (birth WT/1.01)0.599 × (1 + 0.282 × (PNA/17)) × (1/(SCR/0.475)0.525), V = 0.791 × (WT/1.416)0.898	0.3	2.31	−14	46.2

Evaluation of clearance versus serum creatinine.

Individual clearances estimated using maximum a posteriori (MAP) were plotted against serum creatinine and categorized by gestational age and postnatal age quartile groups (Fig. 3). A lack of a direct relationship is evident in the first PNA group of 1 to 10 days, but for PNA greater than 10 days, there is a consistent steep decline of clearance as serum creatinine increases. There are also more patients with serum creatinine above 1.5 mg/dL observed within the first 10 days of PNA, especially in patients also in the first gestational age group of 20 to 27 weeks.

FIG 3.

Individual estimated clearance versus serum creatinine faceted by gestational age and postnatal age quartiles. This plot only includes serum creatinine at baseline and serum creatinine of <2 mg/dL.

Simulation.

Four different groups of serum creatinine, based upon the distributions in the study population, were used to conduct simulations: SCR ≤ 0.25, 0.25 < SCR ≤ 0.5, 0.5 < SCR ≤1, and 1< SCR ≤1.5 mg/dL. SCR greater than 1.5 mg/dL were excluded from the simulations because generalized dosing recommendations cannot be provided for these patients who might be experiencing acute kidney injury. Patients with PNA less than or equal to 10 days were excluded from the original data, which was then sampled so that 1,000 patients were taken from each of these four serum creatinine groups. Therefore, a total of 4,000 simulated patients were created with randomly sampled sets of postmenstrual age and weight and randomly sampled serum creatinine within each SCR group. The summary statistics of this simulation patient population are listed in Table 4. Using the parameter estimates listed in Table 2, 22 weight-based dosing regimens were simulated to determine probability of target attainment (PTA) of an AUC₂₄ of ≥400 and trough of ≤20 mg/liter for each of the four SCR groups.

TABLE 4.

Summary statistics of simulation population

Model and parameter	Mean	Median	IQRa	Min	Max
SCR group (n = 4,000)
SCr (mg/dL)	0.62	0.51	0.26–0.97	0.12	1.5
WT (kg)	3.56	3.59	1.89–5	0.52	9.98
PMA (wk)	42.31	42.81	33.3–48.24	24.83	66.49
PNA <10 days-based simulations (n = 3,000)
SCr (mg/dL)	0.73	0.68	0.5–0.9	0.3	1.5
WT (kg)	1.82	1.61	0.71–2.85	0.44	4.68
PMA (wk)	32.57	32.91	26.53–38.21	20.89	42.09

^aIQR, interquartile range.

The PTA results for all explored dosing regimens are listed in Table 5. The dosing regimen that achieved highest PTA for both AUC and trough was as the following: SCR ≤ 0.25 (25 mg/kg of body weight every 8 h [q8h], 72.7%), 0.25 > SCR ≤ 0.5 (30 mg/kg every 12 h, 75.2%), 0.5 < SCR ≤1 group (30 mg/kg every 24 h, 80.1%), and 1< SCR ≤1.5 mg/dL (15 mg/kg every 24 h, 71.4%).

TABLE 5.

