Cochlear-implant simulated signal degradation exacerbates listening effort in older listeners

Abstract

Objective:

Individuals with cochlear implants (CIs) often report that listening requires high levels of effort. Listening effort can increase with decreasing spectral resolution, which occurs when listening with a CI, and can also increase with age. What is not clear is whether these factors interact; older CI listeners potentially experience even higher listening effort with greater signal degradation than younger CI listeners. This study used pupillometry as a physiological index of listening effort to examine whether age, spectral resolution, and their interaction affect listening effort in a simulation of CI listening.

Design:

Fifteen younger normal-hearing (YNH) listeners (ages 18 to 31 years) and fifteen older normal-hearing (ONH) listeners (ages 65 to 75 years) participated in this experiment; they had normal hearing thresholds from 0.25 to 4 kHz. Participants repeated sentences presented in quiet that were either unprocessed or vocoded, simulating CI listening. Stimuli frequency spectra were limited to below 4 kHz (to control for effects of age-related high-frequency hearing loss) and spectral resolution was decreased by decreasing the number of vocoder channels, with 32-, 16-, and 8-channel conditions. Behavioral speech recognition scores and pupil dilation were recorded during this task. Additionally, cognitive measures of working memory and processing speed were obtained to examine if individual differences in these measures predicted changes in pupil dilation.

Results:

For trials where the sentence was recalled correctly, there was a significant interaction between age and spectral resolution, with significantly greater pupil dilation in the ONH listeners for the 8- and 32-channel vocoded conditions. Cognitive measures did not predict pupil dilation.

Conclusions:

There was a significant interaction between age and spectral resolution, such that older listeners appear to exert relatively higher listening effort than younger listeners when the signal is highly degraded, with the largest effects observed in the 8-channel condition. The clinical implication is that older listeners may be at higher risk for increased listening effort with a CI.

INTRODUCTION

Individuals with hearing loss report expending more effort to communicate than individuals with normal hearing (Kramer et al. 2006). Listening effort has been defined as arising from a combination of the adversity of the situation and the mental resources engaged to overcome adversity and meet the individual’s listening goals (Lemke & Besser 2016; Pichora-Fuller 2016). Increased listening effort may lead to fatigue, which in turn may be responsible for social withdrawal, chronic stress, poor health outcomes (Pichora-Fuller 2016), and a need for more sick leave and post-workday recovery time (Kramer et al. 2006; Nachtegaal et al. 2009). Therefore, gaining an understanding of listening effort adds valuable information about auditory perception; it is possible for an individual to achieve high speech recognition levels while needing to exert significant effort to do so. Speech recognition performance and listening effort are not always aligned (Ohlenforst et al. 2017; Winn & Teece 2021; Winn & Teece 2022). Older cochlear-implant (CI) users in particular may be at a higher risk for increased listening effort (DeRoy Milvae et al. 2022), as these individuals potentially experience compounding effects of aging and auditory signal degradation that occurs during CI signal processing (Winn 2016).

Older adults often exhibit increased listening effort when compared to younger adults (Piquado et al. 2010; Desjardins & Doherty 2012; Bernarding et al. 2013; Perreau et al. 2017; Ward et al. 2017; McGarrigle et al. 2021), even when intelligibility is similar or equated between groups (Gosselin & Gagné 2011; Degeest et al. 2015). There are at least two possible reasons for the association between age and effort in the absence of differences in intelligibility: (1) differences in peripheral auditory function that are not accounted for by experimenters, and (2) declines in cognitive function and central auditory changes that occur from the natural aging process (Lemke & Besser 2016; Philips 2016; Tremblay & Backer 2016). Age-related cortical changes to non-primary auditory processing areas, leading to declines in cognitive processes such as executive function, processing speed, and working memory capacity, may be responsible for part of the increased listening effort in older individuals (Grady 2012; Tremblay & Backer 2016).

Older CI users could be at a greater risk for increased listening effort since CIs convey a highly degraded signal. Aging impacts temporal processing abilities such as gap detection (Fitzgibbons and Gordon-Salant 1994) and temporal cues in speech discrimination (Gordon-Salant et al. 2006). Temporal information is increasingly important as speech is vocoded. Poorer performance on sentence recognition in older adults listening to CI simulations is consistent with an explanation of an age-related temporal processing deficit (Goupell et al. 2017). Age-related increases in listening effort occur with increasing stimulus length and complexity (Piquado et al. 2010; Desjardins & Doherty 2012) and degradation from masking noise (Degeest et al. 2015; McGarrigle et al. 2021). These studies suggest that in addition to an aging effect, there may be an interaction between age and signal degradation such that there is consistently high effort with signal degradation in older listeners (Desjardins & Doherty 2012), but the literature is mixed (Piquado et al. 2010; McGarrigle et al. 2021).

