Pharmacokinetic modelling to predict risk of ototoxicity with intravenous tobramycin treatment in cystic fibrosis

Abstract

Introduction

Further optimization of therapeutic drug monitoring (TDM) for aminoglycosides (AGs) is urgently needed, especially in special populations such as those with cystic fibrosis (CF), >50% of whom develop ototoxicity if treated with multiple courses of IV AGs. This study aimed to empirically test a pharmacokinetic (PK) model using Bayesian estimation of drug exposure in the deeper body tissues to determine feasibility for prediction of ototoxicity.

Materials and methods

IV doses (n = 3645) of tobramycin and vancomycin were documented with precise timing from 38 patients with CF (aged 8–21 years), including total doses given and total exposure (cumulative AUC). Concentration results were obtained at 3 and 10 h for the central (C1) compartment. These variables were used in Bayesian estimation to predict trough levels in the secondary tissue compartments (C2 trough) and maximum concentrations (C2_max). The C1 and C2 measures were then correlated with hearing levels in the extended high-frequency range.

Results

Patients with more severe hearing loss were older and had a higher number of tobramycin C2_max concentrations >2 mg/L than patients with normal or lesser degrees of hearing loss. These two factors together significantly predicted average high-frequency hearing level (r = 0.618, P < 0.001). Traditional metrics such as C1 trough concentrations were not predictive. The relative risk for hearing loss was 5.8 times greater with six or more tobramycin courses that exceeded C2_max concentrations of 3 mg/L or higher, with sensitivity of 83% and specificity of 86%.

Conclusions

Advanced PK model-informed analysis predicted ototoxicity risk in patients with CF treated with tobramycin and demonstrated high test prediction.

Introduction

Aminoglycoside (AG)-induced loss of hearing and balance function (ototoxicity) is a common, permanent disability in those affected.¹ Use of AGs is increasing worldwide because they are effective against MDR Gram-negative pathogens and are relatively inexpensive.^2,3 Strategies to prevent ototoxicity without compromising efficacy of treatments is urgently needed, as there are currently no approved therapeutics to prevent ototoxicity.⁴ People with cystic fibrosis (CF) often have long-term and repeated courses of AG treatment for chronic lung infections starting in childhood.^5,6 Hearing loss due to ototoxicity occurs in >50% of patients with CF before they reach adulthood^7,8 and for those with >10 AG courses, 80% are affected by permanent hearing loss.⁹ The primary bacterium requiring AG treatment in CF is Pseudomonas aeruginosa, but increasingly MRSA, non-tuberculous Mycobacterium abscessus (NTM), Burkholderia cepacia complex (BCC) and other Gram-negative bacteria are colonizing the lungs of patients with CF and require AG combination therapy.^6,10–12 Clearly, in such a complex and chronic multifactorial disease, ‘one size fits all’ antibiotic treatment is not adequate.¹³

The first-line treatment recommended for acute pulmonary exacerbations due to P. aeruginosa is once-a-day, high-dose tobramycin, sometimes combined with the glycopeptide drug vancomycin,¹⁴ utilized in >95% of CF centres in the USA to treat P. aeruginosa and MRSA.¹⁵ Once-daily administration minimizes drug accumulation with lower rates of nephrotoxicity in children compared with multiple doses per day.¹⁶ However, tobramycin has a narrow therapeutic window and requires titration of target concentrations due to variable pharmacokinetics (PK) in children with CF.¹⁷ PK monitoring of AG serum concentrations to achieve drug efficacy and detect nephrotoxicity is routine, but the optimal exposure target to maximize treatment success while minimizing toxicity is not well established.¹⁸ Traditional one-compartment analysis with log-linear regression approaches to estimate the AUC_0–24 are markedly affected by blood sampling times.¹⁹

Two-compartment modelling with Bayesian forecasting more accurately reflects tobramycin disposition and provides a more unbiased exposure estimate. This type of PK modelling also accounts for slower diffusion and accumulation in deeper tissue compartments, such as the kidney and cochlea.^11,20 PK modelling uses patient drug dosage information along with parameters such as age, weight and sex. Bayesian statistics are used to fit the PK model to individual data and to generate the central and peripheral compartment concentration–time profiles. Depicted in Figure 1 are simulated tobramycin concentrations for a linear one-compartment PK model (in red) compared with a two-compartment PK model (in blue). Higher C1 trough concentration beyond 24 h after the dose (blue compared with red line) is indicative of drug exchange and accumulation in the tissues and is theoretically associated with higher risk for toxicity.²¹ This concentration accumulation in the peripheral compartment was reported using post-mortem tissue analysis²² and was also described for other drugs such as digoxin.²³ Using such two-compartment PK models, clinicians can implement personalized dose-optimization strategies to both maximize efficacy and minimize toxicity.^24–26 Renal and cochlear toxicity share similar mechanisms, but inner ear tissue clearance is much slower.^27,28

Figure 1.

