Predicting neural deficits in sensorineural hearing loss from word recognition scores

The current gold standard of clinical hearing assessment includes a pure-tone audiogram combined with a word recognition task. This retrospective study tests the hypothesis that deficits in word recognition that cannot be explained by loss in audibility or cognition may reflect underlying cochlear nerve degeneration (CND). We collected the audiological data of nearly 96,000 ears from patients with normal hearing, conductive hearing loss (CHL) and a variety of sensorineural etiologies including (1) age-related hearing loss (ARHL); (2) neuropathy related to vestibular schwannoma or neurofibromatosis of type 2; (3) Ménière’s disease; (4) sudden sensorineural hearing loss (SSNHL), (5) exposure to ototoxic drugs (carboplatin and/or cisplatin, vancomycin or gentamicin) or (6) noise damage including those with a 4-kHz “noise notch” or reporting occupational or recreational noise exposure. Word recognition was scored using CID W-22 monosyllabic word lists. The Articulation Index was used to predict the speech intelligibility curve using a transfer function for CID W-22. The level at which maximal intelligibility was predicted was used as presentation level (70 dB HL minimum). Word scores decreased dramatically with age and thresholds in all groups with SNHL etiologies, but relatively little in the conductive hearing loss group. Discrepancies between measured and predicted word scores were largest in patients with neuropathy, Ménière’s disease and SSNHL, intermediate in the noise-damage and ototoxic drug groups, and smallest in the ARHL group. In the CHL group, the measured and predicted word scores were very similar. Since word-score predictions assume that audiometric losses can be compensated by increasing stimulus level, their accuracy in predicting word score for CHL patients is unsurprising. The lack of a strong age effect on word scores in CHL shows that cognitive decline is not a major factor in this test. Amongst the possible contributions to word score discrepancies, CND is a prime candidate: it should worsen intelligibility without affecting thresholds and has been documented in human temporal bones with SNHL. Comparing the audiological trends observed here with the existing histopathological literature supports the notion that word score discrepancies may be a useful CND metric.

Thus, audiology best practices for hearing evaluation in adults, i.e. a standard audiogram and a word-recognition task in quiet 19 , may not fully assess the hearing impairment associated with CND. While a number of clinical centers have enhanced evaluation protocols (e.g., speech-in noise testing, evaluation of extended high frequencies), particularly in patients reporting difficulty in noisy environments despite normal or near-normal audiometric thresholds, retrospective studies of CND must rely on traditional audiometric data.
One tool of particular interest is the speech intelligibility curve (SIC), which describes the cumulative distribution of useful speech information as word presentation level is increased 20 . This sigmoidal performance-intensity function is computed from the patient's audiogram combined with a transfer function appropriate to the speech material. In absence of multiple word presentation levels, the SIC can be a useful predictor of the maximal word recognition score (WRS). However, the assumptions behind the SIC treat a hearing loss as if it were simply a frequency-specific attenuation of the incoming speech sounds. This is largely true for a conductive hearing loss (CHL), i.e. one due to dysfunction in the sound transmission apparatus of the external and middle ears. However, for sensorineural hearing loss (SNHL), i.e. impairments arising from pathologies in the transduction of sound-induced vibration by the inner ear's sensory cells into action potentials in cochlear nerve fibers, the pathophysiology is much more complex and cannot, in general, be corrected by simply increasing stimulus level.
In SNHL, the elevation of pure-tone thresholds is typically due to loss of, or damage to, the outer hair cells, which normally function as biological motors, amplifying the sound-evoked vibrations of the sensory epithelium 21 . In addition to threshold elevation, outer hair cell loss broadens the frequency tuning of cochlear neurons and changes their response phase, which can desynchronize the ensemble neural signal 22,23 . These changes in the way sounds are coded in the cochlear nerve cannot be corrected by amplification, and any loss of cochlear neurons will further degrade the information being transmitted to the central nervous system.
