Exposure–Response Relationships of Metronidazole in Infants: Integration of Electronic Health Record Data with Population Pharmacokinetic Modeling-Derived Exposure Simulation

Abstract

Background:

Infants frequently receive metronidazole at variable doses and duration for surgical site infection prophylaxis and treatment of intra-abdominal infections. Seizures are a rare (but potentially devastating) side effect of metronidazole, yet the prevalence of seizures in infants, as well as the relationship with metronidazole dose and exposure, are unknown.

Methods:

We examined the Pediatrix Clinical Data Warehouse for infants in neonatal intensive care units from 1997–2018 who received at least one dose of metronidazole during their first 120 days of life. We used an existing population pharmacokinetic model to simulate exposure parameters, estimating multivariable associations between metronidazole dosing and exposure parameters, and the occurrence of seizure.

Results:

There were 19,367 intravenous doses of metronidazole given to 1,546 infants, and 31 experienced a seizure. Infants with a seizure had a longer median (interquartile values) duration of metronidazole exposure than those without (11 days [6, 15] vs. 7 [4, 11], p=0.01). Each added day of metronidazole (OR=1.06, 95% CI [1.02, 1.10]), and each standard deviation increase in cumulative area under the plasma concentration–time curve (OR=1.27, 95% CI [1.11, 1.45]) were associated with increased odds of seizure. Higher simulated maximum plasma concentration was associated with lower odds of seizure (OR=0.88, 95% CI [0.81, 0.96]).

Conclusions:

Longer metronidazole exposure and higher cumulative exposure could be associated with increased odds of infant seizures. Using a large observational dataset allowed us to identify a rare adverse event, but prospective studies are needed to validate this finding and further characterize metronidazole dose- and exposure-safety relationships.

Metronidazole is an antibiotic commonly used in the neonatal population for surgical site infection prophylaxis and treatment of intra-abdominal infections.^1,2 Despite its widespread use, it is not currently approved for children by the United States (U.S.) Food and Drug Administration (FDA), and pharmacokinetic data are sparse.^3,4 The side-effect profile of metronidazole has been well established in adult populations, but there are no more than case reports in the pediatric population.⁵ Elevated incidence of seizure has been reported as a rare, but potentially devastating, adverse side effect^6–8; it is unknown if seizure risk is related to excessive metronidazole exposure or other predisposing factors.

In a neonatal population, studying the safety of a drug, especially exposure-safety, is a major challenge because: 1) large sample sizes are needed to identify rare adverse events; and 2) measuring exposure parameters requires invasive blood draws, which are challenging to complete in infants, due to difficulty in achieving phlebotomy and limited blood volume. An alternative approach to overcome these limitations is to predict drug exposures in sufficiently large populations exposed to the drug to identify rare events.⁹ Such predictions can be performed using population pharmacokinetic (PopPK) models developed with data from smaller PK trials, as is the case for metronidazole; these predictions can then be applied to a larger dataset that lacks exposure measurements.^10–12

When a drug is administered per standard of care, the electronic health record (EHR) and other real-world data sources can be used to identify rare adverse events. Combining these data with PopPK model-based exposure predictions creates a platform for the study of rare adverse events and their relationship with predicted drug exposure. The Pediatrix Clinical Data Warehouse is a national dataset derived from the EHR of more than 1 million infants in intensive care units across the U.S.¹³ These data capture rare adverse events, but lack drug exposure information. In this manuscript, we combined EHR data with exposure predictions from PK modeling to assess the predicted exposure-safety relationship of metronidazole (defined as occurrence of a seizure). We hypothesized that higher maximal metronidazole concentrations and higher cumulative exposure would increase the incidence of seizure.