Illustration of one- and two-compartment model-predicted AG concentration profiles over time after single drug administration. For the one-compartment model, the PK parameters are: CL, 4.9 L/h; V, 17.5 L. For the two-compartment model, the PK parameters are: CL, 4.25 L/h; Vc, 17.5 L; Q, 0.525 L/h; Vp, 65.6 L. This figure appears in colour in the online version of JAC and in black and white in the printed version of JAC.

Hearing monitoring is currently based on the number or duration of doses, which are not highly predictive of ototoxicity.^29,30 Mathematical multicompartment models of nephrotoxicity and ototoxicity have been developed,¹¹ but empirical validation studies of drug exposure using PK models in relation to auditory measures have not been published previously to our knowledge. The objective of this study was to determine the effect of tobramycin, with or without concomitant use of vancomycin, on hearing function in relation to combined retrospective and prospective measures of drug exposure in patients with CF using two-compartment PK modelling. A sensitive measure of hearing, extended high-frequency (EHF) audiometry, was employed as the gold standard measure of ototoxicity.^31–33

Materials and methods

Participants

The patients were recruited from the paediatric CF centre at Cincinnati Children’s Hospital Medical Center (CCHMC) during their hospital stay for IV AG treatment. Participants were excluded if they were not receiving IV-AG treatment, if they were too ill to complete audiological testing, were younger than 6 years old, or had middle ear (conductive) hearing loss. Only four patients received amikacin, so those cases were excluded.

Ethics

The study protocol, materials and testing methods were reviewed and approved by CCHMC Institutional Review Board (IRB, protocol number 2009-0855). Written or verbal informed parental consent was obtained prior to any study procedures, and assent was also obtained from children aged 11 years or older.

Audiometric procedures

All audiometric testing was completed by licensed audiologists in a double-walled soundproof booth (Industrial Acoustics Company, North Aurora, IL, USA) that met standards for ambient noise for audiometric rooms.³⁴ Prior to hearing tests, otoscopy was done during each visit to ensure a clear ear canal and to document the condition of the tympanic membrane. Tympanometry with a 226 Hz probe tone (Titan, Interacoustics Inc., Middlefart, Denmark) was used to assess tympanic membrane and middle-ear function, e.g. normal acoustic admittance, defined as between 0.3 and 1.5 mmho.^35,36 Participants with purely conductive hearing loss (n = 7) were excluded from further data analyses.

An Equinox audiometer (Interacoustics Inc.) with Sennheiser HDA 300 circumaural earphones (Old Lyme, CT, USA) was used to measure standard and EHF, as recommended for ototoxicity monitoring.³⁰ Bone conduction thresholds were tested using a Radioear Inc. B-71 bone vibrator, with narrowband masking in the contralateral ear if air-conducted thresholds were greater than 15 decibel hearing level (dB HL) at frequencies between 0.25 and 4.0 kHz. Pure-tone air conduction audiometry tests were conducted at conventional frequencies (0.25, 0.5, 1, 2, 4, 6 and 8 kHz) and at extended high frequencies (10, 12.5, 14 and 16 kHz). Test-retest reliability was assessed at 1 kHz in each ear. Speech reception thresholds were measured using recorded spondees from the Central Institute for the Deaf W-1 adult or child word list^37,38 to assess inter-test reliability.

