The relationship and interdependence of auditory thresholds, proposed behavioral measures of hidden hearing loss, and physiological measures of auditory function

Abstract

Objectives:

Standard diagnostic measures focus on threshold elevation but hearing concerns may occur independently of threshold elevation –referred to as ‘hidden hearing loss’ (HHL). A deeper understanding of HHL requires measurements that locate dysfunction along the auditory pathway. This study aimed to describe the relationship and interdependence between certain behavioral and physiological measures of auditory function that are thought to be indicative of HHL.

Design:

Data were collected on a battery of behavioral and physiological measures of hearing. Threshold-dependent variance was removed from each measure prior to generating a multiple regression model of the behavioral measures using the physiological measures.

Study Sample:

224 adults in the United States with audiometric thresholds ≤ 65 dB HL.

Results:

Thresholds accounted for between 21-60% of the variance in our behavioral measures and 5-51% in our physiological measures of hearing. There was no evidence that the behavioral measures of hearing could be predicted by the selected physiological measures.

Conclusions:

Several proposed behavioral measures for HHL: thresholds-in-noise, frequency-modulation detection, and speech recognition in difficult listening conditions, are influenced by hearing sensitivity and are not predicted by outer hair cell or auditory nerve physiology. Therefore, these measures may not be able to assess threshold-independent hearing disorders.

INTRODUCTION

Sensorineural hearing loss is a term that encompasses a plethora of pathologies that can cause hearing loss, effectively from the round window to the auditory cortex. Even within the cochlea, the range of potential pathologies that could affect various aspects of hearing are still unknown. Whereas pathologies that affect outer hair cells (OHCs) can typically be detected by elevated audiometric thresholds, other pathologies, such as inner hair cell (IHC) or auditory nerve (AN) dysfunction or damage may not be detected through the standard diagnostic battery. A phenotype of sensorineural hearing loss that has been a focus of research for the past decade is hidden hearing loss (HHL), which has sustained several definitions, ranging from the phenomenon of anyone with complaints of hearing difficulty despite thresholds within normal limits (e.g., Plack et al., 2014) to the specific pathology known as synaptopathy (Schaette & McAlpine, 2011), introduced by research in animal models. While we are unsure as to 1) what are the perceptual deficits associated with synaptopathy in humans, and 2) all potential causal factors in humans, it cannot be debated that there are people suffering from hearing disorders that cannot be entirely explained by conventional audiometry.

Prevalence of self-reported hearing loss in the United States may be as high as 27% of adults aged 21-74 according to a recent survey (Dillard, 2022). This number includes people with hearing loss per audiometric testing, making it difficult to estimate the number of people who may report hearing loss but have thresholds within normal limits. Studies that have assessed how well self-report aligns with audiometric hearing loss typically agree that most people with moderate or more severe hearing loss do report hearing loss, though a review of ten population studies reported poor agreement between pure-tone audiometry and self-reported hearing loss when thresholds were < 40 dB HL (i.e., mild or no hearing loss; Valete-Rosalino & Rozenfeld, 2005). Oosterloo and colleagues (2020) reported in a survey of 4900 people that ~5% of those with four-frequency pure-tone average (PTA) < 20 dB HL regularly had difficulties understanding what is being said and this prevalence increases to 15% when PTA is 25-35 dB HL. Similarly, a report of the NHANES study found ~2% reported ‘moderate trouble’ or ‘a lot of trouble’ when asked about the status of their hearing but had PTA < 25 dB HL (Choi et al., 2016). If only people with verified normal hearing are surveyed, the incidence of complaints of hearing trouble are 12% according to the Beaver Dam Offspring Study (Tremblay et al., 2015). The prevalence of HHL has also been studied in the audiology clinic. Through a survey of patients at the Audiology Clinic in the Department of Otolaryngology at Washington University School of Medicine in St. Louis, Spehar and Lichtenhan (2018) found that 7% of patients with verified normal audiograms reported they always have difficulty following a conversation in background noise. A 2020 survey by Koerner, Papesh, and Gallun found that over two-thirds of practicing audiologists encounter at least one patient each month with complaints of a hearing disability despite a normal audiogram. More alarmingly, a quarter of audiologists encounter at least four of these patients per month.

The myriad of posited causes of the discrepancy between audiometric thresholds and self-report include, but are not limited to, central presbycusis, obscure auditory dysfunction, King-Kopetzky Syndrome, auditory dysacusis, auditory processing disorders, idiopathic discriminatory dysfunction, cochlear synaptopathy, auditory neuropathy, inner hair cell damage, strial dysfunction, metabolic disorder, tinnitus, neurocognitive disorders, dyslexia, attention deficit disorder, traumatic brain injury, spatial hearing disorders, aging, dementia, cognitive decline or impairment, receptive aphasia, and Alzheimer’s Disease (Beck & Danhauer, 2019). These potential pathologies are not mutually exclusive, and any of these pathologies could be comorbid with a loss of hearing sensitivity. In fact, some of the more recently studied pathologies, such as noise-induced synaptopathy (NIS), would likely exist in parallel with audiometric hearing loss in many patients, as repeated noise exposure can elevate thresholds over time. Yet despite this, most studies in the recent expedition to find a peripheral site of lesion that explains this discrepancy have exclusively recruited normal-hearing participants. Furthermore, most definitions of HHL ignore the possibility of a comorbid audiometric hearing loss. Using NIS as an example, of the 23 studies reviewed by Le Prell (2019) determining effects of noise exposure on the auditory brainstem response (ABR) and speech in noise, only three (Ridley et al., 2018; Valderrama et al., 2018; Yeend et al., 2017) did not have inclusion criteria of audiometric thresholds within normal limits (typically an upper limit of 20 or 25 dB HL between 0.25-8 kHz). There are good reasons to exclude participants with hearing loss in these studies, primarily because loss of hearing sensitivity has confounding effects on the auditory measures tested. For example, both a loss of sensitivity and (theoretically) NIS would decrease ABR wave amplitudes (Verhulst et al., 2016) and reduce speech recognition-in-noise (Hoben et al., 2017; Léger et al., 2012; Summers et al., 2013), making causal inferences difficult. Exclusively recruiting from a normal-hearing population, however, introduces sampling bias and reduces the probability of finding NIS within the sample, which could explain why the body of literature on NIS in humans has been inconclusive to date.

