Prefrontal cortex supports speech perception in listeners with cochlear implants

  1. Arefeh Sherafati
  2. Noel Dwyer
  3. Aahana Bajracharya
  4. Mahlega Samira Hassanpour
  5. Adam T Eggebrecht
  6. Jill B Firszt
  7. Joseph P Culver
  8. Jonathan E Peelle  Is a corresponding author
  1. Department of Radiology, Washington University in St. Louis, United States
  2. Department of Otolaryngology, Washington University in St. Louis, United States
  3. Moran Eye Center, University of Utah, United States
  4. Department of Electrical & Systems Engineering, Washington University in St. Louis, United States
  5. Department of Biomedical Engineering, Washington University in St. Louis, United States
  6. Division of Biology and Biomedical Sciences, Washington University in St. Louis, United States
  7. Department of Physics, Washington University in St. Louis, United States

Abstract

Cochlear implants are neuroprosthetic devices that can restore hearing in people with severe to profound hearing loss by electrically stimulating the auditory nerve. Because of physical limitations on the precision of this stimulation, the acoustic information delivered by a cochlear implant does not convey the same level of acoustic detail as that conveyed by normal hearing. As a result, speech understanding in listeners with cochlear implants is typically poorer and more effortful than in listeners with normal hearing. The brain networks supporting speech understanding in listeners with cochlear implants are not well understood, partly due to difficulties obtaining functional neuroimaging data in this population. In the current study, we assessed the brain regions supporting spoken word understanding in adult listeners with right unilateral cochlear implants (n=20) and matched controls (n=18) using high-density diffuse optical tomography (HD-DOT), a quiet and non-invasive imaging modality with spatial resolution comparable to that of functional MRI. We found that while listening to spoken words in quiet, listeners with cochlear implants showed greater activity in the left prefrontal cortex than listeners with normal hearing, specifically in a region engaged in a separate spatial working memory task. These results suggest that listeners with cochlear implants require greater cognitive processing during speech understanding than listeners with normal hearing, supported by compensatory recruitment of the left prefrontal cortex.

Editor's evaluation

The work establishes use of a specific extra area of prefrontal cortex during word listening by CI users and supports a hypothesis based on the multiple demand network that can be tested using other techniques that look at the rest of the network. The revision provides further points of clarity required and better acknowledges the limitations of the technique.

https://doi.org/10.7554/eLife.75323.sa0

Introduction

Cochlear implants (CIs) are neuroprosthetic devices that can restore hearing in people with severe to profound hearing loss by electrically stimulating the auditory nerve. Because of physical limitations on the precision of this stimulation—including, for example, the spatial spread of electrical current (Garcia et al., 2021)—the auditory stimulation delivered by a CI does not convey the same level of acoustic detail as normal hearing. As a result, speech understanding in listeners with CIs is poorer than in listeners with normal hearing (Firszt et al., 2004). Notably, even in quiet, listeners with CIs report increased effort during listening (Dwyer and Firszt, 2014). Despite these challenges, many listeners with CIs attain significant success in understanding auditory speech. This remarkable success raises the question of how listeners with CIs make sense of a degraded acoustic signal.

One area of key importance is understanding the degree to which listeners with CIs rely on non-linguistic cognitive mechanisms to compensate for a degraded acoustic signal. In listeners with normal hearing, cognitive demands increase when speech is acoustically challenging (Peelle, 2018). For example, even when speech is completely intelligible, acoustically degraded speech can reduce later memory for what has been heard (Cousins et al., 2013; Koeritzer et al., 2018; Rabbitt, 1968; Ward et al., 2016). These findings suggest that to understand acoustically challenging speech, listeners need to engage domain-general cognitive resources during perception. In a limited capacity cognitive system (Wingfield, 2016), such recruitment necessarily reduces the resources available for other tasks, including memory encoding. Importantly, even speech presented in quiet (i.e., without background noise) is degraded by the time it reaches the auditory system of a listener with a CI.