These basic concepts suggest that the discrepancies between predicted and measured word scores should be smaller in those with CHL than those with SNHL of similar magnitude. Furthermore, among those with SNHL of similar magnitude, and thus a similar degree of sensory cell damage, word-score discrepancies should be greatest in those with the most CND. To test these ideas, we analyzed data from nearly 96,000 ears from patients seen at the Massachusetts Eye and Ear, with normal hearing, CHL or SNHL of several etiologies, including age-related hearing loss (ARHL), noise damage, ototoxic drugs, sudden SNHL, Ménière's disease and vestibular schwannomas. We observed that the WRS discrepancies were largest in those SNHL etiologies associated with the greatest degree of CND, consistent with the idea that loss of cochlear neurons is a major contributor to the loss of speech intelligibility in SNHL.

Materials and methods
Study populations. We collected audiological data from a large sample of adult patients who underwent a comprehensive hearing evaluation at the Massachusetts Eye & Ear between 1993 and 2017 (see Table 1). We considered several groups of patients including those with (1) normal hearing thresholds (≤ 20 dB HL across test frequencies, n = 36,855 ears) and (2) ARHL only; i.e., with an idiopathic bilateral and symmetrical SNHL (n = 44,967 ears). All these patients had differences ≤ 10 dB between air-conduction (AC) and bone-conduction (BC) thresholds (air-bone gap) at any test frequency and/or did not have an interaural asymmetry ≥ 10 dB in AC thresholds at 2 consecutive frequencies or 15 dB at one test frequency. As a kind of control for cognitive decline, we added a third group of patients with conductive hearing loss only (n = 7396 ears); i.e., patients with an air-bone gap ≥ 15 dB at any test frequency and normal bone conduction thresholds. To assess the effect of known etiologies with sensorineural deficits we added groups of patients with (1) neuropathy including a diagnosis of vestibular schwannoma (n = 664 ears) or neurofibromatosis of type 2 (n = 282 ears); (2) Ménière's disease (n = 628 ears without and n = 128 ears with trans-tympanic injection of gentamicin); (4) a history of sudden (S) SNHL (n = 156 ears); (3) exposure to ototoxic drugs including carboplatin or cisplatin (n = 1135 ears), vancomycin (n = 32 ears) or gentamicin (n = 42 ears); and (5) self-report of recreational or occupational noise damage Hearing evaluation. Audiometric thresholds were obtained using a number of different audiometers including Grason-Stadler (GS-10, GS-16), Interacoustics AC-30, Virtual 320 and Interacoustics Equinox, running under the same Harvard Audiometer Operating System (AOS) 24 . Only pure-tone air-conduction (AC) thresholds measured at standard audiometric frequencies from 0.25 kHz to 8 kHz (in octave steps) using TDH39 headphones or ER-3A insert earphones were considered. Bone-conduction (BC) thresholds were acquired from 250 Hz to 4000 Hz with a Radioear B-71 vibrator over the mastoid. Word recognition was assessed using a recorded CID (Central Institute for the Deaf) W-22 test, consisting of 50 monosyllabic word lists presented with a contralateral masker (speech-shaped noise). The Articulation Index (AI) was used to predict the speech intelligibility curve (SIC, Fig. 1B), a speech intelligibility performance as a function of presentation level based on the audiogram 25,26 (Fig. 1A), using a transfer function for CID W-22 27 . This procedure was automatically generated by the Harvard AOS software as previously described 28 . The level at which maximal intelligibility was predicted (PB max ) was chosen as the presentation level. If this value, however, fell below 70 dB HL, presentation level remained at 70 dB HL. To assess word recognition deficits that cannot be accounted by a loss of audibility, we measured either (1) the WRS obtained at predicted PB max and (2) the difference between the measured vs. predicted WRS (∆WRS) as defined by the SIC in each patient (Fig. 1B).

Statistics.
Using Matlab R2018a (The Mathworks, Natick, MA), Pearson correlation coefficients were used to assess relationships between pure-tone thresholds and age. Ordinary least square regression was used to estimate the effect of age on pure-tone thresholds and word recognition scores. Visual inspection, medians and non-parametric Wilcoxon rank-sum tests were used to compare the WRS of groups according to pure-tone thresholds.