SCR group-based dosing regimen simulations^a

SCR group (mg/dL) and dosing regimen	AUC24	AUC24 >400, %	Trough	Trough <20, %	P value (AUC and trough), %
SCR < 0.25
10 mg/kg q6h	279.8 ± 73.5	6.7	7.6 ± 2.7	100	6.7
12 mg/kg q6h	335.4 ± 88	22.5	9.1 ± 3.3	100	22.5
14 mg/kg q6h	390.3 ± 103.8	42.6	10.6 ± 3.8	99.1	41.7
15 mg/kg q6h	420.7 ± 110.4	52.9	11.4 ± 4	96.5	49.4
15 mg/kg q8h	316.7 ± 84.5	16.2	7.2 ± 3.2	99.9	16.1
20 mg/kg q6h	566.6 ± 150.8	88.9	15.4 ± 5.6	78.9	67.8
20 mg/kg q8h	421.88 ± 111.2	52.4	9.6 ± 4.2	98.5	50.9
20 mg/kg q12h	280.2 ± 73.3	6.3	4.3 ± 2.7	100	6.3
25 mg/kg q6h	706.7 ± 199.5	96.9	19.1 ± 7.3	58.1	55
25 mg/kg q8h	529.3 ± 142.6	82.1	12.1 ± 5.5	90.6	72.7
25 mg/kg q12h	352.7 ± 94.4	27.5	5.5 ± 3.3	99.9	27.4
0.25 < SCR < 0.5
10 mg/kg q6h	445.5 ± 123.9	59.2	13 ± 3.7	96.2	55.4
10 mg/kg q8h	337.9 ± 100.7	23	9.2 ± 3.4	99.8	22.8
12 mg/kg q6h	540.5 ± 146.5	84.7	15.8 ± 4.4	82.8	67.5
12 mg/kg q8h	393.1 ± 109.6	40.4	10.7 ± 3.7	98.5	38.9
14 mg/kg q6h	619.6 ± 184.1	92.5	18.1 ± 5.4	66.4	58.9
14 mg/kg q8h	465.6 ± 131.9	65.7	12.7 ± 4.5	93.6	59.3
14 mg/kg q12h	310 ± 87.6	14.4	6.9 ± 3.3	100	14.4
15 mg/kg q6h	668.4 ± 187.8	96.2	19.5 ± 5.7	55.9	52.1
15 mg/kg q8h	500.6 ± 142.7	74.3	13.7 ± 4.8	88.9	63.2
15 mg/kg q12h	334.4 ± 96.2	21.6	7.5 ± 3.6	99.9	21.5
20 mg/kg q6h	888.4 ± 263.2	99.7	25.9 ± 8	24.9	25.2
20 mg/kg q8h	666.9 ± 190.3	95.7	18.2 ± 6.5	64.5	60.2
20 mg/kg q12h	445.5 ± 130.2	59.2	10 ± 4.7	96.4	55.6
25 mg/kg q8h	837.8 ± 246.3	99.6	22.9 ± 8.4	39.5	39.1
25 mg/kg q12h	558.5 ± 165.6	82.9	12.3 ± 6	88.5	71.4
30 mg/kg q12h	668.2 ± 185.81	96.2	14.9 ± 6.9	79	75.2
0.5 < SCR ≤1
10 mg/kg q6h	808.2 ± 246.2	99.3	21.7 ± 4.9	38.3	37.6
10 mg/kg q8h	606.6 ± 172.1	90.5	17 ± 4.3	75	65.5
10 mg/kg q12h	405.2 ± 113.2	46.8	11 ± 3.6	99	45.8
12 mg/kg q6h	976.2 ± 282	99.9	26.2 ± 5.8	13.3	13.2
12 mg/kg q8h	728.5 ± 208.7	96.5	20.4 ± 5.2	48.7	45.2
12 mg/kg q12h	485.9 ± 139.7	69.8	13.3 ± 4.3	92.9	62.7
14 mg/kg q12h	563.5 ± 162.9	85.3	15.4 ± 5.1	82.9	68.2
15 mg/kg q12h	605.7 ± 175.6	89.9	16.5 ± 5.5	74	63.9
30 mg/kg q24h	607.4 ± 174.5	91.7	12.1 ± 6.3	88.4	80.1
SCR > 1
10 mg/kg q6h	1,357.7 ± 367.1	100	29.4 ± 5.7	1.7	1.7
10 mg/kg q8h	1,007.7 ± 258	100	24.7 ± 4.9	17	17
10 mg/kg q12h	685.2 ± 171.6	97.6	18.6 ± 4.1	64.5	62.1
12 mg/kg q12h	815.9 ± 220.5	99.4	22.2 ± 5.3	35.1	34.5
15 mg/kg q24h	508.8 ± 130.8	79.2	13 ± 4.5	92.2	71.4
20 mg/kg q24h	674.3 ± 182.1	96.1	17.2 ± 6.2	69.5	65.6
30 mg/kg q24h	1,019.5 ± 275.1	100	26.2 ± 9.5	27.7	27.7