METHODS

From the Pediatrix Clinical Data Warehouse, we included infants from 1997–2018 who received at least one dose of intravenous metronidazole during their first 120 days of life. We excluded infants with incomplete clinical or dosing data and those who experienced a seizure prior to the start of metronidazole. Our analysis included demographic, clinical, and laboratory data, and metronidazole dose, frequency, route of administration, and course duration. For each infant, we included only the first course of metronidazole received. This was done for model simplicity; of note, there was no overlap of a subsequent course of metronidazole with a new seizure. We defined a bolus dose as either a single dose of metronidazole or one dose of metronidazole at the start of the course that was higher than subsequent doses. We classified all other doses as maintenance doses. Our primary outcome of interest was clinically documented seizure. We categorized seizures as starting during or after the course of metronidazole. Seizures occurring after the course of metronidazole were counted if they started on any day within 5 half-lives of the end of medication (288 hours), to be concordant with the data used to create the original PopPK model.

We performed plasma concentration-time profile simulations of metronidazole in Nonlinear Mixed Effects Modeling software (NONMEM, version 7.4, Icon Solutions, Ellicott City, Maryland) using the PopPK model described below, and demographic (postmenstrual age [PMA] and current weight) and dosing information from the Pediatrix Clinical Data Warehouse. We included between-subject variability in metronidazole clearance and residual variability in simulations. We simulated maximum plasma concentration (C_max), minimum concentration (C_min), area under the plasma concentration–time curve at steady state (AUC_ss), and cumulative AUC (AUC_c) using the following equations for each individual patient in the Pediatrix cohort^3,14:

$$C_{max}=\frac{Dose}{\left({CL}_i\mathrm{*}DUR\right)}\mathrm{*}(1-e^{-K_{e,i}\mathrm{*}DUR})$$

$$C_{min}=C_{max}\mathrm{*}{e}^{-K_{e,i}\mathrm{*}(24-DUR)}$$

$${AUC}_{ss}=\frac{Dose}{{CL}_i}$$

$${AUC}_c={\int }_0^T{C}_{p}dT$$

Where K_e,i denotes the first-order elimination rate constant calculated as clearance (CL_i)/volume of distribution (V_i), DUR denotes the infusion duration (which was assumed to be 0.5 hours for all dosing simulations), C_p is the plasma concentration, and T is the time of event in days after the first dose. Of note, AUC_c included all days of metronidazole administration in addition to 5 half-lives (288 hours) after the last dose of metronidazole. The metronidazole PopPK model we used for exposure simulation was previously developed using data from 488 plasma samples from 143 infants 23–48 weeks PMA receiving intravenous metronidazole in two prospective, open-label, multicenter PK trials.^3,14 Consistent with FDA guidance for PopPK model development, this model has been internally validated using visual predictive checks and bootstrapping.¹⁵ The final irreducible PopPK model included the following relationships, where WT is body weight (kg):

$${CL}_i(\frac{L}{h})=0.035\mathrm{*}WT\mathrm{*}\frac{{PMA}^{15.4}}{{25.5}^{15.4}+{PMA}^{15.4}}$$

$$V_i\left(L\right)=0.864\mathrm{*}{WT}^{0.754}$$

The unit of observation for our analysis was an infant day of exposure to metronidazole, including up to 288 hours after the last dose. We conducted statistical analyses using Stata 16.1 (StataCorp LLC, College Station, Texas). We calculated summary statistics including counts (percentages) and medians (interquartile values [IQV]) to describe categorical and continuous variables, respectively. We used Chi-square or Wilcoxon rank sum tests to compare distributions of categorical and continuous study variables across groups, respectively.

We developed five separate multivariable logistic regression models to evaluate associations with a clinical diagnosis of seizure as the outcome variable. In the first regression, we evaluated the association of seizure for each infant with all metronidazole dosing parameters, including average daily dose, maximum daily dose, and course duration, and the following covariates: gestational age (GA) at birth, the diagnosis of small for GA at birth, postnatal age (PNA) at the start of metronidazole, mechanical ventilation, inotropic support, and presence of positive blood cultures on the day of or day prior to the start of metronidazole. We chose these adjustment covariates because they previously have been associated with seizure.^16,17 Alternative covariates were tested and did not improve model fit, such as interventricular hemorrhage, positive cerebrospinal fluid (CSF) cultures, PMA, among others. Model fit was assessed via the Wald statistic. Next, we developed separate regressions for each of the following three primary simulated exposure parameters on the days with and without seizure: C_max, C_min, and AUC_ss. Our final regression evaluated the association between AUC_c and the development of a seizure at any time during or after the course of metronidazole. All regressions were adjusted for the same covariates described above, which were selected a priori.