Hearing threshold levels were used to classify each ear as either normal or impaired (conductive, sensorineural or mixed). Normal hearing was defined as air- and bone-conduction thresholds in both ears ≤15 dB HL at all test frequencies, based on paediatric criteria. Sensorineural hearing loss (SNHL) was classified if hearing levels exceeded 15 dB HL in either ear and the gaps between bone- and air-conduction thresholds were ≥10 dB at any two frequencies or >20 dB at any one frequency.³⁹ If a ≥10 dB difference between air and bone conduction at two or more frequencies or >20 dB at one frequency were detected in an ear with hearing loss, either a conductive or mixed hearing loss (in cases with bone-conduction thresholds elevated above 15 dB HL) was classified.²⁹

A total of 38 participants were subdivided into the following hearing loss categories: (1) normal hearing (n = 17); (2) hearing loss without American Speech-Language-Hearing Association (ASHA)-defined progression (n = 16); and (3) hearing loss progression (n = 5). Those in the ‘normal hearing’ category had no hearing loss at any frequency range in any of their serial audiograms. The ‘hearing loss without progression’ category consisted of participants with SNHL in one or both ears that either did not change relative to the baseline measure or, if a change occurred, it was not persistent on a subsequent test (ASHA criterion). Participants in the ‘hearing loss progression’ category had an increase in hearing level from the baseline measurement (10 dB or more at two or more consecutive frequencies or 20 dB at one frequency) that persisted on two tests relative to the baseline measurement. In addition to the hearing loss category, the average high-frequency (HF) threshold (8, 10, 12.5, 14 and 16 kHz) in dB HL in the poorer ear was analysed as a continuous dependent variable to provide a sensitive measure of ototoxicity in the frequency range most likely to be affected.

IV antibiotic dosing and PK estimation

Patient demographics including age, gender, weight, IV-AG doses and routine therapeutic drug monitoring (TDM) (3 and 10 h plasma concentrations) were obtained for each IV tobramycin and vancomycin treatment course. The precise times of IV infusion and blood draws were obtained from each participant’s medical record in EPIC (Madison, WI, USA). Bayesian estimation software MwPharm++ (Mediware, Prague, Czech Republic)⁴⁰ was used to estimate PK profiles and total AG exposure for each IV-AG course. This user-friendly software has flexibility to fit individual data and has been validated in clinical studies.^41–43 A tobramycin population PK model in the MwPharm++ model library, as published by Schentag,²² was used as a priori information to estimate individual PK parameters and drug exposures. This model was developed based on serum concentrations up to 20 days after the last dose and measured as low as 0.01 μg/mL to characterize the terminal elimination half-life of tobramycin. This model better characterizes the two-compartment PK by including later sampling times and low concentrations.

For illustration purposes, tobramycin PK profiles were simulated either using a one-compartment or a two-compartment structure model. The two-compartment model is the model used for Bayesian estimation in our analysis, whereas the one-compartment model was developed by fitting the PK profile from simulations of the two-compartment model, and is the model in routine clinical use. As shown in Figure 1, the two-compartment model requires samples after 24 h to capture drug distributions to and from the peripheral compartment and to differentiate from a one-compartment model. For this study, as we needed to estimate the concentrations not only in the plasma but more importantly in the tissues, this two-compartment model helped to better achieve this purpose. Figure S1 (available as Supplementary data at JAC Online) illustrates how Bayesian estimation is applied to inform deeper compartment concentrations. As indicated, the two-compartment model is able to reflect the accumulation of the tobramycin in the deeper compartment. This model was selected over more recent PK models in patients with CF, because the recent models were developed using TDM data collected over shorter post-dose time intervals (8–15 h) and using clinical immunoassays with a higher lower limit of quantification (LLOQ), typically 0.3 mg/L.⁴⁴ In our dataset, approximately 4.1% of samples were lower than 0.3 mg/L and were reported as below quantification limit (BQL). The PK parameters values used for the Bayesian estimation were CL (central clearance; mean ± SD) of 4.25 ± 2.04 L/h, Vc (central compartment volume) of 17.5 ± 8.8 L, Vp (peripheral compartment volume) of 65.6 ± 32.8 L and Q (inter-compartmental clearance) of 0.525 ± 0.263 L/h. All the parameters of the population PK model (including the variability components) used for Bayesian forecasting are provided in Table 1. Given potential physiological and disease status changes between treatment courses,⁴⁵ PK fitting (reconstruction of the PK profile) was completed for each AG course.

Table 1.

Parameters of the population PK model used for Bayesian forecasting, and the posterior Bayesian model

Parameter	Base model parameter estimates	Posterior Bayesian estimates (median)a
Fixed effect parameters
CL (L/h)	4.25	6.58
Vc (L)	17.5	25.0
Q (L/h)	0.525	0.75
Vp (L)	65.6	93.6
Inter-patient variability (CV%)
ωCL (%)	48	77.4
ωVc (%)	50	38.4
ωQ (%)	50	138.4
ωVp (%)	50	89.6
Residual error
εprop (%)	30
εadd (mg/L)	0.05

ω, variance for the inter-patient variability; ε_prop, proportional part of the residual unexplained variability; ε_add, additive part of the residual unexplained variability.