Reduced spectral resolution is one way that CI signals are degraded, with CI devices providing a limited number of spectral channels. Depending on the specific electrode array type and frequency range processed, speech recognition often plateaus around eight functional channels (Fishman et al. 1997; Friesen et al. 2001; Berg et al. 2021) or higher (Croghan et al. 2017; Berg et al. 2019), but at a spectral resolution far lower than occurs in normal hearing. Reduced spectral resolution, which can be simulated using vocoded speech, increases listening effort (Winn et al. 2015; Ward et al. 2017; Pals et al. 2020). For instance, Ward et al. (2017) used a dual-task paradigm to measure listening effort in younger and older individuals with age-typical (although not necessarily audiometrically normal; i.e., >25 dB hearing level, HL) hearing sensitivity who were presented vocoded sentences with a varying number of spectral channels. In both age groups, listening effort increased with decreasing spectral resolution. Most notably, based on accuracy and reaction time data for the secondary task, the older listeners exerted greater listening effort than the younger listeners when attending to 4- and 8-channel vocoded speech, demonstrating a significant interaction between age and spectral resolution (Ward et al. 2017). Since threshold differences were not accounted for between the younger and older listeners in that study, it is unclear whether the differences in effort were a result of aging, hearing sensitivity, or both.

In a similar experiment, Winn et al. (2015) assessed effort using pupillometry, a physiological index of listening effort (e.g., Beatty 1982; Winn et al. 2018), with young adult listeners with normal hearing. The size of the pupil reflects the activity of the sympathetic and parasympathetic components of the autonomic nervous system (Rajkowski et al. 1994; Gilzenrat et al. 2010). Unlike behavioral measures, pupillometry is physiologic and time-locked to the stimulus presentation, allowing for measurement of this aspect of effort at different time points of interest (Winn 2023). In Winn et al. (2015), pupil dilation increased as spectral resolution decreased. Additionally, speech recognition was not a significant predictor of pupil dilation. It was concluded that lower spectral resolution leads to increased listening effort, even when speech recognition is high, but pupillometry has not yet been utilized to investigate the potential interaction between age and spectral resolution on listening effort.

In summary, listening effort increases with decreasing spectral resolution and increasing age. However, it is unclear whether spectral resolution and listener age interact because of the mixed results across different methodologies and potential confounds between age and hearing loss. Therefore, the present study aimed to extend the findings of Winn et al. (2015), measuring listening effort as a function of number of vocoder channels with both younger and older listeners. The primary goal was to measure if the effects of varied spectral resolution on listening effort are impacted by age with an attempt to address hearing loss confounds. It was hypothesized that as spectral resolution decreased, older listeners would experience a more rapid increase in listening effort than younger listeners, as well as increased overall effort, based on the previous finding of a main effect of age and an interaction between age and spectral resolution (Ward et al. 2017). With better control on peripheral differences in hearing thresholds across age than in previous studies, an aging effect would suggest a central or cognitive contribution to differences in listening effort with age. Such contributions might depend on working memory capacity (Gordon-Salant & Cole 2016) or processing speed (DeVries et al. 2022), two cognitive processes that are utilized in speech recognition and that may be affected by aging (Tremblay & Backer 2016). As such, the secondary goal was to determine whether individual differences in working memory capacity and processing speed predicted listening effort. It was hypothesized that lower working memory capacity and lower processing speed would be associated with increased listening effort. The results will have implications for older CI listener expectations for listening effort, with the potential to impact clinical practice in terms of setting realistic expectations and potential additional motivation for maximizing spectral resolution in CI listeners.

METHODS

Listeners

Fifteen younger normal-hearing (YNH) individuals aged 18-31 years (mean 21) and 15 older normal-hearing (ONH) individuals aged 65-75 years (mean 70) participated in this study. This sample size was chosen based on previous pupillometry studies investigating effects of spectral resolution (Winn et al. 2015) and age (Piquado et al. 2010). Participants were recruited via an existing participant database, an online research sign-up system (SONA Systems Ltd., 2023), and flyers posted in the surrounding community. Air conduction audiometric thresholds were measured at octave frequencies from 250 to 8000 Hz. All listeners had hearing thresholds at or below 20 dB HL for test frequencies ≤ 4000 Hz in at least one ear with the exception of one ONH listener, who had a threshold of 25 dB HL at 4000 Hz in the tested ear. The older listeners with clinically normal hearing from 250 to 4000 Hz had relatively good hearing for their age. The mean audiometric thresholds for each age group are shown in Fig. 1. Following audiometry, all stimuli were presented monaurally for each individual listener; the right ear was used unless only the left ear met audiometric criteria. Only two listeners, both in the ONH group, were tested in their left ears. A multiple linear regression analysis with factors of age and test frequency and their interactions indicated that the ONH listeners had significantly higher thresholds than the YNH listeners in their test ears from 2000-8000 Hz (p < 0.05).