It is important to accept the possibility of comorbidity of HHL, regardless of cause, with traditional hearing loss. The combination of these diseases could conceivably manifest as poor speech understanding in noise despite amplification treatment, a common complaint in the audiology clinic. Because this does not follow typical definitions of HHL, we henceforth use the term threshold-independent hearing disorder (TIHD), which describes a deficit in any auditory measure that cannot be explained by loss of hearing sensitivity. In the present study, we acknowledge the probability of comorbidity of TIHD with threshold-related disorders. We aim to explain the threshold-independent variance of several behavioral auditory measures using physiological tests. If threshold-independent variance in behavioral measures can be explained by physiological measures of the auditory pathway, then we would argue them to be a clinically useful tool for the assessment of TIHD. Identifying TIHD relies on finding measures of hearing that are not affected by hearing sensitivity, a difficult task, or by accounting for the threshold-dependence within an auditory measure. Ridley et al. (2018) removed the threshold-dependent variance from several auditory measures and used the threshold-independent variance to look for evidence of NIS; however, the method of calculating threshold-independent variance was limited by two assumptions. First, the metric of hearing sensitivity used to remove threshold-dependent variance was pure-tone threshold at a single frequency obtained through a 3-alternative forced choice adaptive task. For example, the auditory brainstem response (ABR) wave I amplitude to a 4 kHz tone was regressed with the pure-tone threshold at 4 kHz. This assumes that hearing sensitivity at other frequencies does not affect wave amplitude at 4 kHz, when it is known that AN fibers outside the stimulated frequency region may contribute to the response especially at high stimulus levels (Henry et al., 2014; Lewis et al., 2015; Strelcyk et al., 2009). The second problem with this method was that it assumed a linear relationship between the threshold metric and the auditory measure. In the present study, we optimize the removal of the threshold variance by comparing four quantifications of hearing threshold and allow for a polynomial relationship.

In addition to improving methods of finding threshold-independent variance, we expand on prior research by recruiting a larger sample size and including more behavioral and physiological measures of hearing. Several behavioral measures have been explored as assessments of TIHD, mostly targeting the theoretical symptoms of IHC or AN disorder, that is, deficits in understanding speech in difficult listening situations and temporal processing. Functioning IHCs are vital to hearing, yet there are no clinical tests of IHC used commonly in the clinic. Pure-tone thresholds in noise (TIN) have been proposed as a candidate for assessing the health of IHCs, as selective IHC loss has been shown to elevate TIN while thresholds in quiet remain unaffected (Lobarinas et al., 2016). For example, the threshold equalizing noise test was developed to detect the presence of “dead regions” in the cochlea (Moore et al., 2000) and has been used to find cochlear damage in tinnitus patients with normal hearing sensitivity (Thabet, 2009). A similar pattern is seen in the ABR wave 1, in that wave amplitude is reduced while threshold is maintained in animal models of IHC loss, NIS, and age-related synaptopathy (Furman et al., 2013; Kujawa & Liberman, 2009; Lin et al., 2011; Valero et al., 2017), though a reduced wave I amplitude has only been found in humans with extreme occupational noise exposure from military service (Bramhall et al., 2017, 2021). In the present study, we hypothesize that the threshold-independent variance of TIN can be explained by ABR wave I amplitude.

Temporal processing has also been shown to be impaired in cases of TIHD (Bharadwaj et al., 2014, 2015; Mehraei et al., 2016) due to decreased phase-locking ability in IHC or AN pathology. Behavioral measures of temporal processing ability may include gap detection as well as detection or discrimination of modulated tones. Frequency-modulation detection threshold (FMDT) is a candidate for assessment of temporal fine structure processing, as it depends on the phase-locking ability of the AN, important for speech recognition but not so important for pure-tone thresholds (Moore & Sek, 1996). The frequency following response (FFR) to a periodic stimulus is a measure of temporal processing involving cortical and subcortical structures of the auditory pathway (Anderson et al., 2013; Marmel et al., 2018; Ruggles et al., 2011, 2012; Skoe & Kraus, 2010). We hypothesize that the threshold-independent variance of FMDT can be explained by ABR wave amplitudes and the FFR of the speech-evoked ABR.

Deficits in suprathreshold speech recognition are a hallmark symptom of TIHD. Subjective complaints of speech-in-noise in some cases of TIHD may be attributed to subclinical OHC loss or dysfunction (Kamerer et al., 2022; Parker, 2020; Zhao & Stephens, 2006). While audiometric hearing loss is attributed mostly to OHC loss, thresholds are only moderately correlated with otoacoustic emissions (OAEs), a physiological measure of OHC integrity (Engdahl et al., 2013; Parker, 2020). A study by Pang and colleagues (2019) found that the primary complaint of self-reported hearing loss was not difficulty understanding speech in quiet situations, but rather speech in background noise. Speech-in-noise deficits have been posited as a result of NIS (e.g., Bharadwaj et al., 2014; Kobel et al., 2017; Le Prell & Brungart, 2016; Liberman et al., 2016; Plack et al., 2014) based on animal models that found selective loss of low- and medium-spontaneous-rate fibers in response to noise exposure (Furman et al., 2013; Song et al., 2016). Similar to studies of ABR and NIS, the majority of human studies of speech-in-noise and NIS recruited young, normal-hearing participants and have found no evidence of a relationship between noise exposure and speech-in-noise (e.g., Fulbright et al., 2017; Grinn et al., 2017; Grose et al., 2017; Guest et al., 2018; Shehabi et al., 2022). In the present study, we included measures of speech recognition in quiet, speech-in-noise, and two temporally-modulated conditions: time-compressed speech and reverberant speech. Besides added noise, reverberation and temporal-distortion (such as time compression) have also been shown to be more detrimental to speech recognition in people at a higher risk for NIS (Liberman et al., 2016). We hypothesized that the threshold-independent variance of speech recognition will be predicted by OAEs, ABR wave amplitudes, and the FFR.

There are many pathologies of the cochlea and auditory nerve that could significantly impact hearing but remain “hidden” to the standard diagnostic battery that focuses on measures of hearing sensitivity over suprathreshold hearing. Several clinically feasible, behavioral measures of hearing have been proposed to be sensitive to hidden cochlear and neural pathologies. Most studies intending to find these hidden pathologies have recruited young, normal hearing participants; however, hidden pathologies are likely to be comorbid with loss of hearing sensitivity in most clinical patients. In the present study, we examine whether several proposed behavioral measures of auditory health (TIN, FMDT, and speech recognition in difficult listening situations) can be predicted by physiological measures of cochlear and auditory nerve health (OAE and ABR), that are specific and objective but time-consuming and require specialized equipment and training. We were specifically interested in whether these measures could be useful to find TIHD in people with both normal hearing and those with hearing loss, so we optimize a method of removing threshold-dependent variance of the sample.

METHODS

Procedures

Participants completed all measures within two months over two visits. Average data collection time for each participant was approximately four hours. All procedures were approved by the Boys Town National Research Hospital Institutional Review Board (IRB# 16–01-XP), and informed consent was obtained from all participants. Participants were paid for their participation.