To assess the relative contribution of each test frequency to WRS, we designed the following model: WRS or ∆WRS ~ AC 250Hz + AC 500Hz + AC 1kHz + AC 2kHz + AC 4kHz + AC 8kHz , whereby AC refers to Air-Conduction threshold. We fitted the regression using the linear model "lm" function in R software environment with the default parameters and the above model. We then computed the respective commonality coefficients and determined the unique and common effects for each predictor variable using R 29 with the statistical package described by Nimon and col 30 .
To assess the impact of hearing sensitivity at each test frequency on word recognition, we used SAS (SAS Institute, version 9.4) to conduct a principle component analysis (PCA) of the thresholds at each test frequency. These analyses provided a set of 6 linearly independent components as well as the proportion of the total variance in the data that is explained by each component. We use these components as predictors in linear regression models with WRS as outcomes.  Figure 1. Illustration of the relationship between audiometric thresholds and the performance-intensity function for speech recognition (i.e., speech intelligibility curve). The speech intelligibility curve in (B) is drawn from the amount of speech material that is audible to a patient with a specific hearing loss (A) combined with a transfer function appropriate to the speech material. In (A), W-22 phonetically balanced words presented at either 20 dB HL (green), 60 dB HL (orange) or 100 dB HL (red). To index the loss of intelligibility in word recognition that cannot be accounted by a loss of audibility, we also calculated the difference between the measured word recognition score and the one predicted by the SIC in each patient (B). The threshold for statistical significance was p = 0.05.

Results
The natural history of ARHL. In total, 81,822 ears (59% female, 41% male) from patients aged 18 to 103 years old, presenting with either normal (n = 36,855) or elevated thresholds (n = 44,967) met our inclusion criteria for the normal hearing or ARHL groups, respectively (see Fig. 2, Table 1). Patients with normal hearing were more often female (64%) than those with ARHL (55%). As shown in Fig. 3A, thresholds worsened progressively with age (r = 0.56; p < 0.001 for PTAs), remaining within normal limits until around age 50 and progressing to a down-sloping, moderate to severe loss in our oldest patients. The correlation between age and threshold was stronger at high frequencies (r = 0.74; p < 0.001 for 8 kHz) than low frequencies (r = 0.41; p < 0.001 for 0.25 kHz). Across all ages, the rate of threshold shift ranged from 0.25 dB/year at 0.25-kHz to 1.12 dB/year at 8-kHz; however, the rate of threshold deterioration also changed with age. As seen in Fig. 3B, threshold shifts increased most rapidly at the highest frequency (8 kHz) until about age 75, after which time any additional threshold shifts were relatively flat across frequency. Correspondingly, threshold shifts at low frequencies (≤ 1 kHz) rose slowly until about age 65, then began to rise rapidly, and more or less in unison, for the next 20 years.
Sex differences in the progression of ARHL are also interesting. Plotting the difference between males and female thresholds as a function of age (Fig. 3C) shows that aging males tend to lose their hearing sensitivity first, in a pattern that peaks at 4 kHz, strongly suggesting an etiology of acoustic overexposure [31][32][33] . The rate of progression of this 4 kHz notch peaks at age 60, at which time females begin to lose sensitivity faster than males, but only at the three lowest test frequencies.  www.nature.com/scientificreports/ Impact of hearing loss configuration on intelligibility deficits. We examined the differences between predicted and measured scores on the standard clinical test of word recognition in quiet. When the measured WRS is lower than the predicted WRS, it means the deficit is not simply one of audibility, and one contributing factor can be loss of auditory nerve fibers, i.e. CND. WRS predictions are derived from the Speech Intelligibility Curve (SIC), which treats a threshold shift like an acoustic filter applied to the words, and quantifies the fraction of the speech spectrum remaining in the filtered output as a function of presentation level 20,25,34 .
