Using Real-World Data to Externally Evaluate Population Pharmacokinetic Models of Dexmedetomidine in Children and Infants

Abstract

Dexmedetomidine is a sedative used in both adults and off-label in children with considerable reported pharmacokinetic (PK) inter-individual variability affecting drug exposure across populations. Several published models describe the population pharmacokinetics of dexmedetomidine in neonates, infants, children, and adolescents, though very few have been externally evaluated. A prospective PK dataset of dexmedetomidine plasma concentrations in children and young adults aged 0.01 to 19.9 years was collected as part of a multicenter opportunistic PK study. A PubMed search of studies reporting dexmedetomidine PK identified 5 population pharmacokinetic models developed with data from demographically similar children that were selected for external validation. A total of 168 plasma concentrations from 102 children were compared with both population (PRED) and individualized (IPRED) predicted values from each of the 5 published models by quantitative and visual analyses using NONMEM (v7.3) and R (v4.1.3). Mean percent prediction errors from observed values ranged from −1% to 120% for PRED, and −24% to 60% for IPRED. The model by James et al., which was developed using similar ‘real-world’ data, nearly met generalizability criteria from IPRED predictions. Other models developed using clinical trial data may have been limited by inclusion/exclusion criteria and a less racially diverse population than this study’s opportunistic dataset. The James model may represent a useful, but limited tool for model-informed dosing of hospitalized children.

INTRODUCTION

Dexmedetomidine is an imidazole α−2 adrenoceptor agonist used for sedation. Due to its strong affinity for the α−2 subtype of α adrenoceptors, it induces a sedative effect more similar to natural sleep than the predecessor drug clonidine.¹ The drug is approved as an intravenous infusion administered for the duration of required sedation (not indicated for infusions longer than 24h). During infusion, dexmedetomidine distributes to tissues rapidly resulting in quickly observed effects.^2,3 Metabolic transformation of dexmedetomidine occurs primarily by N-glucuronidation and hydroxylation by cytochrome P450 2A6 enzyme in the liver, resulting in a short half-life between 6 minutes and 2 hours in adults.² Very little unchanged drug is excreted in urine or feces.² Rapid clearance after the end of infusion results in swift recovery from dexmedetomidine sedation in both adults and children.⁴ In adults, approximately 94% of administered dexmedetomidine was observed to be bound to plasma proteins.² Dexmedetomidine maintenance dose may vary by individual depending on observed sedation level after initial dosing (between 0.2 and 1 μg/kg/hr).² In adults, covariates reported to significantly influence pharmacokinetics (PK) include weight, height, cardiac output, age, serum albumin, fat-free mass, alanine aminotransferase, and liver blood flow.^3–13

While dexmedetomidine is only cleared for marketing by the United States Food & Drug Administration to persons 18 years of age or older, it has been reported to be safe in children when administered for up to 24 hours.¹⁴ Published analyses of dexmedetomidine pharmacology have shown that PK in children varies based on age, reaching similar weight-normalized clearance values to those reported in adults at around 6 years of age.⁴ The elimination pathway undergoes maturation during childhood, which has been addressed in most models including infants and children, by inclusion of age as a maturation factor affecting clearance.⁴ Based on the published parameter estimates for these maturation functions, we expect drug clearance for infants to be reduced relative to adults. While dosing is often titrated at an individual level to confer appropriate sedation, including use of model-assisted target-controlled infusion pumps, initial recommended infusion rates for infants are much lower than adults per kilogram to account for the change in clearance.⁴

Several studies have described the population PK of dexmedetomidine in children.^15–36 Nearly all pediatric dexmedetomidine PK models relied on internal validation only for development. Previous external validation of pediatric dexmedetomidine models focused on application of adjustment factors on clearance to better describe infants receiving extra-corporeal membrane oxygenation using previously published models, and did not include a large number of children for validation (N = 8).³⁷ Instead, the purpose of this external validation of pediatric models, including children and adolescents (and potentially adults and/or infants), is to evaluate the generalizability of published models to the children represented in our dataset to better inform potential model-informed dosing in this population.

METHODS

Ethics

This study was approved by the institutional review boards of Duke University (coordinating center) at all participating sites and was carried out in concordance with the International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use Guidelines for Good Clinical Practice. Written informed consent (and assent when applicable) was obtained from the legal guardians of all study participants in both studies. The PK samples used for the external validation of population pharmacokinetic (PopPK) models described in this report were collected from the study entitled Pharmacokinetics of Understudied Drugs Administered to Children per Standard of Care (POP01) (NICHD-2011-POP01, ClinicalTrials.gov #NCT01431326; IND #: 21–038). The names and locations of all study sites are presented in Table S1.

