This study was embedded in the Rotterdam Study (Hofman et al., 2015), a population-based prospective cohort study. The study started in 1990 with 7,983 participants that were aged 55 years and older. In 2001, the cohort was extended with 3,011 participants of 55 years and older. A third cohort expansion was performed in 2006 with 3,932 participants aged 45 years and older. Follow-up examinations are performed every 3–4 years in all participants.

From 2005 onward, brain MRI has been incorporated into the core study protocol. Moreover, from 2011 onward, we have been performing hearing assessments in all participants visiting the research center. The current study comprises those participants that underwent both a hearing assessments and brain MRI (n = 3,168) until 2014. Ninety-six percent of the MRIs were conducted before or within less than 3 months after hearing assessment [median: 1 month; interquartile range (IQR) -17 months; 2 months].

The Rotterdam Study was approved by the medical ethics committee according to the Population Study Act Rotterdam Study, executed by the Ministry of Health, Welfare, and Sports of the Netherlands. A written informed consent was obtained from all participants.

All audiometric measurements were performed in a soundproof booth by one trained health care professional. A computer-based clinical audiometry system (Decos Technology Group, version 210.2.6 with AudioNigma interface) and TDH-39 headphones were used.

To determine hearing thresholds in decibel (dB) pure tone audiometry was performed. Thresholds were measured according to the ISO-standard 8253-1 (International Organization for Standardization [ISO], 2010). Air conduction [frequencies 0.25, 0.50, 1, 2, 4, and 8 kilohertz (kHz)] and bone conduction (only two frequencies due to limited time: 0.50 and 4 kHz) were tested for both ears. Masking was done according to the method of Hood (Hood, 1960). Thresholds for the 4 kHz bone conduction were increased with 10 dB afterward based on the discussion on the reference value (Lightfoot and Hughes, 1993; Margolis et al., 2013).

We determined the best hearing ear by taking the average threshold over all frequencies. If both ears were equal, we alternately chose right and left. Furthermore, we determined low (average of 0.25, 0.50, and 1 kHz), speech (average of 0.50, 1, 2, and 4 kHz) and high frequency hearing thresholds (average of 2, 4, and 8 kHz).

We excluded subjects with an air-bone gap of 15 dB or more (n = 240), to eliminate clinically relevant conductive hearing loss.

All participants underwent non-contrast enhanced MRI scanning on a 1.5T-scanner (GE Healthcare, Milwaukee, WI, USA). The full MRI protocol has been described extensively before (Ikram et al., 2015). Briefly, the protocol included a T1-weighted sequence, a proton density-weighted sequence, a FLAIR sequence, and a T2^∗-weighted gradient echo sequence. For all sequences the slice thickness was 1.6 mm (zero-padded to 0.8 mm), except for the FLAIR sequence for which this was 2.5 mm. We used an automated brain tissue classification method, based on a k-nearest-neighbor-classifier algorithm (Vrooman et al., 2007), to quantify the following: total intracranial volume (ICV), total brain volume, GM volume, WM volume, and cerebrospinal fluid volume (in cubic millimeters). To facilitate more regionalized analysis of total brain, GM, and WM volumes, we segmented individual lobes, by non-rigidly registering a template image in which the lobes have been manually outlined (Figure 1) (Bokde et al., 2002, 2005).

Voxel based morphometry (VBM) was performed according to an optimized VBM protocol (Good et al., 2001). FSL software (Smith et al., 2004) was used for VBM data processing, all GM and WM density maps were non-linearly registered to the standard ICBM MNI152 GM and WM template (Montreal Neurological Institute) with a 1 mm × 1 mm × 1 mm voxel resolution. Subsequently, a spatial modulation and smoothing procedure with 3 mm (FWHM 8 mm) isotropic Gaussian kernel were applied to all images.