Fig. 1.

Average audiogram for YNH (open circles) and ONH (filled circles) listeners. Error bars denote ±1 standard error. Error bars smaller than the symbol are not shown. The shaded region represents frequencies beyond the low-pass filter cutoff used for the stimuli in this experiment.

All listeners were screened for possible cognitive impairment using the Montreal Cognitive Assessment (MoCA; Nasreddine et al. 2005). Individuals scored at or above 26 (YNH) or 23 (ONH); while 26 is the minimum score considered to be absent of even mild cognitive impairment, later research has suggested that a score of 23 produces fewer false positives in differentiating healthy aging and true cognitive impairment (Carson et al. 2018). Most listeners had no self-reported prior experience listening to vocoded speech except for two ONH listeners. All listeners were native speakers of English and reported no known speech or language disorders.

All research was conducted with procedures approved by the Institutional Review Board. All listeners provided informed consent.

Stimuli

Speech stimuli consisted of sentences taken from the Harvard IEEE corpus (Rothauser et al. 1969). This corpus contains 72 lists of 10 phonetically balanced sentences which include 5 keywords each. This study used 260 sentences with 80 individual sentences during the first test session plus 18 full lists during the second test session. A low-pass forward-backward filter (third-order Butterworth with −36 dB/octave attenuation) was applied to all unprocessed stimuli with a 4000-Hz cutoff frequency to avoid the high-frequency region where hearing thresholds were not restricted by recruiting criteria. Root-mean-square energy was equalized across the filtered sentences. These stimuli were used for the “unprocessed” signal condition. To create the stimuli for the vocoded conditions, the original stimuli were bandpass forward-backward filtered (third-order Butterworth with −36 dB/octave attenuation) into 8, 16, and 32 channels between 200 and 4000 Hz with equal logarithmic spacing. The envelope of each channel was extracted using the Hilbert transform and low-pass forward-backward filtering with a fourth-order forward-backward Butterworth filter with a 160-Hz cutoff frequency to modulate narrowband noise carriers. Root-mean-square amplitude of the vocoded stimuli was equalized to the unprocessed stimuli.

Procedure

Testing was conducted over the course of two 1.5- to 2-hour sessions. At the first session, eligibility was confirmed, cognitive testing was completed, and practice time was given for listening to vocoded sentences. At the second session, pupillometry testing was completed.

At the start of the first session, to confirm eligibility, pure-tone air conduction thresholds were measured in a double-walled sound-treated booth using an audiometer (Audiostar Pro, Grason-Stadler, Eden Prairie, MN). Next, cognitive screening with the MoCA (Nasreddine et al., 2005) and other cognitive tests were completed. Cognitive testing consisted of the List Sorting Working Memory Test and the Pattern Comparison Processing Speed Test from the National Institutes of Health (NIH) Toolbox (Gershon et al. 2013). Listeners completed the test battery using an iPad (Apple, Cupertino, CA) application. The working memory task involves the user recalling in size order a list of animals and/or foods after audio-visual presentation. The processing speed task requires the participant to determine whether two images presented side by side are the same or different as quickly as possible.

After eligibility confirmation and cognitive testing, participants completed a training block to practice listening to vocoded speech in a sound booth using custom software (E-Prime, Psychology Software Tools, Sharpsburg, PA) and audio output from a Chronos device (Psychology Software Tools, Sharpsburg, PA). The sentences were presented monaurally at 65 dB SPL over circumaural headphones (HD-650, Sennheiser Electronic Corporation, Wedemark, Germany). For the training block, a total of 80 sentences were presented, and listeners were asked to repeat each sentence aloud. Twenty sentences were presented from each signal processing condition [unprocessed (but low-pass filtered), 32-, 16-, and 8-channel vocoded]. The conditions were blocked from highest (unprocessed) to lowest (8-channel vocoded) spectral resolution, and this was repeated four times (five sentences per condition in each block). Listeners were provided with auditory and written feedback following each response, which included hearing the sentence played back unprocessed and then in its processed form again with the sentence displayed on the screen (Davis et al. 2005; Waked et al. 2017). This training block took approximately a half hour to complete.