Consistent standard errors were obtained via the robust Huber-White method, with clustering by patient in the models using daily data on patients. As a sensitivity check, we also estimated the regressions with fixed and random effects, but this did not improve model performance according to the Wald statistic, so these results are not reported. Estimates are reported as odds ratios (ORs) for one unit change or one standard deviation change in the independent variable, along with 95% confidence intervals (CIs). A p-value of 0.05 or lower was considered significant.

RESULTS

Our final cohort included 1,546 infants from 106 neonatal intensive care units who received a total of 19,367 doses of metronidazole with complete dosing information. The median (IQV) GA was 28 weeks (25, 33), PNA was 17 days (8, 36), birth weight was 1.1 kg (0.77, 1.83), and 896 infants (58%) were male (Table 1).

Table 1.

Infant Characteristics

N (%)	Infants with Seizure, N=31	Infants without Seizure, N=1,515	All Infants, N=1,546
Male	21 (68%)	875 (58%)	896 (58%)
Gestational age (week)*	27 (24, 32)	28 (25, 33)	28 (25, 33)
Postnatal age (day)*	14 (8, 35)	17 (8, 36)	17 (8, 36)
Race/ethnicity
White	13 (42%)	757 (51%)	770 (51%)
Black	5 (16%)	395 (27%)	400 (26%)
Hispanic	12 (39%)	270 (18%)	282 (19%)
Other	1 (3%)	62 (4%)	63 (4%)
Birth weight (kg)*	0.95 (0.68, 1.62)	1.1 (0.77, 1.84)	1.1 (0.77, 1.83)
Small for gestational age	4 (13%)	245 (16%)	249 (16%)
Ventilatory support†	21 (68%)	835 (55%)	856 (55%)
Inotropic support†	13 (42%)	232 (15%)	245 (16%)
Positive blood culture†	5 (16%)	118 (8%)	123 (8%)

^*Median (interquartile value)

^†On the day of or day prior to start of metronidazole

kg, kilogram; N, number

Seizures were rare: 31 infants (2%) were diagnosed with a seizure in the observation period, 20 (65%) during, and 11 (35%) after metronidazole administration. The median (IQV) time from start of metronidazole until seizure diagnosis was 5 days (1, 13). There were 123 infants (8%) with a positive blood culture, but no infants had a positive CSF culture on the day metronidazole was started or the day prior. The most common organisms isolated from blood cultures were Escherichia coli (23%), coagulase-negative Staphylococci (16%), Staphylococcus aureus (10%), and Klebsiella species (9%). The distribution of positive blood cultures did not differ according to whether an infant had a seizure or not (p=0.95).

Dosing-Seizure Relationship

There were 606 infants (39%) who received a bolus dose of metronidazole; the median (IQV) bolus dose was 15 mg/kg (14.5, 15). There were 1,441 infants (93%) who received a maintenance dose of metronidazole; the median (IQV) dose was 7.5 mg/kg (7.1, 7.5) and the most common administration frequencies were every 12 hours (42%), 24 hours (29%), and 48 hours (22%). The dosing parameters did not differ statistically between infants who experienced a seizure and those who did not for average daily dose, maximum daily dose, and cumulative dose (Table 2). However, infants who experienced a seizure had a longer median duration of metronidazole than those without seizure (11 days [6, 15] vs. 7 days [4, 11], p=0.01).

Table 2.