Because word presentation level (as dictated by the SIC) is well above threshold, a very large degree of hearing loss is needed to reduce the predicted WRS. Accordingly, > 99% of our cases have a predicted word score > 95% correct. In practice at the Massachusetts Eye and Ear, word tests are presented at a level at which the SIC has asymptoted (PB max ), or at 70 dB HL, whichever is lower. The mean WRSs in our normal hearing and ARHL groups declined monotonically (Fig. 4A), in a roughly sigmoidal fashion, with increasing threshold shift, whether expressed as PTA (averaging 0.5-2 kHz) or as mean AC thresholds (averaged across all test frequencies). The WRS discrepancy (predicted vs. measured) rolls over as threshold shift increases (Fig. 4B), because past a certain level of hearing loss, the predicted WRS falls rapidly.
For one set of analyses, we considered only patients from the normal hearing and ARHL groups with predicted WRSs ≥ 99%, then averaged data according to the measured WRS (Fig. 5). The mean audiogram for those with near-perfect WRSs (grey symbols in Fig. 5A) shows that sensitivity at the highest frequencies is not critical for intelligibility of words in quiet. As the measured WRS declines, thresholds deteriorate, most markedly and steadily at low frequencies, expected to be most important for speech (Fig. 5A). This variation in WRS discrepancy was not related to presentation level ( Fig. 5B) or age (Fig. 5C), suggesting that age-related cognitive decline was not a confound in these results. No effect of sex on either relation was found (data not shown).  Mean air-conduction thresholds at standard audiometric frequencies from all patients with sensorineural hearing loss whose word recognition performance is predicted ≥ 99% by the speech intelligibility curve as a function of measured word recognition score. (B) Speech intelligibility curves for each subgroup of patients described in (A) with their associated predicted and measured word recognition score. (C) Intelligibility deficits assessed by calculating the difference between measured word score and score predicted by the speech intelligibility curve in each subgroup of patients described in A as a function of age. Error bars are standard error of mean. www.nature.com/scientificreports/ To further clarify the relative contribution of different cochlear regions to the WRS and ∆WRS, we took two approaches. First, we used a linear regression model to compute commonality coefficients to determine the unique and common contribution of each test frequency to the explained variance (R-squared) in the regression model; i.e., to WRS and ∆WRS (see "Materials and methods"). As shown in Fig. 6, a loss in sensitivity at 1-2 kHz produced the largest impact on WRS and ∆WRS. Second, we conducted a principle component analysis, as summarized in Table 2. The first component, which weights all the frequencies equally (see Table 2, column PC1 of Eigenvectors of the covariance matrix) and thus corresponds to the degree of hearing loss, explains ~ 78% of the variance (see Table 2, Eigenvalues of the covariance matrix). Explained variance rises to ~ 92% if we add the second component, which tracks the slope of the audiogram. Finally, ~ 95% of the variance is explained if we add the third component which tracks the weight of the center vs. adjacent test frequencies ( Table 2).

Impact of age and cognition on intelligibility deficits. Among our subjects with ARHL, intelligibility
worsened monotonically with age ( Fig. 7A): slowly at first (0.08%/year for those 20-50 years), then more rapidly (0.98%/year for ages over 65). We estimated the contribution of cognitive decline to these trends by comparing our data to published metrics of verbal comprehension, perceptual reasoning, working memory and verbal memory 35 . Converting both to z-score measures as a function of age (Fig. 7B) suggests that the WRS decline is more rapid than general cognitive decline.
Further evidence that cognitive decline is not a major confound is seen in the group with CHL, which shows minimal effects of age on WRS (Fig. 8A). Unfortunately, the age range for the CHL group is limited, because most patients past 80 years have at least some SNHL. As confirmation of the utility of the SIC, only a small decline in WRSs was observed in the CHL group with increasing audiometric loss (Fig. 8C), presumably because effects of CHL are indeed similar to those of an acoustic filter. To further support of our interpretation of WRS discrepancies, we compared age-matched patients from the CHL (Supplementary Intelligibility deficits in SNHL of other etiologies. If the WRS discrepancy is associated with CND, it should be large in SNHL etiologies where the underlying pathology destroys cochlear nerve fibers. To test this notion, we identified a neuropathy group with either vestibular schwannoma (VS, n = 664 ears) and/or neurofibromatosis of type 2 (NF2, n = 282 ears), both of which are known to cause damage to the cochlear nerve 36 . The VS sample was slightly younger than the ARHL group (Fig. 9A), and the NF2 patients showed a bimodal age distribution at time of test (Fig. 9B), consistent with epidemiologic data 37,38 . As expected, both groups had significantly larger WRS discrepancies compared to the ARHL group (p < 0.001 for both) or to the opposite healthy ear (Fig. 9C, p < 0.001). The differences in ∆WRS for both etiologies were especially striking when plotted as a function of age or hearing loss (Fig. 9D-F, Supplementary Table 1).