Participants

A sample of 224 adults (119 female) between the ages of 19 and 86 (mean age = 45) years participated in this study. Mean years of education of the cohort was 15, while the minimum was 12 and the highest was 21. Racial identifications were 209 as white, 2 as Asian or Pacific Islander, 6 as Black, 1 as American Indian, and 4 chose not to disclose. Hispanic identifications were 201 as not Hispanic or Latino, 13 as Hispanic or Latino, and 10 did not disclose. Recruitment of participants specifically targeted those with normal hearing, or standard audiometric thresholds ≤ 15 dB HL (n = 101) and those with slight to moderately severe hearing loss (15–65 dB HL; n = 123). Furthermore, an effort to recruit participants with and without a history of noise exposure was made to potentially increase variability in performance on the auditory measures, as many of the auditory measures have been proposed to theoretically identify NIS. Participants were asked whether they had been exposed to loud impulse noise (e.g., explosion or gunfire) without hearing protection ever in their lifetime. This single yes/no question has been shown to predict whether a person will have a high or low score on the Lifetime Exposure to Noise and Solvents Questionnaire because of how impulse noise exposure is weighted in the quantification process (Bramhall et al., 2017; Kamerer, Kopun, Fultz, Allen, et al., 2019). Ninety-five participants answered that they had never been exposed to a high-impact impulse noise while 129 participants answered that they had at least once been exposed to impulse noise in their lifetime. As this question is not a validated measure of noise exposure history, it was not included in the analysis presented in this study. Participants were also asked about tinnitus, though presence of tinnitus was not included in the analysis for this study. Ninety-nine participants reported that they experience buzzing, roaring, static, or ringing in at least one ear that lasts for more than 2 minutes at a time.

Case history, tympanometry, and standard audiometry determined inclusion in the study. Exclusion criteria included a history of genetic or congenital hearing loss, family history of hearing loss prior to presbycusis, history of traumatic brain injury, or history of exposure to ototoxic medication or solvents. For inclusion in the study, participants had to be native speakers of English and at least one ear was required to pass the criteria set for tympanometry and audiometry. Middle ear pressure was required to be between −100 and 50 daPa and static compliance between 0.2 and 2.5 cm³ as measured by 226 Hz tympanometry (Otoflex 100, Madsen). Bone conduction thresholds were measured at octave frequencies between 0.5 and 4 kHz and air-bone gaps were required to be ≤ 10 dB as a cross-check of middle ear function. Participants were also required to have air-conduction thresholds ≤ 65 dB HL at all frequencies between 0.25 and 8 kHz. The upper limit was chosen because other measures, such as TIN and FMDT were presented at an equal sensation level relative to pure-tone thresholds at test frequencies, and we were hesitant to present these stimuli above a certain sound pressure level. All behavioral and physiological measures included in the study were made monaurally. If both ears met the inclusion criteria, the “better” ear (or ear with lower thresholds across all frequencies) was selected for testing. If both ears had similar audiometric thresholds, the test ear was selected randomly, though there was an attempt to include equal numbers of right and left ears.

Audiometry

Pure-tone air conduction thresholds at six octave frequencies (0.25–8 kHz), three inter-octave frequencies (0.75, 3, and 6 kHz), and two extended high-frequencies (11.2 and 16 kHz) were measured via over-the-ear headphones (HD 200 Sennheiser, Wedemark, Germany) and an audiometer (GSI AudioStar Pro, Grason-Stadler) in 5-dB steps following the Hughson-Westlake procedure (ASHA, 1978). Several behavioral and physiological measures in the study were tested at 1.5 and 4 kHz; therefore, audiometric thresholds at 1.5 & 4 kHz were measured in 2-dB steps to increase accuracy of threshold at those frequencies. Equipment limitations only allowed for testing of levels up to 90 dB HL for 11.2 kHz and 60 dB HL for 16 kHz, therefore participants with thresholds greater than these levels were assigned a threshold of 5 dB above the level limit (e.g., 95 dB HL for 11.2 kHz). In total there was one participant with no response to 90 dB HL at 11.2 kHz and 73 participants with no response to 60 dB HL at 16 kHz.

Threshold-in-Quiet

In addition to standard audiometry, a 3-alternative forced choice (3AFC) adaptive procedure was used to measure threshold-in-quiet (TIQ) at 1.5 and 4 kHz. This method mitigates some biases known to occur in standard audiometry such as entrainment to the stimulus interval (i.e., interval bias), and effects of age (Gelfand, 1982; Yost, 1978). For a 3AFC, three intervals were presented with only one interval containing the target stimulus; in this case two intervals were silent and one interval contained a pure-tone (AudioLab MATLAB; developed by Lopez-Poveda). Participants were required to indicate which interval contained the pure-tone and feedback was provided for each response. A 2-down, 1-up adaptive procedure was used. It converges upon the 71% point on the psychometric function that was used as a threshold estimate. The initial stimulus level was 20 dB above the participant’s audiometric threshold (in dB SPL) at the stimulus frequency (1.5 and 4 kHz), rounded up to the nearest 10 dB. The procedure had an initial step size of 5 dB which was reduced to 2 dB after 3 reversals. The final 6 reversals were used to determine threshold (dB SPL). Participants completed one training run to familiarize them with the procedure and then two test runs which were averaged to determine TIQ at a given frequency. Test runs were included when the within-run standard deviation was ≤ 5 dB; and two test runs were only averaged if the thresholds were ≤ 6 dB apart. If the two runs did not meet both criteria, additional runs were completed until two runs met qualifications.

Thresholds-in-Noise

A 3AFC procedure identical to the procedure described for TIQ was used to measure thresholds-in-noise (TIN) at 1.5 and 4 kHz. In this procedure, all three intervals contained a broadband noise from 0.2–8 kHz set at a constant 70 dB SPL, while one interval also contained a pure tone. While the noise remained constant, the level of the tone was varied to determine TIN. A TIN run had a total of 12 reversals, the final eight of which were used to determine threshold. Exclusion of a TIN run followed the same rules as TIQ.