Participant Population

Children who received dexmedetomidine per standard of care (SOC) as administered by their treating provider were eligible for enrollment. Participants were recruited from 14 participating hospitals (median enrollment of 5, range 1–21). Exclusion criteria included known pregnancy, as determined by interview or testing, if available. Pharmacokinetic samples were collected per SOC. Participants receiving dexmedetomidine were enrolled in the study for up to 90 days.

Analytical Methods

Blood was collected (enough to provide at least 200 μL of plasma) in an EDTA-K2 Microtainer and was processed into plasma immediately prior to freezing at the study sites. Pharmacokinetic samples were sent to the Frontage Labs (Exton, PA, USA) for storage and analysis. Dexmedetomidine concentrations in plasma were quantified using a validated high-performance liquid chromatography-tandem mass spectroscopy assay. The chromatography system and mass spectrometer used for sample analysis were the AB Sciex Triple Quad 6500+ with Shimadzu HPLC pump and autosampler. Accuracy and precision were assessed using 4 determinations at theoretical levels 5, 15, 1500, and 3750 pg/mL were within the United States Food & Drug Administration bioanalytical assay validation criteria (e.g., ±15%). The lower limit of quantification for dexmedetomidine was 5 pg/mL and the upper limit without dilution was 5000 pg/mL.

PK Dataset

The PK analysis dataset was generated and formatted by The Emmes Company, LLC (Rockville, MD, USA). Clinical, dosing, and drug concentration information was included in the dataset. Due to the very low number of samples below the lower limit of quantification (BQL), BQL concentrations were treated as missing. Observations for each individual within the generated dataset were censored if dosing or observed plasma concentrations were considered unrealistically high. All censored values within the dataset were confirmed to be accurate according to site records, however due to the observational nature of data collection for this study any outlier dosing or plasma concentration values may represent participants with unique clinical statuses that are not comparable to the participants from the studies being evaluated. Additionally, outlier records may represent inadvertent changes to units (i.e. micrograms instead of milligrams) that were not captured during data entry. To evaluate an appropriate cutoff for censoring, dosing rates (in μg/kg/hr) with an infusion duration > 10 minutes were graphically evaluated for outliers, and a cutoff of 6.5 μg/kg/hr was applied where any observation following a dose above that value would be censored for an individual. For observed plasma concentrations, the upper limit of quantification from the assay without dilution (5000 pg/mL) was used as the cutoff for censoring any observations from an individual. The observed plasma concentration cutoff is based on a reported tendency for potential artificial outlier dexmedetomidine values in the event of contamination of collected samples with dexmedetomidine dosing solution.³⁸

Statistical Analysis

Descriptive statistics for demographic and dosing variables were calculated using the value at the time of first PK sample. Median and range values were calculated for all continuous demographic characteristics of the external validation population, including dosing variables, while count and percentage were used to describe categorical variables such as race and sex. Demographic data was stratified by post-natal age. The range of continuous variables denotes the minimum and maximum values in the data set. Demographics were compared as available between published models and the external validation dataset by general summary information. With the exception of the PK modeling, all statistical analyses were performed using the software R (version 4.1.3), RStudio (version 2022.02.1), and R packages tidyverse (v.1.3.1), xpose4 (v.4.7.1), and ggpubr (v.0.6.0).

External Validation of Population PK Models

A literature search of PubMed used search terms listed in Table S2. A separate search using Google Scholar was conducted with the search terms “dexmedetomidine population model pediatric” and “dexmedetomidine population PK model”. Abstracts of published articles were reviewed for inclusion of a parametric population PK model using pediatric plasma concentrations of dexmedetomidine after intravenous infusion. Models included in the analysis were chosen based on demographic similarity to the external validation dataset and availability of reported covariates. This process is shown in Figure S1.