^aGrey shading highlights the dosing regimens which achieved highest PTA for both AUC and trough in each SCR group.

For neonates ≤10 days postnatal age, 1,000 simulations were done for each of three groups: <28 weeks PMA, 29 to 36 weeks PMA, and ≥37 weeks PMA. Patients with SCR > 1.5 mg/dL were excluded from these simulations as well. Dosing with the highest probability of goal attainment was the following: <28 weeks PMA (8 mg/kg/dose every 8 h, 70.7%), 29 to 36 weeks PMA (30 mg/kg/dose every 24 h, 64.3%), and ≥37 weeks PMA (30 mg/kg/dose every 24 h, 60.1%).

In addition, 3 population PK models of vancomycin from the literature for the same pediatric population were evaluated using the same data set and same process as our model (8–10). These 3 population PK models along with the model developed in this study were evaluated for bias and precision (Table 3). The current study model performed the best with a median prediction error of −13.8% and a median absolute prediction error of 30.9%.

Current Neofax dosing recommendations for vancomycin are 10 to 15 mg/kg/dose every 6 to 18 h depending on PMA and postnatal age. The simulated patients were dosed per the Neofax dosing table, and then they were binned into the four serum creatinine groups to calculate PTA (Table 6) (3). The PTA using Neofax dosing for the SCR ≤ 0.25 group was 35.4%. For the 0.25 < SCR ≤ 0.5 group, Neofax dosing achieved 45.3% AUC and trough target attainment. For the 0.5 < SCR ≤ 1 group, Neofax dosing achieved 42.3% target attainment. For the last SCR group, Neofax dosing achieved 44.6% AUC and trough target attainment. The simulated patients were also dosed per the Neofax dosing table and then were binned into the three groups of neonates 10 days old or younger to calculate PTA (Table 7). The PTA using Neofax dosing were 52.4%, 39.9%, and 46.3% for extremely preterm, preterm, and term neonates, respectively. The Neofax dosing PTAs for all three groups were lower than the PTAs for the new dosing regimens proposed based on our simulations.

TABLE 6.

Comparison of probability of AUC and trough target attainment for SCR-based regimen to Neofax dosing

SCR group	Probability (%) of AUC and trough target attainment for:
	Neofax dosing	Proposed new dose
≤0.25 mg/dL	35.4	72.7
>0.25–0.5 mg/dL	45.3	75.2
>0.5–1 mg/dL	42.3	80.1
>1 mg/dL	44.6	71.4

TABLE 7.

Comparison of probability of AUC and trough target attainment for dosing in neonates ≤10 days old to Neofax dosing

SCR group	Probability (%) of AUC and trough target attainment for:
	Neofax dosing	Proposed new dose
Extremely preterm (<28 wk PMA)	52.4	70.7
Preterm (<37 wk PMA)	39.9	64.3
Term (>37 wk PMA)	46.3	60.1

The dosing in the observed data and the individual clearances estimated were used to calculate an AUC target attainment of 64.2%. The AUC target attainment for proposed new doses were 82.1% for 25 mg/kg q8h in the SCR ≤ 0.25 group, 96.2% for 30 mg/kg q12h in the 0.25 < SCR ≤ 0.5 group, 91.7% for 30 mg/kg q24h in the 0.5 < SCR ≤ 1 group, and 79.2% for 15 mg/kg q24h in the SCR > 1 group. For neonates 10 days of age or younger, the AUC target attainment was 91.3% for extremely preterm neonates, 81.1% for preterm neonates, and 71.8% for term neonates.