^aThe posterior Bayesian estimates represent the median of the individual Bayesian parameter estimates as generated by the clinical software MwPharm++.

From the estimated PK profile simulations, AG exposure per course (AUC_0–24) and the trough concentrations of the drug in the central compartment (C1) and peripheral compartment (C2, i.e. body tissues) were obtained and analysed, along with maximum C2 concentrations in each course (C2_max). The C2 trough was used as the best available measure proportional to the inner ear, as there is no specific inner ear model available. The total number of lifetime doses for each patient was calculated. The total AG exposure of a participant was calculated by summing the AUCs (mg·h/L) for each of the antibiotic courses. This total cumulative value (AUCC) was divided by the total number of lifetime doses to normalize it for comparison between participants, referred to as the standardized AUCC.

Statistical analysis

Audiometric data were documented using REDCap,^46,47 a secure web-based research database platform, and were exported and combined with the EPIC data for drug doses and the PK data exported from MwPharm++, then formatted for analysis using JASP statistical software version 0.13.1 (https://jasp-stats.org/). Descriptive statistics were used to summarize demographics and outcome measurements to identify any errors and outliers. One patient had a blood level markedly out of range. This was checked in the medical record and appeared to be an entry error, so that single dose was excluded from the data. Pearson correlations were used to explore relationships between HF average and age at test, sex, cumulative drug doses and the PK variables. From the correlation matrix, significant univariate variables were selected to enter multivariate linear regression. Mixed models were conducted to study differences between the hearing loss categories, with age as a covariate. Mauchly’s test was used to assess for assumption of sphericity, and where violated, Greenhouse–Geisser corrections were applied to degrees of freedom. Post hoc tests used Bonferroni corrections, and effect sizes (Cohen’s d) were included. Two-sided significance level was set at P < 0.05 for all analyses. Finally, receiver-operator characteristic (ROC) curves were analysed for prediction of hearing loss category using various C2_max criteria with methods developed in R language.⁴⁸ The C2_max criteria were specified a priori in the ROC analysis, to determine whether there were C2 trough levels at which risk for hearing loss increased.

Results

Fifteen male and 23 female participants were included in the analysis, with an average age of 16 years (SD of 3 years). The total number of doses of tobramycin and vancomycin were highly variable, ranging from 3 to 480 doses for tobramycin and 0 to 64 doses for vancomycin. Bayesian predictions of individual central compartment and tissue compartment tobramycin PK profiles for a typical case are shown in Figure 2a. Overall, the model predictions closely correlated with the observations with a coefficient of determination (R²) of 0.967 (Figure 2b). A summary of the posterior Bayesian estimates, including the median of individual PK parameters with the coefficient of variation (CV%), is presented in Table 1. We also conducted a prediction-corrected visual predictive check (pc-VPC) using the posterior Bayesian model, indicating reasonable predictions of the observed data using the Bayesian approach (Figure S2). A representative selection of most recent individual fitted PK profiles is also presented (Figure S3). Descriptive statistics for analysed variables are given in Table 2. No data were missing for any of the main variables.

Figure 2.

(a) Tobramycin plasma concentration measures (red symbols) and reconstructed PK profiles in the plasma (C1, red line) and in the tissues (C2, blue line) from Bayesian model predictions for a 14 day course. The black dotted line is the clinical cut-off for plasma concentrations (2 mg/L). TDM is performed on Days 1 and 7. (b) Predictive performance of this model, with individual predicted levels compared with observed levels. The model predictions were closely correlated with the observations. The coefficient of determination (R²) of the linear regression is 0.967. This figure appears in colour in the online version of JAC and in black and white in the printed version of JAC.

Table 2.

Descriptive statistics for demographics, hearing levels, doses, main (C1) and deep (C2) compartments, and number of doses exceeding 2 mg/L in C2

Parameter	Mean	SD	Minimum	Maximum
Age at test (years)	16.37	3.37	8	21
HF avg (dB HL)	13.28	16.01	−3.50	66.00
Tobramycin doses (n)	98.66	106.50	3	480
Vancomycin doses (n)	7.32	15.08	0	64
Standardized AUCC (mg·h/L)	108.80	25.29	51.29	151.2
C1avg (mg/L)	0.65	1.39	0.15	8.97
C2avg trough (mg/L)	2.03	0.53	0.64	2.88
C2 > 2 mg/L	55.76	63.09	0	265
C2max > 2 mg/L	7.29	6.69	0	24

HF avg, average EHF (8, 10, 12.5, 14 and 16 kHz) thresholds; AUCC, cumulative area under the concentration curve; C1_avg, central concentrations; C2_avg, deeper tissue concentrations; C2_max, maximum C2 level per course.