Metronidazole Dosing Data

Median (IQV)	Infants with Seizure N=31	Infants without Seizure N=1,515	p-value
Average dose (mg/kg)	7.3 (4.4, 14.8)	12.2 (7.2, 15.0)	0.10
Maximum dose (mg/kg)	15.0 (7.5, 17.1)	15.0 (7.5, 15.8)	0.71
Cumulative dose (mg/kg)	72.3 (34.7, 163.3)	65.0 (31.4, 120.0)	0.36
Duration of metronidazole (day)	11 (6, 15)	7 (4, 11)	0.01
Bolus dose (mg/kg)	15.0 (14.2, 15.9)	15.0 (14.5, 15.0)	0.02
Maintenance dose (mg/kg)	7.5 (6.5, 7.5)	7.5 (7.1, 7.5)	0.67
Maintenance dosing frequency*			0.51
4 hour	0 (0%)	6 (0.4%)
6 hour	1 (3%)	23 (2%)
8 hour	1 (3%)	71 (5%)
12 hour	10 (32%)	600 (43%)
24 hour	8 (26%)	405 (29%)
48 hour	11 (35%)	305 (22%)

^*Data presented as number (%).

IQV, interquartile value; kg, kilogram; mg, milligram

On multivariable regression, each additional day of metronidazole was associated with increased odds of seizure (OR=1.06, 95% CI [1.02, 1.10]), but maximum and average doses of metronidazole were not (OR=1.05, 95% CI [0.95, 1.16] and OR=0.94, 95% CI [0.85, 1.05], respectively).

Simulated Exposure-Seizure Relationship

When compared to infants without seizures, those who experienced a seizure had lower median (IQV) plasma C_max (9.4 ng/mL [8.0, 11.4] vs. 9.8 ng/mL [8.6, 11.1], p=0.002) and C_min (4.6 ng/mL [3.4, 7.4] vs. 5.7 ng/mL [3.5, 6.2], p<0.001). However, AUC_ss was higher in infants with seizure (221.9 [ng*hr /mL] [196.4, 431.8] vs. 217.0 [ng*hr /mL] [208.0, 250.3], p=0.008), and AUC_c was similar (3923.6 [ng*hr /mL] [1846.5, 7293.8] vs. 2445.5 [ng*hr /mL] [1327.1, 4140.7], p=0.10) (Table 3).

Table 3.

Simulated Metronidazole Exposure Parameters

Parameter	Infants with Seizure, N=31	Infants without Seizure, N=1,515	p-value
Cmax (ng/mL)
All doses	9.4 (8.0, 11.4)	9.8 (8.6, 11.1)	0.002
Bolus doses	17.4 (16.4, 19.9)	18.7 (16.6, 20.3)	0.24
Maintenance doses	9.4 (8.0, 11.4)	9.7 (8.6, 10.9)	<0.001
Cmin (ng/mL)
All doses	4.6 (3.4, 7.4)	5.7 (3.5, 6.2)	<0.001
Bolus doses	8.5 (6.4, 11.3)	7.1 (6.4, 11.4)	0.24
Maintenance doses	4.4 (3.4, 6.1)	5.7 (3.4, 6.2)	<0.001
AUCss (ng*hr /mL)	221.9 (196.4, 431.8)	217.0 (208.0, 250.3)	0.008
AUCc (ng*hr /mL)	3923.6 (1846.5, 7293.8)	2445.5 (1327.1, 4140.7)	0.10

AUC_c, cumulative area under the curve; AUC_ss, area under the plasma concentration–time curve at steady state; C_min, minimum concentration; hr, hour; mL, milliliter; ng, nanogram

On multivariable regression, a one standard deviation increase in AUC_c was associated with increased odds of seizure (OR=1.27, 95% CI [1.11, 1.45]). On days with the diagnosis of seizure vs. days without, higher C_max was associated with lower odds of seizure (OR=0.88, 95% CI [0.81, 0.96]), but AUC_ss and C_min were not (OR=1.00122, 95% CI [0.99996, 1.00248] and OR=1.00, 95% CI [0.91, 1.11], respectively) (Table 4).