Predictions of dexmedetomidine PK in the external validation population using each published model were conducted in NONMEM (v7.5). Published model structures and parameter estimates were recaptured into model codes for NONMEM (Figure S2). Pharmacokinetic predictions of these models provided sets of concentration-time data to compare with observed values. All model parameters were fixed to published values. Including MAXEVAL=0 in NONMEM under the $ESTIMATION block, PRED was calculated using the demographic and dosing information from each individual in our external dataset. By including the POSTHOC function, an additional set of individualized parameters was generated for each individual by estimation of the random effect(s) or ETA(s) that would provide the best model fit based on the individual’s covariates (such as body weight) and observed plasma concentrations. This secondary dataset is output as IPRED in NONMEM, and represents potential model accuracy when some information is known about an individual’s dexmedetomidine clearance. Quantitative analysis of predictive accuracy was performed in R (v.4.1.3) by comparing the predicted datasets of plasma concentrations (PRED, IPRED) with the observed values in the external validation dataset using Equations 1–4.

$$PE=\frac{1}{N}\sum _{i=1}^N\left(PREDi-OBSi\right)$$ #(1)

$$RMSE=\sqrt{\frac{1}{N}\sum _{i=1}^N{\left(PREDi-OBSi\right)}^2}$$ #(2)

$$MPE=\frac{100}{N}\sum _{i=1}^N\left(\frac{PREDi-OBSi}{OBSi}\right)$$ #(3)

$$MAPE=\frac{100}{N}\sum _{i=1}^N\left(\left|\frac{PREDi-OBSi}{OBSi}\right|\right)$$ #(4)

Where $\mathrm{P}\mathrm{E}$ is the average prediction error of each predicted value $\left(\mathrm{P}\mathrm{R}\mathrm{E}\mathrm{D}\mathrm{i}\right)$ for each observation $\left(\mathrm{O}\mathrm{B}\mathrm{S}\mathrm{i}\right)$ within the external validation dataset. $\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}$: root mean square error, $\mathrm{M}\mathrm{P}\mathrm{E}$: mean percent error, MAPE: mean absolute percent error.

Graphical analysis of predictive accuracy included assessment of model structure and reported population inter-individual variability (IIV). Fixed-effect predictions of dexmedetomidine plasma concentrations were assessed by comparing PRED with observed concentrations (DV) and visualizing any trends in predicted data that would indicate over- or underprediction at high or low concentrations. Conditional weighted residual value (CWRES) was calculated in NONMEM and graphed against PRED to visualize potential trends in prediction error across the concentration range that would not have been apparent in the previous graph. CWRES plots were also explored to determine if the prediction accuracy varied across age, body mass, or sex. Reported population variabilities for published models were compared with variability in the external validation population by prediction-corrected visual predictive check (pcVPC).³⁹ The accuracy of reported variabilities in the pcVPC was assessed by comparison of the lower 5%, median, and upper 95% prediction-corrected POP01 values with the 90% prediction interval for the model output prediction-corrected plasma concentrations. Finally, a graphical and statistical analysis of a large number of prediction datasets (randomized random-effects, N = 1000) was performed to evaluate normalized prediction distribution errors (NPDE) using the npde package for R (v.3.2).⁴⁰

Generalizability of published models was assessed individually for each model. Quantitative analysis values were primarily used to compare across models, but graphical analyses and overall accuracy of predicted plasma concentrations (including proportion of predicted values within 20% and 30% of the observed values: F20 and F30, respectively) were able to provide objective evaluations of the degree of prediction errors within the external population. In this study, certain criteria were evaluated to assess generalizability. To determine model misspecification, the percent of CWRES values within ±2 was calculated for each model in addition to visual inspection of trendlines by PRED. Similarly, the NPDE analysis and pcVPC were used to evaluate bias in the model. Normalized prediction distribution errors outputs are able to identify departures from normality in the distribution of the prediction errors, which can represent biases based on PK parameters or structure and should result in distributions as close to normal as possible to be generalizable to our population. Graphs of pcVPC evaluate distributions of prediction-corrected plasma concentrations to compare population variability represented by the model with that observed in the dataset, where deviations from either the median or 90% prediction interval are identified across the time after last dose to distinguish between structural and variability misspecification. Specific targets for quantitative analyses selected to establish generalizability should lie within arbitrary target ranges suggested by other studies.^41,42

Exploratory Exposure-Response Analyses

The sedative effects of dexmedetomidine were evaluated by comparison of dexmedetomidine exposure as the plasma concentration during an infusion with change from baseline of heart rate (HR), systolic blood pressure (SBP), and diastolic blood pressure (DBP). Concentrations taken during infusions (at least 30 minutes after infusion start and within 5 minutes of infusion end) were considered to be near steady-state, and more likely to represent the value influencing post-dose HR, SBP, and DBP compared with concentrations taken after dosing, either bolus or infusion. Both pre-dose and post-dose measures were taken within 60 minutes of the start and end of dosing, respectively, though these measurements were not available for all doses. Evaluation of change in each pharmacodynamic (PD) measure was conducted by plotting post-dose change in HR, SBP, or DBP against observed concentration during infusion. Any observed trend showing sensitivity of PD measures to dexmedetomidine steady-state concentration would be further evaluated by assessment of published PK/PD model prediction of PD changes in the external validation dataset.¹⁸