Participants returned for a second session. Most sessions were completed within approximately 1 week of each other. There were three ONH listeners who were exceptions, who had testing sessions 2-4 weeks apart. This second session consisted of a practice block to introduce eye tracking to the participants and eight test blocks (2 blocks per condition). The practice block was similar in procedure to the training block, with a pupil dynamic range measure and sentence recall task. However, different sentences were used in the practice block for approximately 10 minutes (20 novel sentences, five per condition) without feedback and with eye tracking. In order to facilitate pupil dilation measurement, listeners were seated with their heads against a forehead bar and approximately 25 inches from a 24-inch computer monitor. An eye tracker (SR Research EyeLink 1000 Plus, Ottawa, ON, Canada) was used to record listeners’ pupil size in one eye throughout each trial. Because pupil dilation in response to mental load decreases with age (Van Gerven et al. 2004; Piquado et al. 2010), pupil dilation measurements were scaled relative to each listener’s individual pupil dynamic range, thus accounting for a potential confounding variable brought about by aging (Piquado et al. 2010).

Measurement of each individual’s pupillary dynamic range was conducted prior to each block in the second session (Piquado et al. 2010). Listeners fixed their gaze on a fixation point (“+”) in the center of the computer screen while the screen color progressively changed from black to white in 10-s intervals. The average pupil size measured during presentation of the black screen throughout testing was an estimate of maximum pupil dilation, and the average size measured in response to the white screen was an estimate of minimum pupil dilation. These measures were used to report individual changes in pupil dilation as a percentage of their measured dynamic range.

After the practice block in the second session, eight blocks of testing were completed. Each block consisted of a dynamic range measurement and one novel sentence list (20 sentences) with the same signal processing (unprocessed, 32-, 16-, or 8-channel vocoded). There were two blocks per signal processing condition for a total of 40 sentences per condition. Sentence list order was randomized across listeners. Blocks were ordered such that the first four blocks heard by each listener represented each of the four signal processing conditions in a randomized order; the other four blocks were administered so the conditions were presented in reverse order, thus counterbalancing within participants.

Within a trial, listeners were asked to maintain fixation on the cross in the center of the gray screen while listening and responding. The cross was red during stimulus presentation and turned green to prompt the listener to respond. After each sentence was presented, listeners repeated each sentence aloud, and their responses were scored live by an experimenter for keywords correct. Recordings were later scored offline by two researchers. Scoring rules allowed for changes in verb tense and noun plurality of keywords to be counted as correct. The interrater reliability, the percent where the live scorer and two offline scorers agreed, was 96%, and all conflicting scores were resolved by deliberation between two offline scorers. In situations where a complete audio recording was not available for offline scoring (occurred with two YNH listeners and one ONH listener), the live score was used.

Sentence offsets were time-aligned by the following procedure. A 1-s fixation interval preceded each stimulus presentation window for collection of baseline pupil data. A 4.6-s window was then used to present stimuli such that they were offset-aligned; 4.6 s was the duration of the longest sentence used. Sentences shorter than 4.6 s were preceded by silence long enough to fill in the 4.6-s stimulus window. After the sentence, there was a 1.5-s retention interval for all but the unprocessed test condition due to a programming error; the unprocessed condition had a slightly longer 1.9-s retention interval before the cue to respond. The slightly longer retention interval in the unprocessed condition is unlikely to have influenced the results and was outside the time window used for pupillometry analysis. A 5-s interval followed response collection.

Behavioral Data Analysis

To test whether performance changed as a function of spectral resolution (unprocessed, 32-, 16-, and 8-channel vocoded), age (YNH and ONH), and their interactions, a generalized linear mixed-effects model (GLMM) was used in R 4.2.1 (R Development Core Team, 2022), tested with the “buildmer” 2.8 (Voeten 2023) and “lme4” 1.1-31 (Bates et al. 2015) packages on trial-level data. The dependent variable was a vector of the number of keywords correct per sentence and number of keywords incorrect per sentence. The unprocessed condition and YNH listeners were used as reference levels and model testing was completed using the buildmer default algorithm (backward-elimination) with the binomial family. The binomial family was chosen because keywords were scored with a binomial distribution (correct or incorrect). Random intercepts and slopes were included in the tested model for participants and sentences.

Pupil Data Preprocessing

Pupil data analysis was conducted on successful trials, where the keywords were repeated correctly, to control for effects of accuracy on listening effort. Prior to this analysis, preprocessing of trials was completed as recommended for measures of listening effort (Winn et al. 2018) and similarly to previous studies (e.g. Piquado et al. 2010; Zekveld et al. 2010, 2011; Kuchinsky et al. 2013; Kuchinsky et al. 2014; DeRoy Milvae et al. 2021) for visualization of the pupil data and for the dynamic range analysis. Preprocessing included a 10.6-s interval starting with the baseline per trial.