Frequency-Modulation Detection

One of several behavioral measures of temporal processing ability may be assessed by a modulation-detection task (Moore & Glasberg, 1989). We implemented the 3AFC adaptive procedure for frequency-modulation detection (FMDT) as described by Strelcyk & Dau (2009) and Johannesen et al. (2016). Stimuli for the FM detection threshold tasks were tones with a duration of 750 ms, 50 ms raised cosine ramps, and a fixed level of 70 dB SPL. In one interval, the tone was frequency-modulated at a rate of 2 Hz and variable excursion frequency. The minimum detectable excursion in Hertz was log-transformed. All three intervals were also sinusoidally-amplitude-modulated with a fixed carrier frequency of 1.5 kHz; instantaneous modulation rate between 1–3 Hz; and modulation depth of 6 dB. The phase of the non-FM tones was 0 rad, while the FM tone phase was set to 1.5π rad. The choice of sinusoidal-AM, carrier frequencies, and modulation rates are intended to capture temporal fine structure processing ability as opposed to detection of change in excitation patterns or place cues (Moore & Sek, 1996). Participants were instructed to listen to all three intervals and then click on the box representing the interval that was frequency modulated. Participants were given feedback in the form of a flashing green box for correct responses and red for incorrect responses. Frequency excursion was adjusted according to a 1-up 2-down rule until 71% threshold was reached. Initial step size was log₁₀(1.5) Hz, which decreased to log₁₀(1.26) Hz after four reversals. The task concluded after 12 reversals in frequency excursion or 100 trials. Each participant completed an un-scored practice run and two scored runs. The FM detection threshold was defined as the mean frequency excursion for the last eight reversals of each run (log₁₀Hz). A run was excluded and repeated if the SD was > 0.2 and/or if the difference in threshold for the two trials was > 0.3. Two sequential runs were completed before moving to the next task.

Speech Recognition

Word recognition was assessed in four listening conditions: (1) speech-in-quiet (Quiet), (2) speech-in-noise (Noise), (3) speech that had been time-compressed by 45% (TC), and (4) speech that had been time-compressed by 45% and a reverberation time of 0.3 sec (TC+Rev) (Noffsinger et al., 1994). The stimuli were four 50-word lists spoken by a male talker per condition for a total of 200 words (NU-6; Auditec, Inc., St. Louis, MO) presented monaurally through ER-3A insert earphones (Etymōtic Research, Elk Grove, IL). The words were presented at 65 dB SPL for participants with a pure-tone average (PTA) at 1, 2, and 4 kHz of ≤ 35 dB HL. For eight participants with a PTA > 35 dB SPL, words were presented at 30 dB Sensation Level (SL) of PTA rounded up to the nearest 5 dB for all conditions to ensure that stimuli were audible. The Noise condition consisted of steady noise, spectrally-weighted using the international long-term average speech spectrum for combined male and female talkers and presented at 0 dB SNR (Byrne et al., 1994). Performance in each condition was measured as the percent words correct of the final 45 words in each list. Given the difficulty of some of the listening situations, the first five words were considered as training in order to familiarize participants with the condition.

Otoacoustic Emissions

DPOAEs were measured using custom-designed software (EMAV, version 3.3; Neely & Liu, 1994). The measurement hardware included a DPOAE probe-microphone system (ER-10X, Etymōtic Research, Elk Grove, IL) and a 24- bit soundcard (BabyFace Pro, RME, Germany). Two primary tones (f₁ and f₂) were generated by two separate channels of the soundcard and sent to separate sound sources housed in the probe-microphone system. DPOAE measurements were made only at the two experimental f₂ frequencies, 1.5 and 4 kHz, at stimulus level L₂ = 55 dB SPL. The level of L₁ was set at 61 dB SPL, in accordance with (Kummer et al., 1998, 2000). This stimulus level results in the most accurate identification of auditory status from DPOAEs (Johnson et al., 2010; Stover & Norton, 1993). The f₂/f₁ ratio was 1.22. Prior to making DPOAE measurements, stimulus levels were calibrated in the ear canal. Although it is known that standing waves may influence estimates of SPL under these conditions, especially at 4 kHz, (e.g., Reuven et al., 2013; Richmond et al., 2011; Scheperle et al., 2008; Siegel & Hirohata, 1994) the decision was made to use SPL calibrations because of their relative ease. DPOAE data were collected into two separate buffers and the level of the DPOAE was obtained by summing the contents of the two buffers in the 2f₁–f₂ frequency bin. The level of the noise was estimated by subtracting the contents of the two buffers and then averaging the level in the 2f₁–f₂ frequency bin along with the level in the five bins on either side of the 2f₁–f₂ frequency bin (resolution of 3.9 Hz). Data collection ended when either the noise floor was < −20 dB SPL or artifact-free averaging time reached 65.6 s. Data collection stopped most frequently on the noise-floor criterion and rarely on the test-time criterion.

Auditory Brainstem Response

Tone-Burst Evoked Response •

Tone-burst-elicited ABR waveforms were recorded (OptiAmp USB, Intelligent Hearing Systems, Miami, FL, USA) at 1.5 and 4 kHz using custom software on a computer equipped with a 24-bit soundcard (i.e., Babyface Pro; RME, Germany). Electroencephalographic (EEG) responses were acquired using surface electrodes placed at the forehead (Fpz, ground), vertex (Cz, noninverting active), and an inverting reference electrode placed in the ear canal (ER3–26A gold foil tiptrodes). Pure-tones at 1.5 and 4 kHz were gated via Blackman window with duration of 1 ms. Stimuli were presented in alternating polarity monaurally at a rate of 11/sec to an ER-3A insert earphone (Etymotic Research, Elk Grove, IL) connected to the soundcard. The stimulus sound-pressure level (SPL) was 110 dB peak-equivalent (pe)SPL. Calibration of the stimulus levels was done using a sound level meter (System 824 and SoundTrack LxT1; Larson Davis, Provo, UT) with the ER-3A connected via a 2cc coupler (G.R.A.S. 60126, Denmark). High levels were chosen to maximize the number of ABR waves observed in participants, especially those with hearing loss (Ridley et al., 2018) and because larger effects of noise exposure history have been seen at such levels (Bramhall et al., 2017, 2021). Electrode impedances were ≤ 5 kΩ in all cases. The EEG signal was amplified (gain = 100,000), filtered (0.01 to 1.5 kHz; Opti-Amp 8001; Intelligent Hearing Systems, Miami, FL), filtered for line interference using a 60 Hz notch filter and directed to the computer via the soundcard for averaging. Responses were separated by even and odd recordings and stored in two buffers which were averaged for the final waveform (total averages = 1500 artifact-free responses). Artifact rejection was based on the peak absolute differences between the buffers and was set at a maximum of ± 20 μV. Two examiners independently identified peaks and troughs of ABR waves I and V. The software allowed for a resolution of 0.02 μV for amplitude and 0.02 ms for latency. Wave amplitudes were calculated as the difference between the positive peak and the following trough. Latencies were used to clarify disagreements between examiners but were not used for any other analyses. Differences > 0.02 ms, which occurred in 72 of the 440 total waveforms (16%), were resolved by a third examiner. ABRs could not be recorded in four participants.