RESULTS

Participant Baseline Characteristics from External Validation Population

A total of 108 participants were enrolled in this drug of interest opportunistic study. Participant demographics are shown in Table 1. The median (range) postnatal age was 3.61 years (0.01–19.9) and median (range) body weight (WT) was 16.3 kg (2.66–112). Median (range) serum creatinine (SCR) and albumin (ALB) were 0.41 mg/dL (0.1–2.3) and 2.9 g/dL (1.5–5.3), respectively. Participants received dexmedetomidine via intravenous bolus or infusion. A total of 188 PK samples were collected, representing median per-subject samples of 1 (1–5). Two participants with all associated PK samples (one each) below the limit of quantification were excluded from the analysis. Additionally, 4 individuals (5 total observations) and an extra 13 observations from other individuals were censored due to either missing or excessively high reported dexmedetomidine dosing records or a DV value greater than 5000 pg/mL, leaving a final analysis population of 102 participants with 168 plasma concentrations, as shown in Figure 1. The median (range) infusion rate reported for participants was 0.6 μg/kg/hr (0.04–4.4), while the median (range) for bolus dosing was 0.5 μg/kg (0.02–1.03). Six infusion doses out of 556 reported for the included individuals were below the typical adult minimum of 0.2 μg/kg/hr. The median (range) number of doses recorded per participant was 6 (1–29). Doses were not given at standard intervals, as dexmedetomidine is used for immediate sedation, though many instances of long-term sedation (>24h) were observed. Descriptive statistics were not distinguished between individuals receiving bolus versus infusions, as all but 1 individual received at least 1 infusion dose.

Table 1.

Demographics of the POP01 external evaluation data set. PNA: Postnatal age

Variable	Median (Range) or N (%)
	PNA <2 years	PNA ≥ 2 to ≤ 8 years	PNA >8 years	Total
N	41	26	35	102
Postnatal age (years)	0.35 (0.01–1.86)	4.56 (2.07–7.93)	14.4 (8.34–19.9)	3.61 (0.01–19.9)
Postmenstrual agea (weeks)	58.3 (37.7–137)	277 (148–454)	793 (473–1080)	226 (37.7–1080)
Body weight (kg)	5.70 (2.66–12.8)	17.27 (9.20–33.0)	55.0 (24.0–112)	16.3 (2.66–112)
Body mass index (kg/m2)	15.7 (10.9–25.1)	15.9 (13.5–25.4)	23.4 (13.8–38.6)	17.8 (10.9–38.6)
Body mass index percentile	NAb	57.9 (4.68–99.2)	88.2 (0.05–99.6)	78.5 (0.05–99.6)
Female	20 (48.8)	12 (46.2)	15 (42.9)	47 (46.1)
Serum Creatinine (mg/dL)	0.3 (0.1–1.0)	0.44 (0.19–2.30)	0.60 (0.1–1.95)	0.41 (0.1–2.30)
Serum Albumin (g/dL)	2.8 (2.0–4.1)	3.2 (2.0–5.3)	2.8 (1.5–4.6)	2.9 (1.5–5.3)
Race
White	31 (75.6)	20 (76.9)	25 (71.4)	76 (74.5)
Black or African American	7 (17.1)	2 (7.7)	7 (20)	16 (15.7)
Asian	1 (2.4)	1 (3.8)	1 (2.9)	3 (2.9)
Native American	0 (0)	0 (0)	1 (2.9)	1 (1.0)
Unknown or not reportedc	2 (4.8)	3 (11.5)	1 (2.9)	6 (5.8)
Ethnicity
Hispanic or Latino	5 (12.2)	2 (7.7)	1 (2.9)	8 (7.8)
Not Hispanic or Latino	36 (87.8)	24 (92.3)	34 (97.1)	94 (92.2)
Obese Statusd
With Obesity	NAb	6 (23.1)	11 (31.4)	17 (16.7)
Had Surgerye
Yes	10 (24.4)	5 (19.2)	7 (20.0)	22 (21.6)

^aPostmenstrual age was calculated as PNA + 40 weeks when GA was missing

^bBody mass index percentile is not available for patients <2 years of age

^cIncludes individuals listed as ‘mixed race’

^dObesity defined by BMI > 95^th percentile for age and sex (body mass index percentile > 95)

^eReported as surgery occurring within 24h prior to the first dose of dexmedetomidine

Figure 1.