Table 4.

Adjusted Odds of Seizure Using Metronidazole Exposure Parameters^*

Parameter	OR (95% CI)	OR (95% CI) for 1 SD change in exposure parameter
AUCc (ng*hr /mL)	1.00007 (1.00003, 1.00011)	1.27 (1.11, 1.45)
AUCss (ng*hr /mL)	1.00122 (0.99996, 1.002486)	1.20 (0.99, 1.45)
Cmax (ng/mL)	0.88 (0.81, 0.96)	0.57 (0.38, 0.84)
Cmin (ng/mL)	1.00 (0.91, 1.11)	1.01 (0.75, 1.35)

^*Each estimate is from a separate regression. All regressions were adjusted for gestational age, postmenstrual age, and (0,1) indicators for small for gestational age, presence of inotrope, ventilator dependence, and positive blood culture.

AUC_c, cumulative area under the curve; AUC_ss, area under the plasma concentration–time curve at steady state; C_max,; maximum concentration; C_min, minimum concentration; hr, hour; mL, milliliter; ng, nanogram; OR, odds ratio; SD, standard deviation

DISCUSSION

This study utilized the Pediatrix Clinical Data Warehouse to examine the occurrence of neurotoxicity leading to seizure in infants receiving metronidazole. Dosing simulations were performed by applying baseline characteristics collected from the EHR to a previously created PopPK model. There were higher odds of seizure with increased duration of metronidazole and AUC_c. Higher C_max was associated with lower odds of seizure on the day of seizure vs. the days without seizure. While these associations cannot imply causation, these findings highlight the importance of investigating exposure-safety relationships in infants.

While metronidazole neurotoxicity (including dizziness, peripheral neuropathy, dysarthria, encephalopathy, ataxia, and seizure) has been reported in the literature, the association with metronidazole exposure parameters is unknown.^{6–8,18–34} Neuroimaging studies have identified characteristic lesions, most commonly in the brainstem and cerebellum, consistent with metronidazole neurotoxicity in adults.^{18,22,23,29,32–34} Axonal degeneration has also been noted in histopathologic examination for metronidazole toxicity, and appears to be reversible.³⁵ No similar radiographic or histopathologic studies have been performed in children, but this study demonstrates that neurotoxicity from metronidazole may also occur in the infant population. Unfortunately, pharmacokinetic and pharmacodynamic clinical trials are exceedingly difficult to perform in children because of few eligible participants, phlebotomy challenges, and the need to account for infant growth and maturation in study design.^3,14,36 Most pharmacokinetic or pharmacodynamic studies of metronidazole include fewer than 100 patients, which is not enough for the study of rare adverse events.^3,14,37,38 Our study offers a novel approach that is both an efficient and cost-effective way to predict the exposure-safety relationship of metronidazole and the development of a seizure in infants.

The true incidence and time course of metronidazole neurotoxicity is unknown. Most of the neurotoxic effects have been reported with prolonged courses (longer than 1 month) of metronidazole,^{18,21,26,27,32,34} and have been reversible with cessation of the drug.^22,25,27 In this study, 31 infants out of 1,546 (2%) who received metronidazole experienced a seizure, which occurred at a median (IQV) of 5 days (1, 13) from the start of metronidazole. Our study also found that longer duration of metronidazole was associated with higher odds of experiencing a seizure (OR=1.06, 95% CI [1.02, 1.10]). Increased length of metronidazole course, but not average dose or maximum dose, increased odds of having a seizure, suggesting that course duration is possibly an important factor to consider when prescribing metronidazole to infants. While we cannot determine the true cause of seizure from this simulation, the adjustments performed in the regressions and the time course of metronidazole toxicity appear consistent with the theory that increased duration of metronidazole can lead to seizure in this population.