Preprocessing included removing blinks, saccades, and looks to the outer third of the screen and replacing these points with linear interpolations. Additionally, the percentage of data removed was examined to determine if a trial should be included in the visualization; if over 33% of the baseline period or over 33% of the trial was interpolated, the trial was removed from the data set. Trials were then smoothed with a 5-point moving window and downsampled to 100 Hz. The pupil size was then scaled from arbitrary units (AU) to percentage of individual dynamic range (%DR) with the following equation (Piquado et al. 2010):

$$\%DR=\frac{AU-Min}{Max-Min}\ast 100$$ (1)

In this equation, “Max” refers to the average pupil dilation when viewing a black screen, and “Min” refers to the average pupil dilation when viewing a white screen. The “Min” and “Max” values were averaged over the 8 test blocks. These minima and maxima are not necessarily physiological minima and maxima, given that the duration of each screen was 10 seconds, but they are estimates of minimum and maximum pupil dilation on a similar timescale to the trial duration. Further, the data were baseline corrected such that the average pupil dilation in the first second, during the baseline period, was subtracted from each time point such that the average pupil dilation during the baseline was zero. Data preprocessed in this way were used for visualization.

Preprocessing for the statistical analysis was the same as for data visualization with the exceptions of including points where gaze was located anywhere on the screen rather than in the middle third of the screen. This was done because the analysis included a model of the effect of gaze position on pupil size, so we could be less restrictive of gaze position for the analysis. Further exceptions were that blinks, saccades, and off-screen looks were removed rather than interpolated and smoothed, and trials were downsampled to 50 Hz (van Rij 2019).

Pupil Data Analysis

Generalized additive mixed models (GAMMs) were used to analyze the pupil data, using the “mgcv” package (Wood 2011) in R. GAMMs are a nonlinear hierarchical regression approach that has been used for pupillometry (van Rij 2019), with the advantage of modeling the autocorrelation in the time series data, modeling changes in pupil dilation with gaze position, and the ability to test when in time significant differences are occurring. The time window for the analysis was limited to −2 to +0.5 s in relation to sentence offset, as restricted in a previous study with similar stimuli and research questions (Winn et al. 2015). An ordered factor modeling approach was used (Sóskuthy 2017; Wieling 2018), such that categorical variables of spectral resolution and age were represented as ordered factors. With this approach, parametric differences in curve height as well as nonlinear differences in curve shape (smooth terms) could be compared to a reference condition (unprocessed spectral resolution of YNH listeners).

Cognitive Analysis

To examine whether the age groups differed in cognitive abilities, linear regression analyses were completed for working memory and processing speed scores. Each analysis had a dependent variable of cognitive standard score (not age corrected) and a factor of age (YNH or ONH).

It was also of interest to test whether cognitive scores predicted pupil dilation. To examine this, working memory and processing speed scores were added to the pupillometry model one at a time as an interaction term with time (a tensor product smooth term).

RESULTS

Sentence Recognition

The mean percent keywords correct for each age group and each signal processing condition, during training and test sessions, are shown in Table 1. Overall, both YNH and ONH listeners’ performances were near ceiling in all conditions, >90% correct on average regardless of spectral resolution. There was a significant interaction between age and spectral resolution (see Table 2). To better understand the interaction, the model was re-referenced. With reference levels of YNH and V32, there was a significant interaction between age and the V16 condition (p = 0.04) and between age and the V8 spectral resolution (p = 0.009). In other words, there was an interaction such that there was a greater decrease in performance from 32 to 16 channels and from 32 to 8 channels in the YNH listeners compared to the ONH listeners.

TABLE 1.

Mean percent keywords correct across each signal processing condition for YNH and ONH listeners during training blocks completed in the first session and testing blocks completed in the second session. Standard error is reported in parentheses.

Training	U	V32	V16	V8
YNH	98.5 (0.3)	98.2 (0.4)	98.3 (0.5)	95.3 (0.5)
ONH	98.3 (0.6)	97.7 (0.5)	97.1 (0.7)	95.1 (0.9)
Testing
YNH	99.0 (0.2)	98.0 (0.3)	97.4 (0.4)	93.6 (0.8)
ONH	98.1 (0.6)	96.1 (0.7)	96.5 (0.9)	92.1 (1.2)

TABLE 2.

Generalized linear mixed-effects model summary for the behavioral speech recognition scores, showing the effects of spectral resolution (SR) and age on performance (significant effects are shown in bold). Factors were coded such that comparisons are relative to the reference level (for example, the significant effect of age is observed in the unprocessed condition).