Scatter plot representation of included observed plasma concentrations after dexmedetomidine dosing. Each black dot represents 1 observed concentration at the time it was taken after the start of the most recent infusion or bolus dose. h, hours.

Population PK Model External Validation

The search of PubMed and Google Scholar (conducted 11/22/2022) identified 22 pediatric PopPK models for review. Of these, 7 were only developed for neonates and/or infants and were not intended for prediction of clearance maturation into childhood.^{21,23,24,28,30,31,32} Two models reported significant covariates (cytochrome P450 1A2 genotype, noradrenaline use, and international normalized ratio [INR] on clearance) that were not available in the external validation dataset.^20,22 Four of the pediatric modeling studies were recaptured into a pooled model by Potts et al.,¹⁵ and were not included individually.^33–36 Finally, 4 studies were conducted in Asian pediatric populations that were not representative of the population included in the external validation dataset,^25,26,27,29 and could influence dexmedetomidine PK due to a higher frequency of cytochrome P450 2A6 gene deletion among Asians.⁴³ Among the 5 included studies (described in Tables 2 and 3), 4 were only intended for PK in children and adolescents^15,16,18,19 while 1 model was intended as a ‘universal’ PK model for estimation of dose-exposure in children and adults.¹⁷ Each of the 4 pediatric-focused models utilized a two-compartment model with allometric scaling by weight for both volume of distribution and clearance. The universal model by Morse et al.¹⁷ utilized a three-compartment model with allometric scaling by fat-free mass on clearance and volume of distribution, and a proportional increase to volume of distribution based on fat mass (based on equations from Al-Sallami et al.).⁴⁴ Four models included a maturation function of age on clearance (represented by Equation 5), while 1 pediatric model by Pérez-Guillé et al. excluded all individuals younger than 2 years and did not include a maturation function.¹⁸

$$\begin{array}{c}TVCL={CL}_{pop}\mathrm{*}\frac{{AGE}^{Hill}}{{AGE}^{Hill}+{TM50}^{Hill}}\end{array}$$ #(5)

Where ${\mathrm{C}\mathrm{L}}_{\mathrm{p}\mathrm{o}\mathrm{p}}$ is the population predicted clearance at a given AGE (post-natal or post-menstrual age). TVCL represents the typical value of clearance after scaling by body weight when AGE ≫TM50, where TM50 is the age at which clearance reaches half maximal. The shape of the maturation curve is described by the exponent Hill value, where a higher number indicates a more rapid transition from immature to mature clearance.

Table 2.

Demographics of populations used for model development in publications chosen for analysis.

Variable	Median (Range) or N (%)
	Potts15	James16	Morse (Five Pooled Datasets)a, 17					Pérez-Guilléb, 18	Wiczling19
Postnatal age (years)	3.83 (0.01–14.4)	1.3 (0.003–22.6)	3.83 (0.01–14.4)	20–70	38 (18–60)	42.3 (23–59)	29 (21–36)	11 (5)	5.83 (0.12–15.7)
Weight (kg)	16.1 (3.1–58.9)	9.4 (2–138)	16.1 (3.1–58.9)	51–110	95 (59–152)	90 (47–126)	72 (52–89)	43 (19)	18.5 (4.7–60)
Female	NR (N = 95)	171 (48%)	NR (N = 95)	NR (N = 18)	14 (35%)	26 (65%)	5 (50%)	21 (70%)	15 (39.5%)
Infusion Dose (μg/kg/hr)	0.2	0.6 (0.03–2.0)	0.2	6	0.25–0.5	0.5	1	2.4–4.2c	0.8–1.4
Bolus Dose (μg/kg)	1–6	1.0 (0.06–4.21)	1–6	NA	NA	NA	NA	NA	NA

NR: Not reported, NA: Dosing type not included in analysis

^aFrom left to right, the pooled datasets are from studies by Potts et al.¹⁵, Hannivoort et al.⁷, Cortinez et al.⁴⁵, Rolle et al.¹³, and Talke et al.⁴⁶. Only one study included pediatric data, though the model is described as a ‘Universal Population PK Model’, indicating it is intended for use in children as well as adults.

^bReported as mean (SD), age range reported to be limited to 2–18 years old

^cGiven as 10–15 minute infusions only

Table 3.