We also included a group with Ménière's disease, because massive CND was reported in a case of unilateral Ménière's using electron microscopy 39 . In our study, we separated patients who did not receive gentamicin to control vertigo (n = 628 ears) from those who did (n = 128 ears). The Ménière's groups both had similar age distributions to the ARHL group (Fig. 10A,B), but their WRSs were dramatically poorer (Fig. 10C, p < 0.001 for both groups). When compared as a function of age or hearing loss, both groups had worse word scores than ARHL patients (Fig. 10D-F, Supplementary Table 1). In addition, Ménière's patients who received gentamicin scored worse than those who did not (Fig. 10C,D, p < 0.001). However, these differences (gentamicin vs. no gentamicin) disappeared when the degree of hearing loss was considered (Fig. 10E,F, p > 0.05), suggesting that differences were related to loss in audibility from gentamicin-induced outer hair cell loss.
In a recent study, we showed that patients who recovered from idiopathic SSNHL had poorer WRSs than predicted by the residual loss of audibility 40 , and CND is one histopathological feature of SSNHL 41 , as well as of viral infection and ischemia, the main two candidate etiologies of SSNHL [42][43][44][45][46] . Here, we identified a SSNHL group (n = 156 ears) with similar age distribution to the ARHL group (Fig. 11A). SSNHL patients had significantly worse WRS in the affected ear compared to either their own other ear or the ARHL group (Fig. 11B,   www.nature.com/scientificreports/   Fig. 4) are compared to patients presenting with a CHL. Intelligibility deficits were assessed either by calculating the difference between measured word score and score predicted by the speech intelligibility curve (B) or by measuring word recognition scores at PB max (C). Error bars are standard error of mean.  Table 1). To these three SNHL etiologies, we added patients who received systemic ototoxic drugs, either (1) platinum-based cancer therapeutics (cisplatin and carboplatin, n = 1135 ears) or an ototoxic antibiotic, either (2) vancomycin (n = 32 ears), a glycopeptide, or (3) gentamicin (n = 42 ears), an aminoglycoside. All these drugs are known to cause hearing loss, as measured in a standard audiogram 47 . In general, the WRS discrepancies in these groups were similar to those in the normal-aging group. In the gentamicin group, word scores were statistically indistinguishable from patients with age-matched controls (Fig. 12D, p > 0.05), whether considered as a function of age or hearing loss (Fig. 12E-G, Supplementary Table 1). Likewise, ∆WRSs were similar overall in patients treated with vancomycin (Fig. 12D, p > 0.05) including as a function of age (Fig. 12E, Supplementary Table 1); however, significantly poorer word scores were noted when considered as a function of hearing loss (Fig. 12F,G, Supplementary Table 1). In the chemotherapeutics group, both WRS and ∆WRS were poorer overall (Fig. 12D, p < 0.001) and when considered as a function of age (Fig. 12E). Scores were better than the ARHL group when considered as a function of hearing loss (Fig. 12E-G, Supplementary Table 1), perhaps because the drugs destroy outer hair cells, thereby increasing thresholds, without worsening the CND that comes with normal aging.