Fixed Effects	Estimate	SE	z	p
Intercept (YNH, U)	4.96	0.20	25.19	< 0.001
SR (V32 > U)	−0.63	0.22	−2.83	0.005
SR (V16 > U)	−0.95	0.21	−4.50	< 0.001
SR (V8 > U)	−1.92	0.19	−9.93	< 0.001
Age (ONH > YNH)	−0.61	0.22	−2.75	0.006
SR (V32) × Age (ONH)	−0.13	0.27	−0.49	0.62
SR (V16) × Age (ONH)	0.31	0.27	1.16	0.25
SR (V8) × Age (ONH)	0.36	0.24	1.48	0.14
Random Effects	Variance	SD
By-sentence intercepts	0.91	0.96

Pupillometry

Based on overall high performance for all listeners in all conditions, and in order to isolate listening effort from accuracy, analysis of pupillometric data was limited to correct trials only. Mean pupil dilation as a percentage of individual dynamic range is plotted to visually compare spectral resolution (Fig. 2) and age (Fig. 3). Fig. 2A shows pupil dilation in arbitrary units before conversion to units of dynamic range or baseline correction so that the impact of these transformations on the data are apparent; Fig. 2B shows the data with these transformations, which are replotted to compare age in Fig. 3. Statistical analysis of the time window from −2 to +0.5 s from stimulus offset (Table 3) showed that relative to the unprocessed condition with YNH listeners, ONH listeners had larger pupil dilation in the 8-channel (p = 0.04) and 32-channel (p = 0.046) vocoded conditions, and nearly significantly increased pupil dilation in the 16-channel vocoded condition (p = 0.08). Additionally, the curve shape was significantly different for the ONH listeners in the V8 condition (p = 0.03), with a sharper peak in pupil dilation. In general, there was an interaction between spectral resolution and age such that the ONH listeners had increased pupil dilation, especially in the most adverse spectral resolution.

Fig. 2.

Mean pupil dilation over time for the four signal processing conditions with age, in arbitrary units before percent dynamic range and baseline correction transformations (A) and after these transformations (B). Only correct trials are included. The gray shaded region represents the analysis window. The white line indicates the sentence offset.

Fig. 3.

Mean pupil dilation over time for YNH and ONH listeners in each signal processing condition. Only correct trials are included. Error shading denotes ±1 standard error. The gray line indicates the sentence offset.

TABLE 3.

Generalized additive mixed model summary for the pupillometry test data recalled accurately, showing the effects of spectral resolution (SR) and age on pupil dilation (significant effects are shown in bold; edf = estimated degrees of freedom; Ref.df = reference degrees of freedom).

Parametric Terms (Curve Height)	Estimate	SE	t	p
Intercept (YNH, U)	−0.51	1.11	−0.46	0.64
SR (V32)	−0.06	0.96	−0.07	0.95
SR (V16)	0.65	0.72	0.91	0.36
SR (V8)	1.68	1.14	1.48	0.14
Age (ONH)	−1.67	1.54	−1.09	0.28
SR (V32) × Age (ONH)	2.74	1.37	2.00	0.046
SR (V16) × Age (ONH)	1.80	1.04	1.73	0.08
SR (V8) × Age (ONH)	3.40	1.63	2.09	0.04
Smooth Terms (Curve Shape)	edf	Ref.df	F	p
Time	3.70	4.07	0.65	0.71
Time × SR (V32)	2.84	3.32	0.20	0.86
Time × SR (V16)	1.84	2.03	0.44	0.72
Time × SR (V8)	1.02	1.02	0.91	0.34
Time × Age (ONH)	3.74	4.20	0.61	0.74
Time × SR (V32) × Age (ONH)	4.37	5.16	1.18	0.32
Time × SR (V16) × Age (ONH)	3.81	4.39	0.31	0.89
Time × SR (V8) × Age (ONH)	6.51	7.32	2.33	0.03
Pupil Position (X, Y)	293.19	298.67	103.68	< 0.001
Random Effect of Time by Subject	199.03	300.00	5.11	< 0.001
Random Effect of Time by Subject by V32	142.03	298.00	1.62	< 0.001
Random Effect of Time by Subject by V16	160.10	298.00	1.66	< 0.001
Random Effect of Time by Subject by V8	156.53	298.00	1.88	< 0.001
Random Effect of Time by Sentence	882.67	1600.00	2.24	< 0.001

Because hearing thresholds were higher for the older listeners, which has been shown to affect vocoded speech recognition (e.g., Turner et al. 1995; Souza & Boike 2006; Shader et al. 2020), high-frequency pure-tone average (HFPTA; mean threshold at 1000, 2000, and 4000 Hz) was also included (HFPTA interaction with time) in this analysis in order to estimate the extent to which differences in peripheral hearing status may have contributed to differences in pupil dilation. However, HFPTA was not predictive of pupil dilation over time (p = 0.23), suggesting that small differences in hearing thresholds across groups were not a confounding variable.