Population PK Parameter Values Reported by Models Chosen for Analysis

Variable	Median (range) or N (%)
	Potts15	James16	Morse17	Pérez-Guillé18	Wiczling19
CL (L/h/70kg)	42.1	27.3	53.8	20.8	41.6
V1 (L/70kg)	56.3	161	25.2	21.9	52
Q,Q2 (L/h/70kg)	78.3	26.0	100.8, 37.2	75.8	56.8
V2,V3 (L/70kg)	69	7903	34.4, 65.4	81.2	70.4
IIV (%CV)a
CL, Q, Q2	30.9, 37.0	103, 82	35.8, 63.2, 89.7b	27.5, 25.3	30.9, 37
V1, V2, V3	61.3, 47.0	138, 624	103.9, 41.8, 61.6b	20.2, 21.8	61.3, 47
Residual errora
Additive	0 Fixed	0 Fixed	0.004 (μg/mL)	0 Fixed	0 Fixedc
Proportional (%CV)	16.2	50.5	19	14.2	0 Fixedc
Covariates
Covariate	Age on CL Surgery on CLd	PMA on CL	Age on CL Fat mass on V	NA	Age on CL
Effect size	TM50=44.5 Hill = 2.56 Finf = 0.73	TM50=41.9 Hill = 7.04	TM50=52.4 Hill = 1 Fixed Ffat = 0.293	NA	TM50=44.5 Hill = 2.45

^aAll IIV and Residual Error values (as %CV unless otherwise noted) shown as reported in each publication.

^bReported as population parameter variability percent or PPV%.

^cModel utilized an exponential error model where EPS = 0.1.

^dWhere F_inf represents the proportion of pre-surgery clearance that is observed for dexmedetomidine infusions after surgery within the study populations.

Cl, clearance; CV, coefficient of variation; EPS, epsilon or the variance of the normally distributed residual error values; F_fat, fractional influence of fat mass on volume of distribution compared to lean body weight; F_inf, proportional clearance observed within 24 hours after surgery; Hill, exponent describing the slope of the maturation curve; IIV, Inter-individual variability; PPV%, population parameter variability percent; PMA, Post-menstrual age; Q, inter-compartmental clearance; TM50, age in weeks at 50% of adult clearance; V, Volume of distribution.

Demographic summaries of individuals included in the development of each published model are shown in Table 2. Published model parameters and covariate effects are shown in Table 3. Quantitative analysis of predictive accuracy for each published model is shown in Table 4 and Figure 2. Assessment of trends in prediction errors are shown in Figures 3 and 4. Visualizations of predicted dexmedetomidine concentrations in the external population using reported model IIV are shown as pcVPCs in Figure S3.

Table 4.

Quantitative Analysis of Published Model Predictive Accuracy

Model	F20 (% of total)		F30 (% of total)
	PRED	IPRED	PRED	IPRED
Potts15	17.9	44.0	29.2	58.3
James16	26.2	68.5	34.5	78.6
Morse17	18.5	63.1	28.0	75.0
Perez-Guillé18	8.9	29.8	15.5	39.3
Wiczling19	4.8	22.6	11.9	36.9

F20, proportion of simulated values within 20% of the observed values; F30, proportion of simulated values within 30% of the observed values; IPRED, individualized predicted values; PRED, population predicted values.

Figure 2.

Comparison of quantitative evaluations of prediction error in pg/mL (PE), mean percent error (MPE), mean absolute percent error (MAPE), and root mean square error (RMSE) across published models. Dashed lines represent criteria for acceptable model predictions, which are 15% for MPE and 30% for MAPE.⁴¹ PRED, population predicted values; POSTHOC, individualized predicted values.

Figure 3.

Graphical assessment of population predicted values (PRED) against observations. Visualized trends for PRED versus observed concentrations (DV) are shown in red (geom_smooth, method = “lm”). The black line is the line of identity for DV. Conc., concentration.

Figure 4.

Graphical assessment of conditional weighted residual error (CWRES). Visualized conditional weighted residual errors (CWRES) between population predicted values (Pred.) and observed concentrations (DV). Three outliers (> 40) were removed for all graphs for ease of visualization. Black circles represent CWRES values. Black, dashed lines represent ± 2 on the y-axis. Red dashed lines represent a linear regression by geom_smooth function (method = “lm”).