We collected detailed information on relevant covariates by interview, physical examination, and blood sampling (Hofman et al., 2015). Assessment of systolic and diastolic blood pressure, Body mass index (BMI), cognition (Mini-Mental State Examination (MMSE)), smoking, and alcohol consumption was done at the research center the day the hearing tests were conducted. Smoking was categorized as never, former, or current. Alcohol consumption was categorized as none-drinker, light consumer (1 unit of alcohol per day for women and 1–2 units of alcohol per day for men), or above average (more than 1 unit of alcohol per day for women and more than 2 units of alcohol per day for men) (Dawson and Room, 2000). Level of education was categorized as having completed primary educational level, secondary educational level, or higher education. Fasting blood samples were obtained and serum total cholesterol and high-density lipoprotein (HDL) cholesterol were measured using an automatic enzymatic procedure (Hitachi analyzer, Roche Diagnostics). Glucose was determined enzymatically by the Hexokinase method. We calculated the cholesterol ratio by the quotient of the total and high density lipoprotein cholesterol. Diabetic mellitus was defined as fasting glucose was 7 mmol/L or more, or non-fasting glucose was 11 mmol/L or more, or subjects used antidiabetic medication, diabetes was stated present.

We performed audiometric measurements and brain MRI in 3,168 subjects. From these, 240 subjects with conductive hearing loss and 20 with an incomplete pure-tone audiogram were excluded, leaving 2,908 subjects for analyses.

To allow direct comparison of effect estimates, we calculated Z-scores for the different brain volumes (total brain volume, GM volume, and WM volume). Missing data on covariates in 305 (10.5%, maximum 7.1% per variable) subjects were entered using multiple imputation (iterations = 5). The difference in standardized brain volume and 95% confidence intervals (CI) were calculated per decibel increase in hearing threshold. This was done for the all tone average, low, speech, and high frequency average. Higher thresholds indicate worse hearing.

Model 1 was adjusted for age (linear and quadratic terms), sex, time between the conduction of MRI and hearing assessment, and ICV. Model 2 was additionally adjusted for MMSE, educational level, systolic and diastolic blood pressure, presence of diabetes mellitus (DM), cholesterol ratio, BMI, smoking, and alcohol consumption.

In the VBM analysis the same linear regression models were fitted with voxel values of GM and WM density as the dependent variable. Statistical significant threshold for family wise error correction was calculated by performing 10.000 random permutation tests, resulting to a p-value threshold of 3 × 10^-7.

Data analysis was done using IBM SPSS Statistics version 21 (IBM, Armonk, NY, USA), and R version 3.1.2 (R Foundation for Statistical Computing, Vienna, Austria).

The mean age at the time of the hearing assessment was 64.9 years (SD 7.3) and 56% of the subjects were female. Table 1 displays all relevant characteristics of the study population.

Table 2 shows how hearing impairment is related to brain volume. We found that a higher hearing threshold (i.e., worse hearing) was significantly associated with a smaller brain volume (difference in standardized brain volume per decibel increase in hearing threshold in the age-sex adjusted model: -0.003; CI 95% -0.004; -0.001).

When brain tissues were analyzed separately, the association with the pure-tone audiogram appeared to be driven by WM volume only (difference in standardized brain volume per decibel increase in hearing threshold -0.004; CI 95% -0.006; -0.001). Above associations were present in all frequencies, however, the effect size was strongest in the lower frequencies.

Adjustment for cardiovascular risk factors, alcohol consumption, education and MMSE score did not alter the associations (Table 2, model 2).

Apart from total brain volume, the volumetric parameters of WM in the frontal, temporal, parietal, and occipital lobe were analyzed (Figure 2).

We found similar and significant associations among all four brain lobes for low frequency hearing impairment, independently from the side of the hemisphere. Alike the first analysis, a higher hearing threshold (worse hearing) was significantly associated with smaller WM volume.

For high frequency hearing impairment, we also found associations for the frontal, temporal, parietal, and occipital lobe. Although these were not all significant, they point toward a similar effect: a higher hearing threshold (worse hearing) seems to be associated with smaller white volume.

After exploring the brain lobes, we conducted exploratory voxel-based analysis to identify if age-related hearing impairment was associated with certain GM and WM regions on voxel level.

The analysis showed association for smaller WM in certain areas such as right pre and postcentral gyri, the right insula and the left posterior temporal lobe (Figure 3). Mainly, stronger associations were shown in the right hemisphere, while in the posterior temporal lobe this was in the left hemisphere. Again, associations were more pronounced in the lower than the higher frequencies. For GM, we saw association for both smaller and larger volumes in especially the superior frontal gyrus and medial orbital gyrus right, and the superior parietal gyrus left (Figure 4). However, none of the voxels reached the multiple-testing correction threshold (Supplementary Tables 1 and 2).