Exposure parameters, rather than just dosing parameters, should be examined to fully elucidate a drug safety profile.^10–12 This is especially important in pediatric PK studies because interindividual variability of volume of distribution, maturation of enzymes and transporters, and disease effects result in high variability in drug dose-exposure relationship.^3,14 For example, in our study, neither higher cumulative nor maximal doses of metronidazole were associated with increased odds of seizure, but higher AUC_c was (OR=1.00007, 95% CI [1.00003, 1.00011]). Furthermore, if a one standard deviation increase in AUC_c occurred, there was a 27% greater odds of having a seizure. A study that examined only dosing parameters would miss the association of seizure with increased exposure (AUC_c in this case).

Exposure parameters should also be examined in the appropriate compartment of the body that the drug is affecting or causing adverse events. Metronidazole has a wide volume of distribution and is easily able to cross the blood-brain barrier.^20,39 Animal studies have suggested that there is a marked amount of ¹¹C-metronidazole bound to RNA in the nervous system, along with notable degeneration of Purkinje cells, after metronidazole administration.^33,35 In this study, a higher plasma C_max was associated with lower odds of having a seizure (OR=0.88, 95% CI [0.81, 0.96]), which suggests that the blood concentration may not be predictive of neurotoxicity. Instead, lower plasma concentrations may in fact be associated with higher CSF concentrations, which could be associated with seizure risk. One case report in the pediatric population described that metronidazole readily enters the CSF, and at times, metronidazole concentrations in the CSF exceeded those in plasma.⁴⁰ CSF sampling is difficult to conduct in children because it is not ethical to obtain for research purposes alone, and would likely only be permissible when there is concern for a neurologic condition for which CSF sampling would otherwise be performed as per standard of care. Therefore, confirmatory PK study of the CSF in children would be challenging. Additionally, as with all retrospective, non-randomized studies, unidentified covariates may drive the associations observed.

A major limitation of this study is reliance on the EHR to identify dosing information and characteristics that affect PopPK simulation, rather than being able to individually measure the PK data from a prospective clinical trial database. While a prospective trial would be ideal, the sample size required to capture this rare adverse event would be too large to be feasible. Incomplete dosing data sharply reduced the usable sample size. Overall, 4,318 infants received metronidazole, but only 1,546 had complete dosing data usable for simulation. These gaps highlight the importance of complete documentation of dosing data in large datasets so that future studies can be more informative. While the incidence of seizure was small in this cohort, which could create some bias in the simulated data, the incidence of seizure in the overall group was also 2%, suggesting that the cohort used for our simulation is representative of the larger population. This study does not aim to investigate causation, but does report associative relationships discovered. Another limitation of the study is that seizure was a clinical diagnosis rather than one confirmed with electroencephalography, which could have resulted in inconsistencies in diagnosis. Additionally, there are other clinical factors that could lead to seizure that are not be captured from this clinical data warehouse, such as concomitant fever, family history, and possibly other events that were occurring on the day of seizure.

Per design, our study utilizes model predictions of PK parameters to simulate exposures in a large dataset where metronidazole is not quantified in plasma. Despite the strengths of this approach, it is not equivalent to prospective PK analysis using observed drug concentration data when it comes to definitively characterizing exposures in an individual infant. Finally, we used a population PK model developed and internally evaluated per FDA guidance to predict individual estimates of CL and V. We then used them in the exposure calculations in a population that mirrored the population in which the models were developed. While the internal evaluation of two prior iterations of this model done in accordance with FDA guidance were reassuring,^3,14 we acknowledge that formal external validation in an independent dataset would further enhance our confidence in model simulations.

In conclusion, using EHR data to perform dosing simulations with a previously created PopPK model is a potential cost-effective and feasible way to identify rare adverse events in infants. Using this modeling strategy, we identified that metronidazole neurotoxicity could cause seizure in infants, especially with prolonged use. Clinicians should be cautious of placing infants on prolonged courses of metronidazole, but further prospective pharmacovigilance studies are required to validate these findings and discern the safest prescribing practices for metronidazole in infants.