A final statistical and visual assessment was conducted to determine if model prediction errors demonstrated normal distribution. Mean, variance, skewness, and kurtosis of prediction errors are shown in Table S3, while statistical test p values are shown in Table S4. Three models (James et al., Pérez-Guillé et al., and Wiczling et al.) demonstrated significant departures from normality through the t test, and all models did so from the Shapiro-Wilk test. Only the model from James et al. provided a p value for the Fisher variance test above 0.05. This suggests that while 2 of models predicted errors centered around 0, they demonstrated a larger, flatter spread of values than what would be considered a normal distribution. On the other hand, the model by James et al. predicted a distribution closer to normal, but precisely centered around 0.38 instead of 0. Graphical representations of error distributions are shown in Figure S4. A plot of prediction errors from the model by James et al.¹⁶ are shown in Figure S5.

Exploratory Exposure-Response Analyses

Dexmedetomidine concentrations were compared to changes in multiple PD measures after dosing. Observations included in the external validation process were used to establish steady-state concentrations for the PD analysis. The number of reported doses that included steady-state concentrations and measured pre- and post-dose values for HR, SBP, and DBP were 17 each. This data was not available for all participants, or for all doses for included participants. The majority (70–77%, not shown) of recorded PD measures were taken within 20 minutes of the start or end of the dose. When exposure was compared to changes in HR, SBP, and DBP, trends were observed for each (Figure S5). A minor positive result for HR, and similar negative trends for SBP and DBP. The regressions for SBP and DBP may have been strongly influenced by solitary concentrations near 4000 pg/mL. Due to the minimal amount of PD measures at high concentrations, and the lack of observable trends in the data at lower concentrations, no PK/PD fitting or external validation of the data was conducted.

DISCUSSION

Dexmedetomidine is prescribed in children for sedation. Given that the intended use of dexmedetomidine is for the duration of anesthesia and is generally indicated for less than 24 hours, there is limited opportunity for dose adjustments should the initial dose cause over- or under-sedation. Instead, PK models can be utilized to predict dose-exposure in a given population, including one with high PK IIV such as children.⁴ While limited, there have been several reports regarding dexmedetomidine disposition and dosing in the pediatric population, including population PK models.^15–36 To determine the appropriateness of using any given model to predict dose-exposure in the local population receiving dexmedetomidine, external validation was conducted for only those relevant to the represented demographics in our “real-world” clinical dataset. In the present study, dexmedetomidine plasma concentrations from children (0.01 to 19.9 years of age) were used to externally validate relevant published PK models of dexmedetomidine and determine generalizability to our pediatric population. The included models are varied in their structure, parameter estimates, covariate inclusions, and modeling methodologies, though each of them were developed in similar pediatric populations to the external validation dataset.

The predictive accuracy of the models was evaluated objectively with overall prediction accuracy (F20 and F30) and graphical analysis as well as comparatively using quantitative analysis of prediction errors without and with random-effects (PRED and IPRED, respectively). While no formal standard has been established for external validation analysis values that would indicate acceptability, a recent publication by Zhao et al.⁴¹ arbitrarily set some standards with which to evaluate their own models. These include median PE percent (MDPE) ≤ 15%, mean absolute percent error (MAPE) ≤ 30%, F20 > 35%, and F30 > 50%. One study conducted an external validation in neonates using these standards, and found that 1 tested model was able to meet the F20 and F30 criteria, which provides some confidence in using these as a baseline for acceptability.⁴²

The models from Potts et al.¹⁵, James et al.¹⁶, and Morse et al.¹⁷ showed comparable accuracy from PRED predictions, though IPRED analysis slightly favored predictions from James et al. and Morse et al. over the other three. Overall prediction accuracy and graphical analyses tended to favor the model from James et al., including over 68% of IPRED predictions within 20% of observed, over 90% of CWRES values within ±2 and an improved trendline across PRED values, and accurate prediction of median plasma concentrations in the PC-VPC plot. Among all models, no trends were observed for predictive accuracy when considering age, BMI, or sex (not shown). The 2 models that did not result in comparable PRED values were likely limited by their lack of inclusion of infants to the validation dataset (Pérez-Guillé et al.¹⁸) and narrow dose range (Wiczling et al.¹⁹). No model was able to meet the model evaluation criteria above for all values (e.g., MDPE/MAPE/F20/F30), though when individualized clearance and volume of distribution were estimated (IPRED), the model by James et al. was very close, where only MDPE (MPE) was out of range at 16.3% (>15%).