DISCUSSION

In this study, we developed a vancomycin population PK model for neonates and infants, which we used to determine the optimal starting dose for these patients to increase the probability of attaining an AUC₂₄ of ≥400 and a trough concentration of <20 mg/liter. This was the largest evaluation of vancomycin PK in neonates and infants to our knowledge, with a total of 934 patients. We also found that the model developed in this study best predicts PK of our data compared to the Frymoyer et al., Lo et al., and Zhou et al. models (8–10). Due to the large amount of data available, we also assessed, using our model, when serum creatinine becomes a marker for vancomycin clearance, which, to our knowledge, has not yet been explored using population PK modeling.

The vancomycin data in neonates and infants was characterized well by a one-compartment model. Although vancomycin’s pharmacokinetic profile is generally described using a two-compartment model, the sparse sampling nature of the data potentially did not enable adequate estimation of a two-compartment model. However, the use of a one-compartment model for this patient population is appropriate and has similar predictive ability, as this age group is reported to have a short distribution half-life and long elimination half-life (11). The covariates that were found to be significant in this study were weight, serum creatinine, and postmenstrual age. The use of all three of these covariates has a valid physiological basis, and these covariates have been reported to have a relationship with vancomycin clearance and volume of distribution in previous studies (11). Although maternal serum creatinine is present in the initial days of life, the use of postmenstrual age as an additional covariate on clearance likely accounts for the change in relationship between serum creatinine and clearance at this point of life (2, 12–15).

The mean parameter estimates from our population PK model are similar to estimates from previous studies. We report a mean clearance estimate of 0.237 liters/h and a mean volume of distribution estimate of 2.98 liters for a 3.5-kg patient. These estimates were similar to what was reported by Chen et al. in their study, which developed a population PK model in a similar population of neonates and infants. Their mean estimates were 4.87 liters/h for clearance and 40.7 liters for volume for a 70-kg patient, which, if normalized for a 3.5-kg patient, would result in mean clearance of 0.524 liters/h and mean volume of distribution of 2.04 liters (16).

Using the population PK model, we determined four starting dosing regimens for four different serum creatinine groups that are most likely to reach a target AUC₂₄ of ≥400. Because we found that serum creatinine is not a valid marker for renal function during the first 10 days of life, neonates 10 days old or younger were not included in the simulations to determine these four starting dosing regimens. These doses would provide a substantial improvement in maximizing PTA at the initiation of vancomycin therapy compared to the commonly used Neofax dosing algorithm and local institution dosing practices. These four starting dosing regimens had a higher probability of AUC and trough target attainment than the Neofax recommended dosing (72.7%, 75.2%, 80.1%, and 71.4% versus 35.4%, 45.3%, 42.3%, and 44.6%). Lexicomp is another commonly used database for vancomycin dosing. Although a formal comparison was not performed here, Lexicomp’s dosing recommendations, based on gestational age and serum creatinine, range from a total daily dose of 7.5 mg/kg to 30 mg/kg, which is much lower than our recommendations (4). Based on the AUC₂₄ target attainments seen in Table 5, opting for the lower dosing would result in failure to meet the therapeutic target. The probability of AUC target attainment with the four proposed doses was also higher than the percent AUC target attainment observed in the patient data collected from Texas Children’s Hospital (82.1%, 96.2%, 91.7%, and 79.2% versus 64.2%).