Although aging causes cortical GM atrophy (Raz et al., 2005) and is also strongly associated with hearing impairment (Gates and Mills, 2005), we found an association between hearing impairment and a small total brain volume independent of age. Additional adjustments for cognition, alcohol consumption, and cardiovascular risk factors did not influence this relation. When investigating brain volume more specifically, the relation between hearing impairment and smaller total brain volume was solely driven by WM volume. This is in line with various other studies that reported age- and sex adjusted associations between hearing impairment and WM volume (Chang et al., 2004; Lin et al., 2008, 2014; Husain et al., 2011). Interestingly, one other study (Lin et al., 2014) did find an significant association between hearing impairment and GM. They found hearing impairment to be related to a decrease of GM volume in specifically the right temporal lobe. We explain these differences in results through differences in study design. By performing a longitudinal study, Lin et al. (2014) could correct for possible intra-individual brain volume differences, thereby eluding the effect of heterogeneity. Unfortunately, this is not to overcome by a cross-sectional design.

Profant et al. (2014) found results similar to ours, as they showed hearing impairment to be not associated with smaller GM volume. This was only influenced by age. Other cross-sectional studies that did find associations with hearing impairment and brain volumes mainly focused on GM in specific brain regions (Eckert et al., 2012), and even found smaller and larger GM volumes contemporaneously in hearing impaired subjects (Husain et al., 2011; Boyen et al., 2013), something we found in our VBM analysis as well.

Several potential mechanisms may be underlying the relation between hearing impairment and brain volume. First, hearing impairment and smaller brain volumes might be the results of mutual risk factors, which is often referred to in the literature as the “common cause” hypothesis (Dawes et al., 2015). This mechanism fits with our findings that suggest a generalized rather than a singular effect in specific brain areas. Such effect might be of microvascular origin, as this is a known process in both cerebral and peripheral auditory systems (Lin et al., 2011).

Second, brain atrophy might induces hearing impairment. This implicates that hearing impairment – as measured by the pure tone audiogram – also reflects certain central auditory changes. It has been generally adopted that the pure tone audiogram only reflects to peripheral auditory function, as it does not require higher central auditory processing (Pickles, 2008). It is difficult to justify that atrophy of the brain would lead to peripheral effects only.

Third, peripheral hearing impairment may induce brain atrophy through neural deprivation of the auditory system by loss of sensory input (Emmorey et al., 2003). This causative effect has already been described in animal studies, although its role is not clear for the aging human auditory system (Ouda et al., 2015). In this case we might expect to find initial atrophy in specific brain lobes that are involved in auditory processing such as the temporal lobe. Although we initially found comparable associations across all brain lobes in our main analysis, our voxel-based analysis revealed stronger associations in certain brain regions.

Poorer hearing was associated with smaller WM volume in the left temporal lobe, which is the location of the auditory cortex and thus involved in high-level auditory processing. The same goes for the right insula, which participates in key auditory processes (Bamiou et al., 2003) and the right pre and post gyri (speech relevant regions). This suggests that volume differences exist in specific brain regions that are more related to the process of hearing impairment than other regions. The associations in the voxel-based analysis were not significant, but we assume this is a power related problem.

It would be interesting to study the effect of hearing aid use in relation to the hypothesis above, as this might reduce the process of neural deprivation. Unfortunately, only 6% (N = 170) of the participants in our study wore hearing aids and after excluding them, results did not change.

As has been described previously, different types of age-related hearing impairment arise from different origins (Schuknecht and Gacek, 1993). Strial presbycusis involves atrophy of the stria vascularis especially in the apex and thus affects the lower frequencies. Sensory presbycusis involves loss of outer hair cells especially in the base of the cochlea and thus affects the higher frequencies. As such, stronger associations between WM volume and low frequency thresholds suggest at least some influence of a vascular factor. Although we adjusted for known traditional cardiovascular factors, these factors may not reflect well the more subtle differences in (micro)vascularization that might be involved in the case of age-related hearing impairment. Realistically, the above mentioned hypotheses are not mutually exclusive and could all take (a minor) part in the etiology of age-related decline in hearing and brain morphology.