Further investigation into the prediction errors of the model from James et al.¹⁶ demonstrated a number of points that were not well predicted, including some apparent bias in the mean error from predicted concentrations used for NPDE analysis. While this model was developed using similar real-world data, it is important to note that the external validation dataset was collected prospectively by opportunistic collection during standard of care dexmedetomidine administration, which is a different sampling method than the one reported by James et al. (retrospective data and scavenged samples). The reported model parameters describe a two-compartment model with large IIV on clearance and volume of distribution, along with a very rapid maturation of clearance based on the high reported Hill value (>7) for the maturation function. This resulted in a reasonable approximation of the variability observed in the external validation dataset (Fisher test, pcVPC), but there is potential for extreme over- or underpredictions at moderate-to-high plasma concentrations. While there is no clear demographic factor influencing model predictive accuracy, the model appears to most accurately predict concentrations below 1000 pg/mL within 100 hours of the first recorded dose. External validation of PK models should be considered a first step in understanding potential knowledge gaps in similar populations of critically ill children. Before these models are deployed to guide individualized dosing, additional external datasets can continue to improve understanding of potential patterns in predictive errors as well as build up available data sources for possible re-estimation of PK parameter values, influential covariate effects, and population IIV, that can increase confidence in the capacity for a model to predict individual PK.

In exploratory PD analyses, there appeared to be a lack of sensitivity to dexmedetomidine exposure for HR, SBP, and DBP at low-to-moderate steady-state plasma concentrations within the external validation population. This may be due to the varying clinical environments and interventions across this “real-world” dataset that could not be accounted for as significant covariates in the PK/PD model by Pérez-Guillé et al.¹⁸ Additionally, less than 20% of administered doses included both predose and post-dose measurements of HR, SBP, or DBP, with no obvious trends of missingness by infusion duration or dose rate (not shown). A large number of individuals contributed at least 1 PD observation, though there may be potential confounding factors influencing measurement availability. The available PD data appears generally in line with reported lack of sensitivity for SBP and DBP to dexmedetomidine infusion, though the effects on HR are less clear.⁴⁷

The study is limited by the overall sparseness of the data which prevented development of a de novo model with the given external validation dataset alone. Additionally, the observational nature of the study collecting data from participants receiving dexmedetomidine per SOC may result in dosage and duration of drug administration and data collection to be dependent on individual tolerance and response, which could introduce a selection bias. While several published models of dexmedetomidine PK in children cover similar demographics, the dose range in our dataset was not represented in any publication, which could confound prediction for individuals at the higher dose levels if dexmedetomidine PK is nonlinear in children at this unstudied range. Published model recapture required translation of multiple models to NONMEM code from either MLXTran, WinBUGS, or modeling code that was not available aside from details within the relevant publication, which can be associated with minor misspecifications and differences to random effect estimation during IPRED calculation. Recapture of the published models in NONMEM code was based on published or received model codes written for NONMEM, Mlxtran, or WinBUGS. Actual codes used for model development were available for all models except the one from Pérez-Guillé et al., though the model structure was completely described in the relevant publication.¹⁸

Additionally, comparison of real-world data with models developed using clinical trial data is more likely to identify issues with IIV and possibly fixed-effect estimation given the highly heterogeneous nature of this data. For example, the external validation dataset included children with obesity, while it is unclear if models by Potts et al.¹⁵, Pérez-Guillé et al.¹⁸, and Wiczling et al.¹⁹ were validated with similar individuals. Translation of specific parameter effects, especially the effect of surgery on dexmedetomidine clearance from Potts et al., may not be reliable to reported surgeries included in our dataset, though fixed-effect prediction was comparable to other models. Overall, this study represents an important first step to applying these published models. It should be built upon by including additional populations and potentially further steps towards linking clinical decisions and model predictions that were beyond the scope of this preliminary evaluation.

CONCLUSIONS

Dexmedetomidine is used regularly off-label in children for both short and long-term sedation. Published estimates of PK parameters in children vary between PopPK models, which provide opportunities for external validation and comparison of PK predictions in unique pediatric populations. The model published by James et al. had slightly better overall predictive performance, but still indicated bias for our real-world dataset. Nevertheless, IPRED estimates (utilizing both demographics and observed plasma concentrations) from the James model may represent a useful, but limited, tool for incorporation of model-informed dosing or dose adjustments of dexmedetomidine in hospitalized children, including neonates, infants, and adolescents.

Data Sharing and Data Availability

To help expand the knowledge base for pediatric medicine, the PTN is pleased to share data from its completed and published studies with interested investigators. For requests, please contact a PTN Program Manager (PTN-Program- Manager@dm.duke.edu).