Chen et al. previously performed a similar study in this age group and developed an AUC-guided dosing algorithm that was binned by postmenstrual age (16). The patients in Chen et al. had PMA ranging from 28 weeks to 48 weeks. In this study, we had a larger sample size of patients, and we developed a dosing algorithm that works for all ages in our population, from 21 weeks to 67 weeks PMA, and was instead binned by serum creatinine. This results in a more simplified dosing algorithm to initialize vancomycin therapy in neonates and infants. In addition, the doses found to achieve the best targets in this study (25 mg/kg every 8 h, 30 mg/kg every 12 h, 30 mg/kg every 24 h, and 15 mg/kg every 24 h) indicate that a more frequent interval of 6 to 8 h is not always necessary. With the exception of 25 mg/kg every 8 h, three out of the four doses are less than or equal to the upper range of 60 mg/kg total daily vancomycin dose that has been frequently reported. Although it is higher than the typically reported upper range, 25 mg/kg every 8 h could be necessary for patients with SCR of ≤0.25 mg/dL, as these patients have higher individual clearance values. In addition, this higher total daily dosage still results in 90% of the trough levels being less than 20 mg/liter. The remaining three dosage regimens indicate that the total daily dose upper limit of 60 mg/kg does not necessarily need to be divided into frequent every 6 h or every 8 h dosing, which could lead to higher troughs and therefore increased risk of toxicity. As suggested by the standard deviation of the troughs from the simulations of the four optimal doses, the vancomycin trough may not necessarily need to be above 10 mg/liter. One standard deviation below the mean trough for all four doses goes below 10 mg/liter and as low as 5.8 mg/liter. This is further illustrated by Fig. 4, in which the AUC₂₄ versus trough relationship is shown. An AUC₂₄ slightly above 400 is associated with a range of troughs, often less than 10 mg/liter. Therefore, a trough target should be used more as a tool to avoid toxicity than to achieve efficacy.

FIG 4.

AUC₂₄ versus trough concentration faceted by dosing interval for simulated patients dosed based on proposed dosing.

Ideally, an AUC target should be used primarily to monitor efficacy of vancomycin therapy. However, many institutions have used trough-based dosing mainly as a surrogate for AUC, because for many years direct measurement of AUC has been more cumbersome to achieve. There has been a shift from trough-based dosing to AUC-based dosing in recent years due to the availability of more Bayesian dose optimization software or, in the case of some institutions, more willingness to adapt to a first-order pharmacokinetic equation-based method (17–19). Although the future of vancomycin dosing points in that direction, institutions that still monitor vancomycin therapy based only on trough need to employ a dosing algorithm that maximizes the probability of attaining an AUC₂₄ of ≥400. With the population PK analysis in this study, we offer a dosing algorithm that maximizes the probability of attaining an AUC₂₄ of ≥400 in the neonate and infant population.

To further aid the optimization of vancomycin dosing, we studied the trends of individual clearance estimates from the final model versus serum creatinine to understand when serum creatinine becomes a valid marker for renal function in the neonate population. Vancomycin is almost exclusively cleared by renal elimination, so relating serum creatinine to individual estimates of vancomycin clearance would act as a surrogate for evaluating renal function. Visual inspection of the clearance versus serum creatinine plot, faceted by quartiles of gestational age and postnatal age, showcased a clear distinction between the first PNA group of 1 to 10 days and the remaining PNA groups, indicating that during the first 10 days of life is when serum creatinine does not act as a valid marker for renal function. Newborns have been reported to have higher serum creatinine levels at birth with no regard to change in renal function (2, 12). Elevated levels at birth were also observed in this study, with more patients having serum creatinine greater than 1.5 mg/dL for PNA group 1. This increased serum creatinine at birth has been attributed to the transfer of maternal creatinine to the fetus at birth (13).

This significant amount of developmental change occurring during these first 10 days of life may explain why the three starting dosing regimens we determined to reach highest probability of target attainment in this population have lower probability of target attainment than the SCR-based dosing regimens we recommend for older neonates and infants. Nevertheless, these three starting doses also had higher probability of AUC target attainment than the probability of AUC target attainment observed in the data (88.6%, 79.5%, and 71.8% versus 64.2%). Although we found SCR to not be a clear marker for renal function, some elevated SCR levels may still be associated with acute kidney injury. Hence, an SCR cutoff of  >1.5 was used to err on the side of caution for these dosing recommendations proposed for neonates within the first 10 days of life.