Speech-Evoked Response •

The ABR to a speech stimulus was recorded immediately following the tone-burst-elicited ABR. A 170-ms synthetic /da/ was chosen as the stimulus because it has been used extensively in complex ABR research (see Skoe & Kraus, 2010). The stimulus used in the present study was developed by the Auditory Neuroscience Laboratory at Northwestern University as part of their Brainstem toolbox. The /da/ is a six-formant syllable synthesized at a rate of 20 kHz. The duration is 170 ms, with a voicing onset at 10 ms (100 Hz fundamental frequency). Additional details of the formant frequencies and transitions can be found in Song et al. (2011). The stimulus was played at a level of 90 dB SPL at a rate of 4/s. The EEG was band-pass filtered at cutoff frequencies of 0.1–3 kHz and the processing delay of the soundcard was taken into account when analyzing the data. The analyses performed on the speech ABR were directed at the response of the periodic (vowel) portion of the stimulus, therefore the response to the initial transient portion of the stimulus and formant transition portion were removed and analyses were performed over the portion of the response that was delayed 60–170 ms relative to onset of the stimulus (Song et al., 2011). From the steady-state portion of the response, the frequency-following response (FFR) was calculated as the strength of the spectral components of the response relative to the noise floor.

Analyses

The goal of the analysis was to assess whether the physiological measures could predict the behavioral outcome measures after removing threshold-dependent variance. Physiological predictor variables included DPOAE level (dB SPL) at 1.5 and 4 kHz, ABR waves I and V amplitude (μV) to 1.5 and 4 kHz tones, and the speech-evoked FFR. Behavioral outcome variables included TIN (dB SPL) at 1.5 and 4 kHz, FMDT at 1.5 kHz, and speech recognition (%) in four listening conditions. Statistical analyses began with assessment of distributions and identification and removal of outliers greater than 3 standard deviations from the mean for physiological variables, as a deviation that size typically indicates equipment or testing error (the total n for each measure after removal of outliers is reported in Table 1). The next step was removal of threshold-dependent variance from each outcome and predictor variable. Then, regression models for each threshold-independent behavioral outcome variable were built using the threshold-independent physiological predictor variables. If any predictive models were significant, the final step was to explore possible mediating factors in our heterogenous sample, such as age, sex, impulse noise exposure, tinnitus, and extended high frequency thresholds, on the relationship between the physiological and behavioral variables.

Table 1.

Summary statistics

			n	mean	SD	min	max
Threshold Quantification	PTA	dB HL	224	12.79	10.24	−6.5	52.25
	SII	%	224	89	15	22	99.88
	C1	a.u.	224	45.17	34.98	−15.86	161.35
	TIQ1.5	dB SPL	223	10.2	11.44	−7.5	64
	TIQ4	dB SPL	223	21.51	18.05	−8.5	69.33
Behavioral Outcome Variables	TIN1.5	dB SPL	222	54.35	2.65	50.25	72.12
	TIN4	dB SPL	222	57.19	3.7	52	76.25
	FMDT	log10Hz	224	1.01	0.22	0.49	1.94
	Quiet	% correct	224	98.2	2.84	82.22	100
	Noise	% correct	224	79.26	10.09	37.78	100
	TC	% correct	224	89.07	9.19	46.67	100
	TC+Rev	% correct	224	61.74	12.51	15.56	84.44
Physiological Predictor Variables	DPOAE1.5	dB SPL	224	4.26	8.97	−29.75	23.05
	DPOAE4	dB SPL	224	−3.4	11.23	−31.33	19.62
	Wave I1.5	μV	220	0.20	0.18	0.003	0.76
	Wave I4	μV	220	0.24	0.17	0.008	0.87
	WaveV1.5	μV	220	0.37	0.17	0.04	0.83
	Wave V4	μV	220	0.30	0.16	0.005	0.66
	FFR	a.u.	220	9.79	9.79	−8.6	30.62

PTA, pure-tone average; SII, speech intelligibility index; CI, principal component 1; a.u., arbitrary unit; subscript _1.5 and ₄, stimulus frequency in kHz; TIQ, threshold-in-quiet; TIN, threshold-in-noise; FMDT, frequency-modulation detection threshold; TC, time-compressed words; Rev, reverberation; DPOAE, distortion-product otoacoustic emissions; FFR, frequency following response; SD, standard deviation

Removal of threshold-dependent variance involved regressing threshold-related metrics with each variable and calculating the residual error (Ridley et al., 2018). The residual error of each variable was used as the threshold-independent representation of that variable in further analysis. Quantification of hearing sensitivity into a single, yet meaningful, metric is not ideal. Therefore, to maximize removal of threshold-dependent variance, four quantifications of threshold and three orders of polynomial regression in the explained variance (R²) of each behavioral outcome and physiological predictor variable were compared. The regression model with the highest R² for each variable was used to calculate residual error for that variable. The first quantification of threshold was four-frequency pure-tone average (PTA) using audiometric thresholds in dB HL at 0.5, 1, 2, and 4 kHz. The second quantification was speech-intelligibility index (SII) calculated for “normal” speech levels of 62.35 dB SPL according to ANSI S3.5–1997. Input thresholds for SII included all octave and inter-octave frequencies from 0.25–8 kHz, with interpolation between frequencies, and critical band SII procedure constants (‘sii’ package R; Warnes, 2013). The third quantification was a principal component analysis (PCA; ‘prcomp’ R) on all available audiometric thresholds between 0.25 and 8 kHz. Principal components were rotated orthogonally to minimize collinearity. The first component (C1) of the PCA was used in further analysis. The fourth quantification was TIQ at 1.5 and 4 kHz.

Individual third-order polynomial regression models were created for each behavioral and physiological variables using each of the threshold quantifications as predictors. The models with TIQ used the TIQ at the same stimulus frequency of the variable (e.g., TIQ at 1.5 kHz was regressed with DPOAE at 1.5 kHz). When no specific stimulus frequency was used, one model for each TIQ frequency was compared. For example, speech recognition in quiet compared models of PTA, SII, C1, TIQ_1.5, and TIQ₄. The cubic, quadratic, and linear versions of each model were compared for best fit, defined as the highest variance explained (R²), Akaike Information Criteria (Akaike, 1973), and Sawa’s Bayesian Information Criteria (Sawa, 1978). The best fit polynomial of each threshold quantification was compared to the other threshold quantifications for that behavioral or physiological variable. The threshold quantification that accounted for the most variance was used to calculate residual error. As the goal was best fit to the data more so than finding significant relationships, we were not concerned with correcting for multiple comparisons for the removal of threshold-dependent variance.