We also examined the effect of our transformations on these data. With a percent dynamic range transformation, we assumed that task-evoked effects on pupil size are proportional to the pupil dynamic range as in Piquado et al. (2010). The mean dynamic range measured for the YNH listeners was 1603 arbitrary units (SE = 122) in comparison to 818 arbitrary units (SE = 68) for the ONH listeners. This physiological reduction in pupil reactivity in older adults has been observed previously (e.g., Piquado et al 2010; Zhao et al. 2019). Pupil size was scaled within each individual’s measured dynamic range in an effort to compare younger and older listeners (Piquado et al 2010). With this transformation, baseline pupil dilation observed was on average 32 percent dynamic range for YNH listeners (SE = 2.93) and 37 percent dynamic range for ONH listeners (SE = 3.49). A linear mixed-effects model was used to examine if there were significant differences in baseline (in units of percent dynamic range prior to baseline correction) with age, spectral resolution, and their interaction included as fixed effects and by-subject intercepts included as random effects. Trial-level baseline averages were used as the dependent variable. Model testing with “buildmer” (Voeten 2023) revealed no significant effect of age, spectral resolution, or their interaction, with all fixed effects removed from the model during model testing (LRT > 0.05).

Cognitive Factors

Mean working memory and processing speed standard scores are shown in Table 4. The analyses showed significant effects of age for working memory (p < 0.001) and processing speed (p < 0.001). Working memory and processing speed scores were added to the pupillometry model (described in Table 3) as an interaction with time, one at a time, to test whether these cognitive factors predicted pupil dilation. Working memory (p = 0.39) and processing speed (p = 1.00) did not predict pupil dilation during this speech recognition task.

TABLE 4.

Mean working memory and processing speed standard scores for each age group. Standard error is reported in parentheses.

Cognitive Score	Working Memory	Processing Speed
YNH	112.80 (2.17)	137.27 (4.97)
ONH	98.20 (2.44)	100.47 (3.13)

DISCUSSION

This study used pupillometry to evaluate whether aging and reduced spectral resolution interact to cause increased listening effort on a sentence recall task. In an attempt to mitigate the effects of hearing threshold differences across age, strict hearing threshold criteria were used (Fig. 1) and stimuli were restricted to frequency regions where thresholds were clinically normal. In addition, to control for the impact of differences in performance across age and condition, pupillometry was analyzed for sentences recalled correctly.

It was hypothesized that older listeners would experience greater and more rapidly increasing listening effort as spectral resolution decreased. This hypothesis was supported by an interaction between age and spectral resolution, where older listeners demonstrated significantly increased pupil dilation in the 8- and 32-channel vocoded conditions (Fig. 2 and 3, Table 3), and a greater peak in the response in the 8-channel condition (Table 3).

Sentence Recognition

Speech recognition in normal-hearing listeners using 8 or more channels in a vocoder is typically excellent (e.g., Ward et al. 2017; Winn et al. 2015). Direct comparisons across studies are difficult because these studies had small but perhaps consequential differences in the vocoder implementation, notably including the frequency range. However, similar patterns of results were observed for the listeners in this study, who demonstrated near-ceiling performance across all four levels of spectral resolution in both the training and testing sessions (Table 1). Statistical analysis revealed an interaction between age and spectral resolution (Table 2). The largest average difference between YNH and ONH listeners at the same spectral resolution level was 1.9%; between spectral resolution levels within the same age group, the largest difference was 6.0%. These values reflect poorer performance in the ONH listeners and with poorer spectral resolution, respectively. These are relatively small functional differences in performance, but in the direction expected based on previous studies. Older adults have shown poorer vocoded sentence recognition (Shader et al. 2020) and poorer speech understanding with CIs (Sladen & Zappler 2015). Poorer performance on sentence recognition in older adults listening to CI simulations is consistent with an explanation of an age-related temporal processing deficit (Goupell et al. 2017), since temporal information is the primary cue for recognition of vocoded speech. Interestingly, the YNH listeners showed greater reductions in performance with poorer spectral resolution than the ONH listeners (Table 2), but the absolute differences are so small that while the interaction was statistically significant, it may not be functionally meaningful.

Pupillometry

For sentences recalled accurately, ONH listeners demonstrated relatively greater pupil dilation for speech recognition when there were lower levels of spectral resolution than YNH listeners; there was an interaction between age and spectral resolution (Table 3). The 8- and 32-channel vocoded conditions led to higher pupil dilation in the ONH listeners relative to the unprocessed condition in the YNH listeners; the increase in the 16-channel vocoded condition in the ONH listeners did not reach statistical significance. These findings suggest that there may be a need to recruit greater mental resources to recognize degraded speech (e.g., when listening to speech through a CI) and this may be exacerbated for older individuals.