A limitation of our study is that the vancomycin concentration data collected were sparse, with most patients having only 1 to 3 samples. However, the wide range of time points at which the concentrations were collected after dose and the large number of patients in our study allowed for good precision of our model. In addition, additional work was done to evaluate the impact of the single sampled subjects on the analysis (see Table S1 in the supplemental material). The analysis was redone to include an internal validation, where 80% of the data (458 patients with >1 level and 289 patients with only 1 level, for a total of 747 patients) was used for model building, and the remaining 20% (187 patients with one level) was used as a test data set. The model estimation using the training data set resulted in very similar PK parameter estimates as the parameter estimates when performing estimation using all 934 patients (Table S1). The shrinkage did not change much from using less of the single-sampled subjects in the model estimation step. The shrinkage from analysis using data from all 934 patients was 19.4% for CL and 57.9% for V, and these values dropped to 17.7% for CL and 54.1% for V with model estimation using only the training data set of 747 patients. When model estimates from the training data were used to perform simulations to predict the test data, the median prediction error was 12.6% and median absolute prediction error was 30.1%. These values were similar to the validation done using all data (Table 3), which resulted in median prediction error of −13.9% and median absolute prediction error of 30.9%. Therefore, the entirety of the data was utilized for this work.

Another limitation of our study is that the data were collected from retrospective review, which could be associated with inaccuracy in records as well as confounding factors that cannot be controlled for. However, this type of data is reflective of the clinical practice for which the described dosing algorithms are being proposed. This study is also based on data from a single center. To ensure generalizability, the current study model was validated using goodness-of-fit plots and comparison of predictive performance to other models. Although the final analysis did not include a split of the data to perform internal validation, a separate analysis that had an 80-20 split resulted in similar model estimates and had good predictive performance on the test data set. However, further external validation of the data is warranted to fully ensure generalizability of the model and simulations.

In conclusion, a vancomycin population pharmacokinetic model was developed for pediatric patients ≤180 days of age that has a greater probability of attainment of an AUC₂₄ of ≥400 and trough of ≤20 mg/liter than current dosing regimens. Serum creatinine is not an effective marker of vancomycin clearance in patients ≤10 days postnatal age.

MATERIALS AND METHODS

A retrospective, population pharmacokinetic modeling study was developed. The hospital electronic medical record was queried from 1 January 2011 to 30 August 2018 to identify patients and collect data. Inclusion criteria consisted of hospitalized pediatric patients ≤180 days of age who received intravenous vancomycin and had one or more serum vancomycin concentrations sampled. Patients were excluded if they were missing a weight measurement, did not have a height measured within 1 month of vancomycin initiation, or were undergoing renal replacement therapy or mechanical circulatory support at the time of vancomycin therapy. Vancomycin serum concentrations were excluded if they were below the limit of quantification (<5 mg/liter), were sampled during the time of infusion, were values outside physiologic possibility, or did not have a serum creatinine measured within 48 h of the serum concentrations.

The following patient data were collected for model development: vancomycin dose (milligrams), vancomycin serum concentrations (milligrams/liter), timing of dose and serum concentrations, postnatal age (PNA), gestational age (GA), serum creatinine, sex, weight, and height. Postmenstrual age (PMA) was calculated by adding gestational age to postnatal age. All of the above-described collected covariate data were time-varying, meaning the covariates did not remain constant over the course of each subject’s therapy.