Regression models for the threshold-independent behavioral outcome measures (TIN, FMDT, and speech recognition) were built using threshold-independent physiological measures (DPOAE, ABR, and FFR). An unintended benefit of the threshold removal procedure is that it centers the data. A PCA on the physiological predictor variables was implemented to reduce collinearity and the number of model variables. We predicted the physiological variables would load into two primary components, one related to OAEs and one related to ABRs. We were unsure if the FFR would load together with the ABR wave amplitudes or separately. Linear regression models were built for each threshold-independent outcome measure using n principal components for physiological measures. Statistical analyses were performed in R.

RESULTS

Distributions

To provide a sense of variability in audiometric thresholds, the mean (solid line) and range (shaded region) of thresholds are shown in Figure 1. Due to the inclusion criteria, all thresholds below 8 kHz are ≤ 65 dB HL, except at 8 kHz, where two participants had thresholds at 70 and 75 dB HL. Because they met all other inclusion criteria, these participants were included in the study. As expected from recruitment efforts, there were approximately equal numbers of participants with normal hearing as those with hearing loss, resulting in a distribution skewed toward the lower thresholds. Distributions of each threshold quantification, behavioral outcome measure, and physiological predictor measure are presented in Table 1. There were ceiling effects in SII and speech recognition in three of the four listening conditions: quiet, noise, and time-compressed speech. A rationalized arcsine unit (Studebaker, 1985) conversion was attempted but did not reduce the skew and was therefore not used. In one participant, a Wave I at 1.5 kHz could not be determined and another participant was excluded because their WaveI_1.5 amplitude was 1.86 mV, greater than six standard deviations above the mean which was determined to be due to human error. In one participant, a Wave I at 4 kHz could not be determined. Three participants had ratios above 60, more than three standard deviations above the mean and were excluded from further analysis that involved FFR.

Figure 1.

Means (solid black line) and range (shaded grey) of audiometric thresholds of the 224 participants. Participants with no response at 60 or 90 dB HL for 11.2 and 16 kHz, respectively, were assigned a threshold value 5 dB above (i.e., 65 or 95 dB HL).

Removal of Threshold-Dependent Variance

The first step in the analysis was to quantify and remove the variance in each outcome and predictor that could be accounted for by behavioral pure-tone thresholds. Four potential metrics of threshold were in competition to account for the most variance: PTA, SII, TIQ_SF, and C1. Linear and quadratic polynomial models were built for each outcome and predictor variable and each threshold variable. The variance (% R²) explained by each model can be found in Table 2. The relationships between the threshold metrics and each behavioral and physiological measure are shown in Figures 2 and 3.

Table 2. Percent variance (R²) explained by each threshold measure.

x indicates linear model and x² the quadratic model. Bold values indicate best fit for each audiological variable.

		C1			PTA			SII			TIQ1.5			TIQ4
		n	x	x 2	n	x	x 2	n	x	x 2	n	x	x 2	n	x	x 2
Behavioral Outcome Variables	TIN1.5	222	13	16	222	14	20	222	14	17	221	25	36
	TIN4	222	40	44	222	41	45	222	42	42				221	43	60
	FMDT	224	17	17	224	17	18	224	13	15	223	20	21
	Quiet	224	27	40	224	30	39	224	33	40	223	30	37
	Noise	224	29	33	224	30	33	224	32	32	223	25	26
	TC	224	37	43	224	40	45	224	39	39	223	40	41
	TC+Rev	224	45	46	224	44	45	224	43	44	223	33	33
Physiological Predictor Variables	DPOAE1.5	224	36	36	224	38	38	224	31	31	223	36	36
	DPOAE4	224	47	50	224	41	45	224	38	48				223	51	51
	Wave I1.5	220	16	17	220	13	14	220	10	13	219	6	8
	Wave I4	220	21	25	220	17	20	220	12	20				219	22	25
	WaveV1.5	220	4	6	220	4	5	220	2	4	219	1	3
	Wave V4	220	11	20	220	10	17	220	5	15				219	15	17
	FFR	220	2	5	220	1	4	220	1	2				219	3	3

CI, principal component 1; PTA, pure-tone average; SII, speech intelligibility index; TIQ, threshold-in-quiet at 1.5 and 4kHz; TIN, threshold-in-noise; FMDT, frequency-modulation detection threshold; TC, time-compressed words; Rev, reverberation; DPOAE, distortion-product otoacoustic emissions; FFR, frequency following response

Figure 2.

Scatterplots of each threshold metric: principal component 1 (C1 [a.u.]), 4-frequency pure-tone average (PTA [dB HL]), speech intelligibility index (SII), and thresholds-in-quiet (TIQ [dB SPL]) at 1.5 and 4 kHz; and each behavioral measure: thresholds-in-noise (TIN [dB SPL]) at 1.5 and 4 kHz , word recognition in quiet (Quiet [% correct]), word recognition in background noise (Noise [% correct]), time-compressed word recognition (TC [% correct]), and time-compressed word recognition with reverberation (TC+Rev [% correct]). Solid black lines are the linear or quadratic regression lines, whichever was a better fit.

Figure 3.

Scatterplots of each threshold metric: principal component 1 (C1 [a.u.]), 4-frequency pure-tone average (PTA [dB HL]), speech intelligibility index (SII), and thresholds-in-quiet (TIQ [dB SPL]) at 1.5 and 4 kHz; and each physiological measure: distortion-product otoacoustic emissions (DPOAE [dB SPL]) at 1.5 and 4 kHz , auditory brainstem response wave amplitudes for waves I and V (μV) at 1.5 and 4 kHz, and the frequency-following response (FFR) to a speech stimulus (a.u.). Solid black lines are the linear or quadratic regression lines, whichever was a better fit.

The threshold quantification with the strongest correlation to TIN and FMDT was TIQ when regressed with a quadratic term. TIQ accounted for 36% (F = 62 [2, 218], p <0.001) and 60% (F = 165 [2, 218], p <0.001) of the variance in TIN at 1.5 and 4 kHz, respectively, and 21% (F = 29.21 [2, 220], p <0.001) of FMDT. Speech recognition not surprisingly had the strongest relationship with thresholds quantified across multiple frequencies, that is, C1 and PTA. C1 accounted for 40% (F = 71.16 [2, 221], p <0.001) of word recognition in quiet, 33% (F = 52.5 [2, 221], p <0.001) of word recognition in noise, and 46% (F = 93.29 [2, 221], p <0.001) of time-compressed speech with reverberation. PTA accounted for 45% (F = 90.38 [2, 221], p <0.001) of word recognition score in time-compressed speech. DPOAEs were best fit with PTA and TIQ. PTA accounted for 38% (F = 68.21 [1, 221], p <0.001) of the variance in DPOAE at 1.5 kHz, while TIQ accounted for 51% (F = 225.1 [1, 221], p <0.001) of the variance at 4 kHz. C1 accounted for 17% (F = 22.68 [2, 216], p <0.001) of the variance in ABR Wave I amplitude at 1.5 kHz. TIQ accounted for 25% (F = 35.46 [2, 216], p <0.001) of the variance in ABR Wave I amplitude at 4 kHz. C1 accounted for 6% (F = 6.99 [2, 217], p = 0.001) and 20% (F = 27.38 [2, 217], p <0.001) of the variance in Wave V amplitude at 1.5 and 4 kHz, respectively. C1 also accounted for 5% (F = 5.16 [2, 214], p = 0.006) of the variance in the FFR. The residual errors of the best-fit model for each auditory measure were computed and used in the subsequent analyses.