PK measures

Univariate correlations are given in Table 3. For demographic factors, age was significantly correlated with hearing level, but sex was not. For drug doses, cumulative doses of tobramycin were significantly correlated with hearing level, but cumulative vancomycin doses were not. In terms of PK variables, the number of C2 trough concentrations >2 mg/L was significantly correlated with HF hearing, as was the number of courses with a maximum C2 level >2 mg/L. However, the cumulative standardized AUCC for tobramycin exposure, average C1 trough and C2 trough concentrations were not correlated with HF hearing.

Table 3.

Univariate Pearson correlations for prediction of hearing loss severity

Variable	Pearson’s r	P value
Tobramycin C2max > 2 mg/L	0.510	<0.001
Age at audiogram	0.496	<0.001
Tobramycin C2 > 2 mg/L	0.368	0.011
Cumulative tobramycin doses	0.345	0.017
Standardized AUCC	0.235	0.078
C2 average trough	0.235	0.078
Sex	0.203	0.110
Cumulative vancomycin doses	0.121	0.235
C1 average trough	−0.119	0.762

Age and maximum deep compartment treatment courses with concentration for tobramycin exceeding 2 mg/L were included in the final predictive model. Significant P values are indicated in bold type.

A multivariate linear regression model was next performed to determine the most important predictive variables for severity of HF hearing loss using the significant univariate factors (age, number of tobramycin doses, C2 > 2 mg/L and C2_max > 2 mg/L). The final model is displayed in Table 4 and includes age and C2_max > 2 mg/L. These two factors together significantly predicted hearing level (r = 0.618, P < 0.001).

Table 4.

Multiple regression model (r = 0.618, P < 0.001)

Factor	Unstandardized	Std error	Standardized	t	P value	95% CI
Intercept	−22.184	10.606		−2.092	0.044	−43.714 to −0.653
Age at test	1.750	0.668	0.368	2.621	0.013	0.394–3.106
C2max > 2 mg/L	0.934	0.336	0.390	2.776	0.009	0.251–1.617

Age and maximum deep compartment treatment courses with concentration for tobramycin exceeding 2 mg/L were included in the final predictive model. Significant P values are indicated in bold type.

Additional PK analysis was performed for tobramycin dosing. The number of C2 trough concentrations and C2_max concentrations exceeding a certain criterion were further analysed according to progressive cutpoints set at 2, 3, 4 and 5 mg/L. Analysis was completed using repeated-measures mixed model ANOVAs with hearing loss category as the between-subject variable, and age as a covariate (Table 5). Each of the C2 cutpoints were significant for hearing loss category, with a significant interaction between C2 cutpoint and hearing loss category. Post hoc comparisons were performed for each of the cutpoints, and each of them were significantly different from each other (Table 6). Effect sizes were large (Cohen’s d > 0.8) for C2 concentrations of 2–5 mg/L and for C2_max concentrations of 2–5 mg/L.

Table 5.

Repeated-measures mixed ANOVA results for deep compartment (C2) concentrations across all tobramycin doses, and for the peak levels during each inpatient treatment

Within-subject effects	Type III sum of squares	df	Mean square	F	P value
C2 concentration	2356.9	1.055a	2234.56	1.396	0.247
C2 concentration × hearing loss category	19637.2	1.055a	18617.94	11.628	0.001
C2 concentration × age at test	101.6	1.055a	96.3	0.060	0.821
Residual	59109.6	36.916a	1601.19
C2max concentration	34.5	1.445a	23.847	1.997	0.157
C2max concentration × hearing loss category	272.9	1.445a	188.835	15.813	<0.001
C2max concentration × age at test	2.3	1.445a	1.605	0.134	0.806
Residual	603.9	50.573	11.942

Significant P values are indicated by bold type.

^aMauchly’s test of sphericity (P < 0.05), Greenhouse–Geisser correction.

Table 6.