In addition, the growth of pupil dilation over time in response to the sentence was higher for older listeners in the 8-channel condition with the same reference, indicating a more rapid increase in effort in response to the most degraded spectral resolution (Table 3). This suggests that older individuals could be recruiting greater mental resources in a shorter amount of time than younger individuals during vocoded speech perception with poor spectral resolution. Increased effort in the ONH listeners could arise from poorer cortical representation of the degraded auditory signal despite hearing thresholds within the normal range, leading to an increase in engagement of mental resources to understand the degraded sentence. The consequence of this increase in effort during sentence processing could be that older adults with CIs sacrifice resources to processing that could otherwise be devoted to processes such as memory and processing a running speech stream, decreasing communication quality and leading to increased fatigue. Further research is needed to explore these potential consequences.

Cognition

Working memory and processing speed scores did not have a significant impact on listening effort exerted in this study as measured by pupillometry, contrary to the hypothesis that declines in these cognitive processes would contribute to increased listening effort in older listeners (Tremblay & Backer, 2016). Previous studies that have explored the relationship between listening effort and cognition show mixed results; some studies showed no significant relationship (DeRoy Milvae et al. 2021; Perreau et al. 2017) but others did observe a relationship (Desjardins & Doherty 2012; Ward et al. 2017). Given that different indices of listening effort may assess different dimensions of effortful listening (Alhanbali et al. 2019), it is possible that another measure of listening effort might better correlate with cognitive abilities. Additionally, performance on this task was near ceiling. This high performance may have impacted our ability to observe a relationship between cognitive abilities and listening effort due to lower task demands; it is possible that a more challenging task would lead to greater individual differences in effort that might better relate to individual differences in cognitive ability.

Clinical Implications

If degraded spectral resolution affects older adults more adversely than younger adults, requiring greater listening effort, this could be important to share with CI candidates when setting realistic expectations. Additional strategies that could be used in aural rehabilitation would be to recommend limiting time in difficult listening environments and to monitor fatigue and effortful listening as rehabilitation outcomes. Maximizing spectral resolution is also important for all CI listeners; however, since adults experience age-related declines in temporal processing, it may be particularly important for them to receive signals with optimal spectral resolution to minimize the listening effort required to understand degraded speech. Additional research is needed to confirm these potential strategies.

Limitations and Future Directions

The primary aim of this study was to compare listening effort across age while minimizing differences in hearing thresholds between younger and older listeners to reduce the hearing loss confound. While all listeners had normal hearing throughout the range of frequencies included in test stimuli, a statistically significant difference remained between the YNH and ONH thresholds (Fig. 1). Effects of peripheral hearing status were not completely eliminated despite stringent recruitment criteria and low-pass filtered stimuli at 4 kHz. There remains the possibility that part of the aging effects observed on listening effort are related to the YNH listeners having better hearing sensitivity than their ONH counterparts. Because of this, a role of differences in peripheral function across age cannot be entirely ruled out (Lough & Plack 2022). Ideally, this study would be run with hearing threshold-matched listeners, although even this strategy would not account for changes in the peripheral auditory system with age that are not reflected in the audiogram. Testing CI users themselves would also partially address this confound. The groups were also not matched on factors such as education level or other demographic factors. Additionally, with the high performance on this task, it remains to be seen if the same patterns in pupil dilation would be observed with greater overall task demands, such as with the introduction of noise, where it is possible that aging effects could be more pronounced. Although significant effects were observed in this study, replication with a larger sample size is an important step to confirm those effects. The role of these factors on the results could be explored in future studies.

Another limitation of this study lies in its use of normal-hearing listeners who were presented vocoded speech as a proxy for CI listeners. Research that is not conducted on CI users is no more than an acute simulation of sound degradation and does not represent the true experience of listening through a CI. Adjusting the number of channels delivered by a CI would be most comparable to this experiment but has drawbacks such as adaptation issues with frequency allocation (Fu & Galvin 2003; Pals et al. 2020; Rosen et al. 1999). An alternative approach would be to examine changes in effort in actual CI listeners across other types of signal degradation, such as comparing speech understanding in quiet and in noise.

CONCLUSIONS

1. When sentences were recalled accurately, there was a significant interaction between age and spectral resolution such that older adults with normal hearing thresholds up to 4 kHz exhibited greater listening effort as indexed by pupil dilation in two of three spectral resolution levels compared to the unprocessed condition in the younger adults with normal hearing.

2. Older adults appear to be at risk for increased listening effort with highly degraded speech, which has the potential to impact the CI listening experience.