Vancomycin serum concentrations were collected in either a 1-by-0.6-mL amber microtainer with Gel or 1-by-1-mL red/black serum separator vacutainer. The vancomycin assay was performed by using the VITROS Chemistry Products VANC reagent in conjunction with the VITROS Chemistry Products Calibrator kit 11 on the VITROS 5600 integrated system (Ortho Clinical Diagnostics, Raritan, NJ). The assay is based on competition between vancomycin in the sample and vancomycin labeled with glucose-6-phosphate dehydrogenase (G6P-DH) for antibody binding sites. Activity of G6P-DH decreases upon binding to the antibody; therefore, vancomycin concentration in the sample can be measured in terms of G6P-DH activity. The analytic measurement range was 5 to 50 mg/liter. The coefficient of variation was <6%.

Population pharmacokinetic modeling.

Population PK analysis was performed with Pumas v1.0.5 (www.pumas.ai) using the first-order conditional estimation (FOCE) method (20). Data exploration and visual inspection of results were done in R (v 3.4.4) software. Both one-compartment and two-compartment models were tested for the structural model along with an additive, multiplicative error model and combined additive and proportional error models. The covariate selection process was done by plotting potential covariates against parameter estimates and η estimates obtained from the base model. If a relationship between a covariate and a parameter was found, the covariate would be included in the model if the covariate effect on the given parameter had valid physiological reasoning and if including the covariate resulted in a significant objective function value (OFV) drop as well as significant change in between-subject variability of the parameter.

Model evaluation.

The goodness of fit of the model was evaluated using the following diagnostic plots: observed versus population-predicted concentrations (DV versus PRED), observed versus individual predicted concentrations (DV versus IPRED), conditional weighted residuals versus time (CWRES versus TIME), and conditional weighted residuals versus population predicted concentrations (CWRES versus PRED). Prediction-corrected visual predictive check was also used to evaluate the model. The model was further evaluated by first using the final model estimates to simulate the predicted concentration profiles for each patient and then performing these simulations using previously published vancomycin models for a similar population to assess how well the current study model performs compared to other models (8–10). Model bias and precision were calculated using the predicted concentrations from internal validation and the observed concentrations as shown: prediction error (bias) is (predicted – observed)/(observed); absolute prediction error (precision) is (|predicted – observed|)/(observed).

Evaluation of clearance versus serum creatinine.

The final covariate model was used to estimate individual pharmacokinetic parameters using the maximum a posteriori (MAP) probability Bayesian approach. The individual clearances estimated using MAP were used to explore the relationship between clearance and serum creatinine. Serum creatinine was plotted against clearance within different PMA and PNA quantile groups to observe at what age group a clear relationship between serum creatinine and clearance became apparent.

Simulation.

The final population parameters and estimates between-subject variability from the final covariate model were used to conduct simulations using the R package mrgsolve (v 0.9.2) to evaluate the optimal dosing regimens for different covariate groups. The original data were randomly sampled to obtain 4,000 virtual patients with covariate values that would act as the patient population being simulated. If a certain age group was found to not have a clear relationship between serum creatinine and clearance, covariate values from this patient population were not included in the 4,000 virtual patients. The simulated patients were grouped based on serum creatinine from the patient population. Optimal dosing regimens were determined for each group by evaluating different dosing regimens and assessing the probability of target attainment of an AUC₂₄ of ≥400 and trough of ≤20 mg/liter. The doses explored included 8, 10, 12, 14, 15, 20, 25, and 30 mg/kg administered every 6, 8, 12, or 24 h. Simulations were performed separately for the age group without relationship between serum creatinine and clearance and were assessed for the same AUC₂₄ and trough probability of target attainment as the serum creatinine-grouped simulations.

In addition to evaluating new dosing regimens to identify optimal dose, dosing recommended by Neofax, an online evidence-based database commonly used to guide dosing in the clinical setting, was also evaluated by assessing the probability of target attainment of an AUC of  ≥400 and trough of ≤20.

The starting dosing regimen used in the original data set was also evaluated for percent AUC₂₄ target attainment using the individual clearance estimated for each patient using the MAP probability Bayesian approach. The percent target attainment from the original data was compared to the probability of AUC₂₄ target attainment determined from the simulations to assess each dosing regimen.