Post hoc comparison

Concentration	Mean difference	Standard error	t	Cohen’s d	P valueBonferroni
C2 (mg/L)
2–3	28.842	4.920	5.862	0.951	<0.001
2–4	47.842	8.572	5.581	0.905	<0.001
2–5	53.842	9.923	5.426	0.880	<0.001
3–4	19	3.878	4.900	0.795	<0.001
3–5	25	5.314	4.704	0.763	<0.001
4–5	6	1.558	3.852	0.625	0.003
C2max (mg/L)
2–3	2.289	0.427	5.367	0.871	<0.001
2–4	5.395	0.760	7.094	1.151	<0.001
2–5	6.816	1.015	6.717	1.090	<0.001
3–4	3.105	0.506	6.140	0.996	<0.001
3–5	4.526	0.774	5.851	0.949	<0.001
4–5	1.421	0.351	4.048	0.657	0.002

Significant P values are indicated by bold type.

Results of the ROC analysis for prediction of hearing loss using multiple C2_max tobramycin concentrations are shown in Table 7 and Figure 3. This analysis showed equal test performance for C2_max concentrations of 2 or 3 mg/L, which had higher area under the ROC curve than C2_max concentrations of 4 or 5 mg/L. Sensitivity was 83% and specificity was 86%, while positive predictive value was 83% and negative predictive value was 86%. These are uniformly high values, indicating excellent ability of the C2_max metric to predict hearing loss after treatment with tobramycin. Patients with six or more courses exceeding C2_max of 3 mg/L were 5.8 times more likely to have hearing loss than those with lower concentrations.

Table 7.

ROC curve performance measures for C2_max exceeding 3 mg/L during multiple courses of tobramycin

Measure	Value	Lower limit	Upper limit
Sensitivity	0.824	0.566	0.962
Specificity	0.857	0.637	0.970
Positive predictive value	0.824	0.824	0.824
Negative predictive value	0.857	0.626	0.970
Positive likelihood ratio	5.765	1.976	16.815
Negative likelihood ratio	0.206	0.073	0.073

Optimal cut-off method: Youden; optimal cut-off point: 6; optimal criterion: 0.680.

Figure 3.

ROC curve for the number of C2_max concentrations exceeding 2, 3, 4 or 5 mg/L during multiple courses of tobramycin. The black dotted line is the chance cut-off, while each coloured line represents successive cutpoints tested for sensitivity (y-axis) and false alarm rate on the x-axis (1 − specificity). This figure appears in colour in the online version of JAC and in black and white in the printed version of JAC.

Discussion

Although Bayesian strategies to estimate true drug exposure have been available for over 20 years, they are not standard practice in the clinical setting.⁴⁹ The AUCC has been suggested as an exposure metric suited to nephrotoxicity monitoring in the CF population due to their highly variable PK,⁴⁹ and the need for individualized target concentration intervention.¹⁷ Our data showed that standardized AUCC and C1 trough concentrations did not have a significant correlation with hearing loss. C2 trough and C2_max did show a significant, predictive relationship to severity of hearing loss, presumably because these concentrations are more reflective of prolonged drug exposure in deeper body tissues, such as the inner ear. Excellent clinical test performance was found for C2_max concentrations >2 mg/L with a large effect size and high sensitivity, specificity, positive predictive value and negative predictive value for prediction of ototoxic hearing loss. Bayesian exposure algorithms deployed in software platforms such as we employed in MwPharm++ and other commercial platforms (e.g. InsightRX and DoseMe), provide a means to implement C2 trough monitoring in clinical practice and thus could be deployed to adjust dosing to avoid exposure to toxic levels, particularly in children.⁵⁰ In fact, recent evidence demonstrates that a single course of IV tobramycin causes progression in ototoxic hearing loss in 39% of people with CF,⁵¹ which supports the need for ongoing ototoxicity monitoring and management in this clinical population.

The analysis of cumulative courses of tobramycin supported previous studies, demonstrating a weak but significant relationship between the number of tobramycin doses and HF hearing loss.⁸ Age was also significantly related to HF loss, beyond the variance due to the greater number of doses expected in older patients.