Finally, we identified patients with putative noise-induced hearing loss, either presenting with an audiometric notch ≥ 20-dB at 4 kHz (n = 1448 ears) or with an explicit history of occupational or recreational noise exposure (n = 2091 ears). In these younger and male-dominated groups (Fig. 13A,B), ∆WRS was overall significantly poorer only in patients presenting with an audiometric notch (p < 0.001). However, small but significantly poorer WRSs were noted as a function of age or hearing loss when measured with the PTA (Fig. 13D-F  www.nature.com/scientificreports/

Effect of age and sex on hearing thresholds
Here, we gathered data on hearing sensitivity from a large number of adult patients with ages covering the lifespan. Our cohort might not reflect hearing status of the general population, because it includes only people who sought the care of an otologist or audiologist. Nevertheless, our value for hearing-loss prevalence (55%) is similar to that in the Beaver Dam study (80%) of a more random sample 48 if we restrict our age range to match theirs (i.e. 48 to 92 years) and use similar definitions of hearing loss. The overall hearing loss prevalence in the Framingham study was even higher (83%) [49][50][51] . Clearly, hearing loss is rampant and may be underestimated here, since these studies included more females than males (58% vs. 42% in the present study), and females tend to have better hearing at most ages than males (Fig. 1). It is worth noting that 45% of patients in our cohort had "normal" hearing. While some of this arises because some of the audiometric referrals were unrelated to an ear or hearing complaint (e.g., a physical exam or a hearing evaluation for a dizziness complaint), this rate emphasizes that many patients report a hearing issue that is not captured by a standard hearing test.
Hearing sensitivity decreased with age and increasing test frequency (Fig. 3), following the classic downsloping pattern of ARHL (for review, see 52 ). At the highest frequencies, threshold decline was significant even from the youngest ages examined (18 yrs), and mean thresholds fell outside the "normal" range (20 dB HL) by age 45, consistent with previous studies 53, 54 .
The sloping audiometric pattern of ARHL is consistent with the basal-turn loss of outer hair cell loss in aging humans 3 . However, the most comprehensive histological analyses show prominent apical and basal loss of outer hair cells, with a relative sparing of mid cochlear regions 3,55,56 . Apical OHC degeneration is not well reflected in the audiogram, because (1) the lowest frequency tested (250 Hz) is far (~ 14%) from the apical extreme of the human cochlea 57 , (2) the contribution of OHC electromotility to cochlear amplification decreases with increasing frequency 21 and (3) low-frequency auditory-nerve fibers are so broadly tuned that apical losses are hidden by neighboring cochlear regions with minimal increase in sound pressure level 58 . Despite the minimal effect on the audiogram, the loss of information due to the silencing of apical neurons must decrease the fidelity of stimulus representation especially in difficult listening environments. www.nature.com/scientificreports/ As shown in Fig. 3C, there were significant sex differences in hearing loss progression across the lifespan. Before age 65, loss was more prominent in males, particularly at 4 kHz, a configuration seen in patients with a history of noise exposure 32,59 . Increased hearing loss in males has been linked to occupation, noise exposure history and income level 49,50,60,61 . However, even after adjusting for these risk factors, sex differences remained significant, and smoking, atherosclerosis and other factors have been suggested as possible contributors 48 . Past age 65, hearing loss accelerated at the lower frequencies (≤ 1 kHz), with larger losses in females (Fig. 3C). This low-frequency hearing loss in older women has been associated with vascular disease 62 and may be related to the degeneration of the stria vascularis, which is worse in apical cochlear region in aging humans, where it would be expected to cause a low-frequency hearing loss 3 .

Impact of hearing thresholds and cognition on intelligibility.
The speech intelligibility index, i.e. the proportion of the total speech information audible to the listener 63 , can be used to predict the WRS, based on the audiogram and the spectrum of the speech tokens used. However, despite the corrections applied to the speech intelligibility curve 64 to account for upward spread of masking and the negative impact of high presentation levels 65 , the loss of outer hair cells that typically underlies threshold elevation also degrades frequency tuning and critical cochlear non-linearities 22,66 that can reduce intelligibility. Correspondingly, in a purely conductive hearing loss, where damage to outer hair cells is minimal, the measured and predicted WRSs are quite similar (Fig. 8). Nevertheless, the large range in word scores seen with similar patterns and degrees of mild to moderate SNHL (Fig. 5) argues against an exclusive role for outer hair cell damage in the loss of intelligibility and leaves CND as a prime candidate. The mapping of the hearing loss frequencies on WRS (Fig. 6) and the principle component analysis (Table 2) identify the 1 kHz region as the frequency band most important to the WRS, which is consistent with the peak innervation density in humans 2 , as well as the peak of the spectrum for word tests.