Physiological Models of Behavioral Measures

Residual errors of each variable and best-fit threshold quantification were used as the threshold-independent portion of each variable and denoted with the suffix ‘-R’. Several threshold-independent physiological predictor variables were correlated (i.e., DPOAE-R at 1.5 and 4 kHz); therefore, a PCA was implemented to reduce the number of predictor variables that were collinear. The first two principal components accounted for 58% of the variance. As expected, the ABR-R variables (residual wave amplitudes and FFR-R) loaded together in the first component (PC1) while the DPOAE-Rs loaded together in the second (PC2; Table 3). The two principal components modeled the threshold-independent TIN-R, FMDT-R, and residual speech recognition conditions. There were no significant relationships between the physiological measures and the behavioral measures after removal of threshold-dependent variance (Table 4). Furthermore, because physiological measures did not predict behavioral measures, the mediation analyses were not performed. discussion

Table 3.

Principal component loadings for threshold-independent physiological variables

	PC1	PC2
DPOAE-R1.5	−0.148	−0.689
DPOAE-R4	−0.108	−0.647
Wave I-R1.5	−0.501	−0.042
Wave I-R4	−0.493	0.006
WaveV-R1.5	−0.416	0.276
Wave V-R4	−0.408	0.171
FFR-R	−0.365	0.014

‘−R’ denotes residual error values of regression with threshold, or threshold-independent values

Table 4.

Regression of behavioral measures of hearing using physiological measures after removal of threshold-dependent variance.

		Est.	SE	t	p			Est.	SE	t	p
TIN-R1.5	n = 221					Quiet-R	n = 224
	(Intercept)	0.01	0.17	0.07	0.943		(Intercept)	0.01	0.15	0.04	0.968
	PC1 (ABR-R)	0.06	0.07	0.91	0.366		PC1 (ABR-R)	−0.06	0.06	−1.01	0.315
	PC2 (OAE-R)	−0.03	0.03	−1.14	0.256		PC2 (OAE-R)	0.03	0.02	1.50	0.135
TIN-R4	n = 221					Noise-R	n = 224
	(Intercept)	0.01	0.16	0.08	0.940		(Intercept)	0.17	0.56	0.30	0.766
	PC1 (ABR-R)	0.01	0.06	0.09	0.928		PC1 (ABR-R)	−0.26	0.22	−1.16	0.246
	PC2 (OAE-R)	−0.01	0.03	−0.39	0.694		PC2 (OAE-R)	0.08	0.09	0.95	0.345
FMDT-R	n = 223					TC-R	n = 224
	(Intercept)	0.00	0.01	−0.02	0.987		(Intercept)	0.26	0.41	0.63	0.529
	PC1 (ABR-R)	0.01	0.01	1.44	0.151		PC1 (ABR-R)	−0.01	0.16	−0.09	0.928
	PC2 (OAE-R)	0.00	0.00	−1.63	0.105		PC2 (OAE-R)	0.10	0.06	1.55	0.124
						TC+Rev-R	n = 224
							(Intercept)	0.01	0.60	0.02	0.987
							PC1 (ABR-R)	−0.40	0.23	−1.71	0.090
							PC2 (OAE-R)	0.16	0.09	1.74	0.083

‘−R’ denotes residual error values of regression with threshold, or threshold-independent values; Est., estimated coefficients; SE, standard error; TIN, threshold-in-noise; FMDT, frequency-modulation detection threshold; TC, time-compressed words; Rev, reverberation; PC, principal component; OAE, otoacoustic emissions; ABR, auditory

The goal of the study was to assess whether clinical measures of hearing could identify TIHD resulting from cochlear or neural pathology. Hearing sensitivity has been shown to affect virtually all the behavioral and physiological measures observed in studies of TIHD. To reduce effects of hearing sensitivity on their measures of interest, most studies of NIS to date have excluded people with audiometric hearing loss. Unfortunately, TIHD and its underlying pathologies will likely be comorbid with loss of hearing sensitivity in humans, as age-related hearing loss begins as early as age 30. Only including young, healthy, normal hearing participants in a study will lower the incidence of TIHD in the experimental sample and reduce the probability of identifying TIHD or its underlying pathological cause. Our approach was to include a wide range of people with and without a history of impulse noise exposure and loss of hearing sensitivity and remove threshold-related variance from the sample before regression analyses. We hypothesized that the threshold-independent variance in several behavioral measures of hearing could be predicted by physiological measures of hearing. The behavioral measures implemented in this study were pure-tone thresholds-in-noise (TIN) at two frequencies (1.5 and 4 kHz), frequency-modulation detection threshold (FMDT) at 1.5 kHz, and word recognition score in four listening conditions: quiet, background noise, time-compressed words, and time-compressed words with reverberation. The measures were chosen for two reasons: (1) they have been empirically shown or theoretically posited to be affected by pathologies underlying TIHD, and (2) they are easily implemented in a clinical setting. On the other hand, physiological measures of outer hair cell and neural function require specialized equipment and are less likely to be implemented during the typical diagnostic appointment, though DPOAEs and ABRs are clinically available and often used in cases where behavioral measures are not reliable. The data presented does not support a relationship between the physiological and behavioral measures after removal of threshold-dependent variance.

Threshold-dependent variance in behavioral and physiological measures of hearing

The variance accounted for by hearing sensitivity varied greatly among the behavioral and physiological measures of hearing. In general, hearing sensitivity was more strongly correlated with the behavioral measures than physiological measures. The C1, PTA, and TIQ performed, or were correlated with the auditory measures, at a similar level. Speech intelligibility index was the poorest predictor of auditory measures, likely because there was a ceiling effect due to half of the sample having thresholds within normal limits. Neither centering nor rationalized arcsine transformation mitigated the issues the distribution had on the regression analyses. TIN and DPOAEs were highly correlated with PTA and TIQ. A strong linear relationship at 1.5 and 4 kHz was expected, as this relationship has been established (e.g., Boege & Janssen, 2002; Gorga et al., 2003). More variance was explained by thresholds at 4 kHz than at 1.5 kHz, again to be expected. However, there was significant variance in OAEs left unexplained by audiometric thresholds, suggesting the audiogram does not entirely capture OHC integrity. Audiometric thresholds accounted for 20- 28% of the variance in ABR Wave I amplitude, and less for wave V and FFR, which is expected given the decreased proximity of generators for each consecutive response.