The most important limitation of this study was the inability to obtain in situ measurements due to inaccessibility of the inner ear. Thus, we had to rely on central compartment levels to estimate exposure in the deeper tissues. Such estimates will largely reflect the population parameters, as adjusted by the individual peak and trough levels. We chose a two-compartment model to predict the C2 trough, using the concentration results that were available from our routine TDM practice with sampling at 3 and 10 h after multiple doses. With these concentrations we obtained reasonable estimates of drug distribution into the tissues. However, later sampling times would better reflect drug accumulation into the tissues. In addition, for PK parameter estimation for a two-compartment model, ideally data from four sampling points should be collected to inform all four PK parameters (total body clearance, CL; inter-compartment clearance, Q; V for the central compartment, Vc; and V for the peripheral compartment, Vp). Given the identification of the 24 h timepoint, a more sensitive assay with a lower LLOQ is required. Thus, we recognize that a more sensitive biochemical measurement approach and later sample collections will help to estimate tissue concentrations more accurately. Other limitations of this study include the relatively small sample size, retrospective drug dose collection and sparse audiometric assessment. We have recently obtained funding for a larger, prospective study that will allow us to study AG exposures related to hearing measures and to validate these results across two major CF centres (Cincinnati, OH and Portland, OR). In the new study, we will improve the bioassay approach with a lower LLOQ and take samples at later timepoints (for example 24 h post-dose).

All patients were prescribed outpatient nebulized tobramycin treatment cycles to treat chronic P. aeruginosa, which has not been analysed separately. However, additional tobramycin concentration from this route of administration should have been accounted for in the PK analysis. Further, delivery of the drug via nebulizer treatment is considered to mostly remain in the lung and should not significantly affect systemic concentrations.^52,53 As discussed, the PK model for the C2 trough represents peripheral concentrations once the drug has diffused from the blood. We theorized that trough levels within the inner ear could follow a comparable pattern to that of the peripheral compartment.

In addition, there has been a recent decline in the treatment of CF infections treated with vancomycin, which made the sample size for vancomycin analysis much smaller than that of tobramycin. Further analysis of vancomycin PK would need to have an increased sample size to validate relationships between trough concentrations, total drug exposure and hearing levels. Amikacin is an AG that was rarely used in our sample but is frequently used for NTM patients. Given the promising results for tobramycin secondary compartment models, these other ototoxic drugs could be explored using Bayesian estimation and HF hearing loss as a sensitive outcome measure. Developing models that are specific to drug types and patient populations would enable personalized medicine approaches for dosage. Being able to prescribe the right dose based on patient characteristics could decrease drug toxicity while maintaining treatment efficacy. To confirm the results of this study, there is a need for validation in a larger prospective sample. Prospective studies are needed to develop improved physiological-based PK (PBPK) models to predict C2 trough for the inner ear. With a specific inner ear compartment model, the C2 trough would be more accurately estimated and would provide a more definitive answer as to whether C2 trough concentrations correlate with the amount of hearing loss a patient develops.

These results also emphasize that the effects of AG exposure carry a high risk of permanent hearing loss, stressing the need to develop earlier, more reliable methods of detection in these patients before there is irreversible decline. Hearing testing should be carried out at least annually, or ideally after every 2–3 AG courses to detect changes in hearing, including the EHF ranges (8–16 kHz) that are not often included in standard hearing evaluations. Finally, the variability inherent in hearing loss associated with AG exposure suggests that there are individual and/or genetic predispositions. Such genetic studies are needed to discover factors that may predispose or protect against AG toxicity to better understand ‘tough’ versus ‘sensitive’ ears. Genetic variation in the mitochondrial genome (A1555G and C1494T mutations) is associated with severely increased susceptibility to AG ototoxicity⁵⁴ although A1555G has an allele frequency of only ∼1 in 1111 in the USA. Genetic variants related to three AG-permeant cation channels (TRPA1,⁵⁵ TRPV1⁵⁶ and TRPV4⁵⁷) revealed individual SNPs at these loci associated with either SNHL (or otoprotection) in subjects with CF. We are currently collecting genetic data in a multisite study funded by the Cystic Fibrosis Foundation (PI: A. Garinis) to identify genetic protective or susceptibility factors. Such information would be highly useful in otoprotectant drug discovery for individualized patient care and could help further the development of less ototoxic medications.

Funding

Supported by Place Outcomes Research Award (L.L.H.), Cincinnati Clinical Translational Research Center, National Institutes of Health (NIH) Clinical and Translational Science Award (CTSA) program, grant 2UL1TR001425-05A1, National Institutes of Health (NIH) grant R01 DC017867 (M-PIs L.L.H. and M. P. Feeney).

Transparency declarations

None to declare.

Supplementary data

Figures S1 to S3 are available as Supplementary data at JAC Online.