Word recognition is linked to cognition 67 , thus a decline in performance may be related to age-related cognitive decline 35,68 . Our results argue against cognitive factors as a major contributor to the continuous decline in WRS with age given that: (1) deficits in WRS with age are minimal in patients with conductive hearing loss when compared to age-matched patients with sensorineural hearing loss (Fig. 8); (2) age-related cognitive declines are not as steep as the decline in WRS with age (Fig. 7); and (3) WRS is highly variable (< 40% to 100%) in patients of similar age (Fig. 5C).

Association of intelligibility deficits with CND.
We know from animal studies that destruction of cochlear nerve fibers has little impact on audiometric thresholds, until it becomes extreme, i.e. > 80% loss [4][5][6] . Our working hypothesis is that this CND contributes more prominently to degradation of WRSs, and that differences in the degree of CND are key contributors to the differences in WRSs we observe in different etiologies. Although histopathology of human CND is not extensive, it is informative to consider the present results in light of the available data. www.nature.com/scientificreports/ Recent work on normal-aging humans shows that cochlear neural degeneration precedes the loss of the inner hair cells they innervate 2,69 . The linear best-fit between age and CND (r 2 = 0.8) in a sample of 26 cases suggested a mean neural loss of 46% by age 60 years (averaged across all the audiometric frequencies), and an extrapolated loss of 69% by age 90 2 . Referring to the mean WRSs for normal-aging 60 y.o. found here (Fig. 5C), suggests that 46% loss of cochlear nerve fibers has minimal impact (< 5%) on word recognition in quiet, whereas 69% neural loss diminishes word scores by 63%. These two points begin to populate a summary of the estimated relation between word score and CND (black curve in Fig. 14). These inferred relations between CND and WRS suggest an upper bound on the CND contribution, because the WRS discrepancies also include contributions from (1) the degradations in auditory nerve response introduced by loss of frequency tuning, (2) other effects of hair cell damage that are not compensated by raising presentation levels and (3) response alterations in the central auditory pathways elicited by these changes in the periphery 6,12,70,71 . Indeed, the differences in ∆WRSs (Fig. 8A) between the normal-aging cases (with outer hair cell damage) and the CHL cases (without) may provide a measure of the magnitude of that effect.
Patients with explicit neuropathies, i.e. vestibular schwannoma (VS) and/or neurofibromatosis of type 2 (NF2) showed poorer WRSs than the normal-aging group (Fig. 9). VSs are benign tumors in the vestibular nerve that typically arise unilaterally. They can cause hearing impairment, presumably by affecting the blood supply from pressure-induced vascular changes or release of cytokines 36,72,73 . Many VS patients have tinnitus and elevated or absent middle-ear muscle reflexes, both of which have been associated with CND [74][75][76][77] . Recent histopathological analysis of temporal bones from five unilateral VS cases (mean age 79 years) revealed a 34% mean loss of the peripheral axons of cochlear nerve fibers in the affected vs. contralateral ears 78 . Assuming the latter are normal for their age, as the present WRS data suggest (Fig. 9C), we estimate 13% survival of cochlear neurons in an 80 y.o. VS patient (i.e. 34% of 39%), which was associated with an average word score of 24% 78 . These estimates fit well on the relation between WRS and CND extrapolated from the normal aging group (Fig. 14).
Of all the etiologies we studied, patients with Ménière's disease showed the worst WRSs (Fig. 10). This debilitating disorder causes episodic vertigo, fluctuating hearing loss and roaring tinnitus 79,80 , with relatively little loss of either inner or outer hair cells 41  www.nature.com/scientificreports/ the WRS in the affected ear of this patient was 34%, compared to a predicted score of 99.3%, based on the speech intelligibility curve and the audiogram. These values also fit well on the summary curve in Fig. 14.