There are two limitations of this method to remove of threshold variance: generalizability and overfitting. The change from the raw variable to the residual variable is dependent on several factors that are particular to this data sample. Residuals depend on the relationship between threshold and the variable across the sample, not the individual; therefore, the same participant placed within a new sample will likely have a different residual value. Furthermore, because our sample is not randomly drawn from the U.S. population, and rather dependent on factors such as our targeted recruitment efforts for people with noise exposure and equal numbers of people with and without audiometric hearing loss the relationship would likely be different in a new sample conducted by a different laboratory. However, it may still be a viable way to reduce the effects of audiometric hearing loss on auditory measures within a particular study, as it is similar to building a hierarchical regression model. The benefits of finding threshold-independent variance before building a model is that it centers the data for group comparisons in a meaningful way, such that the residual values show how much of that variable cannot be accounted for by audiometric hearing loss. Moreover, it removes collinearity of audiometric hearing loss from the predictor variables. A potential issue with removing the variance accounted for by audiometric loss is overfitting and removal of too much variance, interfering with a potential relationship between the physiological and behavioral measures. To examine this possibility, we also performed the regression analyses on the raw variables and included an audiometric variable. There were no significant predictors of TIN_1.5 except for TIQ and DPOAEs and no significant predictors of FMDT other than TIQ. Interestingly, TIQ, DPOAEs, and waves I and V amplitudes of the 1.5 kHz (but not the 4 kHz) ABR were statistically significant in predicting TIN₄; however, wave amplitudes were not significant after correcting for multiple comparisons.

Thresholds in noise, frequency-modulation detection, and speech recognition scores are not predicted by physiological measures

In the present study, TIN, FMDT, and speech recognition scores were not explained by measures of sensorineural physiology after removal of threshold-dependent variance, suggesting these behavioral measures cannot act as a proxy measure of threshold-independent cochlear pathology. Most studies of NIS and HHL recruit participants with audiometric thresholds within normal limits and therefore do not account for audiometric thresholds in their analyses. Drawing on the example mentioned earlier, if we did not remove the threshold-dependent variance, we would have found a significant relationship between TIN₄ and wave I amplitude (though only of the 1.5 kHz tone-burst response). Figure 4a shows the relationship between TIN₄ and wave I amplitude before removal of threshold-dependent variance. Participants with the highest (or worst) TIN had the smallest wave I amplitudes (F = 14.94, p < 0.001). This relationship disappears when either participants with audiometric hearing loss (defined as >15 dB in this study) are removed (F = 0.07, p = 0.79; Fig. 4b), or when threshold-dependent variance is removed from TIN and ABR wave amplitude from all participants (F = 0.22, p = 0.63; Fig. 4c).

Figure 4.

The significant relationship between TIN at 4 kHz and ABR wave I amplitude (a; n = 220) dissolves when either the participants with audiometric thresholds >15 dB HL are removed (b; n = 99) or when threshold-dependent variance is removed from both variables (c; n = 220).

The lack of relationship between TIN, FMDT, and speech recognition scores and physiology are limited to the physiological measures chosen for this study. There are additional measures that have been proposed to identify site-of-lesion for TIHD, including the SP/AP ratio of the electrocochleogram (Liberman et al., 2016), ABR wave I curvature (Bao et al., 2022; Zhang et al., 2023), the envelope-following response (Bramhall et al., 2021; Shaheen et al., 2015), high-stimulus-rate ABR (Schirmer et al., 2024), and middle ear muscle reflex (Bramhall et al., 2022; Mepani et al., 2018; Shehorn et al., 2020; Valero et al., 2016; Wojtczak et al., 2017). These were not explored in the current study. Furthermore, certain behavioral and physiological measures of hearing are more reliable than others. In a prior study we assessed the test-retest reliability of threshold-independent variance in all the measures used in the present study (Kamerer, Kopun, Fultz, Neely, et al., 2019). In general, we found that the physiological measures had better reliability than the behavioral measures. Threshold-independent OAE amplitudes had good-to-excellent reliability and ABR wave I amplitudes had moderate reliability. On the other hand, ABR wave V amplitudes and the FFR had poor test-retest reliability. TIN, FMDT, and speech recognition in the four listening conditions had poor-to-moderate reliability. Poor within-subject reliability will affect the ability to find significant and meaningful between-subject relationships, and all studies of HHL and TIHD are suffering from this issue.

Implications for research

The aim of this study was to investigate whether several proposed behavioral measures of HHL, NIS or TIHD, which can be easily implemented clinically, can be linked to physiological measures of sensory and neural function. As hearing sensitivity affects many auditory measures, it was important to remove this effect before investigating any relationships. After removal of threshold-dependent variance, we found no relationship between TIN, FMDT, or word recognition in difficult listening conditions and OAEs or ABR metrics. This suggests that the remaining variance in TIN, FMDT, and word recognition that cannot be explained by thresholds also cannot be explained by our current measures of OHC, AN function, or some higher order neural function (contributing to wave V and the FFR). We know that some of this unexplained variance can be accounted for by cognitive capacity (e.g., Bosen & Barry, 2020; Humes, 2021; Merten et al., 2022; Roque et al., 2019). Our previous study found that 17% of the variance in TIN at 4 kHz could be explained by short term and working memory and processing speed, while 41% of the FMDT was explained by attention and working memory capacity. Sixteen percent of TC and 22% of TC+Rev could be explained by working memory and executive function (Kamerer, Aubuchon, Fultz, Kopun, et al., 2019). As we move forward in developing clinical tests for people with HHL, NIS, or TIHD, we will need to determine what these tests are actually measuring. It is only once we learn the mechanisms underlying these disorders that we can develop a truly effective treatments for the many patients who suffer from unexplained hearing deficits.

Conclusions

1. Thresholds accounted for between 22–68% of the variance in our behavioral measures of hearing (TIN, FMDT, and word recognition), and between 5–54% of the variance in our physiological measures of hearing (DPOAES, ABR wave amplitudes, and FFR).

2. There was no evidence that TIN, FMDT, or word recognition in quiet, noise, time-compression, or reverberation could be predicted by DPOAEs, waves I and V amplitudes of the ABR, or the FFR after removal of threshold-dependent variance.

3. TIN, FMDT, and word recognition in quiet, noise, time-compression, or reverberation may not be able to detect outer hair cell or auditory nerve dysfunction that is independent from threshold.