Patients with SSNHL also showed discrepancies between measured and predicted WRSs that were larger than in normal-aging controls (Fig. 11C), but the differences were not as dramatic as those in Ménière's or VS patients. The etiology of SSNHL is poorly understood, and likely includes a number of underlying causes, including viral infection and cochlear vascular compromise 82,83 . A survey of cochlear histopathology in SSNHL cases with unrecovered thresholds concluded that roughly half (7/15) showed marked CND 41 ; however, the neural loss has not been well quantified enough to make any meaningful comparison to the other etiologies.
Of all the etiologies studied, the smallest differences between measured and predicted word scores were seen in cases of ototoxic drugs (Fig. 12) and putative noise-induced damage (Fig. 13). We found no relevant quantitative data in the literature on CND after hearing loss from ototoxic drugs; however, a recent comparison of human temporal bones with vs. without a history of noise damage 3 found a small increase in CND (~ 15%) across all audiometric frequencies in the latter, but only among patients aged 50-74 years and not among individuals from 75 to 100 years. This age effect is opposite to that seen here, where differences from normal-aging were maximal in the older group. However, both studies agree that the added effects of noise on CND in the aging ear are small. Furthermore, both studies acknowledge the difficulties in identifying normal-aging humans who have not suffered ear abuse, and in relying on self-report as a reliable metric of cumulative noise exposure. That problem is evidenced here by the presence of a 4 kHz notch in the mean audiograms of the normal-aging group (Fig. 3), for which a history of acoustic overexposure was not noted. Nevertheless, the histopathological study showed that including the degree of CND in a multivariable linear regression of histopathological metrics, including hair cell loss and strial damage, helped predict the measured WRS.

Summary and conclusion
We examined audiometric thresholds, word recognition in quiet and the differences between predicted and measured word scores in nearly 96,000 ears with either normal hearing, ARHL (i.e., idiopathic) or SNHL attributable to diagnosable disease or cochlear insult of several etiologies. Intelligibility worsened monotonically with age and degree of hearing loss. Results from patients with conductive hearing loss suggested that cognitive decline was not a major confound. In SNHL, larger intelligibility deficits were observed in etiologies known, or suspected, to cause greater cochlear nerve loss. After accounting for age and degree of hearing loss, word performance deficits were greatest (in descending order) in Ménière's disease, in those with acoustic tumors and after sudden sensorineural hearing loss. Intelligibility deficits were considerably smaller in those with a history of exposure to noise or ototoxic drugs (either aminoglycosides or platinum-containing chemotherapeutics). This ranking is consistent with the fragmentary data on CND in human temporal bones. A compilation of those data suggests that auditory nerve loss must exceed 60% before word scores in quiet fall below 90%. The relation between CND and performance on more difficult listening tasks could be much steeper.
These findings agree with a number of studies linking speech perception or signal-in-noise detection with neural deficits measures by ABRs/electrocochleography [84][85][86][87] , middle-ear muscle reflex 76,77 , envelope following responses 88 , in vivo imaging of auditory nerve diameter 89 and computational models 13 . While the evidence implicates CND in the intelligibility challenges of hearing impairment, CND also likely serves as a peripheral trigger for excess central gain and other maladaptive plasticity in central auditory pathways that further degrade perceptual accuracy for complex sounds beyond that predicted by viewing the audiogram as a simple acoustic filter 12,70,90 .  Figure 14. Prediction of word scores as a function of cochlear neural survival. Neuronal survival as a function of age was estimated from a best fit straight line to the data in Fig. 6A from a histopathological study of normal aging humans 3 . Word scores as a function of age were obtained from the present study (Fig. 6A). See text for further details.

Data availability
All relevant data are within the paper and its Supporting Information files. Raw data that support the findings of this study are available from the corresponding author, upon reasonable request. www.nature.com/scientificreports/