Cognitive demands during speech understanding are supported by several brain networks that supplement classic frontotemporal language regions. The cingulo-opercular network, for example, is engaged during particularly challenging speech (Eckert et al., 2009; Vaden et al., 2017) and supports successful comprehension during difficult listening (Vaden et al., 2013). Recruitment of prefrontal cortex (PFC) complements that in the cingulo-opercular network and varies parametrically with speech intelligibility (Davis and Johnsrude, 2003). Activity in PFC, particularly dorsolateral regions, is associated with cognitive demands in a wide range of tasks (Duncan, 2010), consistent with domain-general cognitive control (Braver, 2012). We thus hypothesized that listeners with CIs would rely more on PFC during listening than listeners with normal hearing, and in particular regions of PFC associated with non-linguistic tasks. However, the functional anatomy of PFC is also complex (Noyce et al., 2017), and dissociating nearby language and domain-general processing regions is challenging (Fedorenko et al., 2012).

A central question concerns the degree to which listeners with CIs rely on cognitive processing outside core speech regions, such as dorsolateral PFC. Obtaining precise spatially localized images of regional brain activity has been difficult in listeners with CIs, given that functional MRI (fMRI) is not possible (or subject to artifact) due to the CI hardware. Thus, optical brain imaging (Peelle, 2017) has become a method of choice for studying functional activity in CI listeners (Anderson et al., 2017; Lawler et al., 2015; Lawrence et al., 2018; Olds et al., 2016; Zhou et al., 2018). In the current study, we use high-density diffuse optical tomography (HD-DOT) (Eggebrecht et al., 2014; Zeff et al., 2007), previously validated in speech studies in listeners with normal hearing (Hassanpour et al., 2015; Hassanpour et al., 2017; Schroeder et al., 2020). HD-DOT provides high spatial resolution and homogenous sensitivity over a field of view that captures known speech-related brain regions (White and Culver, 2010). We examine the brain regions supporting single word processing in listeners with a right unilateral CI relative to that in a group of matched, normal-hearing controls. We hypothesized that listeners with CIs would exhibit greater recruitment of PFC compared to normal-hearing controls.

Results

Multi-session single-subject results

Due to the variability across CI users and difficulties in defining single-subject regions of interest (ROIs), we performed a small multi-session study from one CI subject for six sessions (Figure 1). We collected two runs of spoken word perception per session (for six sessions) and one run of spatial working memory task per session (for four sessions). This multi-session analysis enabled localizing the left and right PFC based on the non-verbal spatial working memory task for this subject (Figure 1C). It also revealed the engagement of regions beyond the auditory cortex, including PFC, during the word perception task (Figure 1B). Time-traces of oxyhemoglobin concentration change show a clear event-related response for four selected regions in the word perception results.

Single-subject data from one cochlear implant (CI) user across multiple sessions.

(A) A CI user wearing the high-density diffuse optical tomography (HD-DOT) cap. (B) Response to the spoken words across six sessions (36 min of data). Hemodynamic response time-traces are plotted for peak activation values across six sessions for four brain regions. The seed colors match the plot boundaries with error bars indicating the standard error of the mean over n=12 runs of data. Gray shaded region indicates period during which words are presented. (C) Response to the spatial working memory task for the same CI user across four sessions (32 min of data).

Mapping the brain response to spoken words

We first investigated the degree of auditory activation in both control and CI groups by assessing the activity in a block design single word presentation task. We found strong bilateral superior temporal gyrus (STG) activations in controls similar to our previous studies using the same paradigm (Eggebrecht et al., 2014; Sherafati et al., 2020), as well as a strong left STG and a reduced right STG activation for the CI users (Figure 2A–B). In addition, we observed strong left-lateralized activations in regions beyond the auditory cortex, including parts of the PFC, in the CI group (Figure 2B).

Figure 2 with 1 supplement see all
Spoken word recognition group maps.

Response to the spoken words in (A) 18 controls and (B) 20 right ear cochlear implant (CI) users. (C) Differential activation in response to the spoken words task in CIs>controls highlights the group differences. The first column shows unthresholded β maps and the second column shows t-statistic maps thresholded at voxelwise p<0.05 (uncorrected) for each group.

Figure 2—figure supplement 1 provides the β maps of oxyhemoglobin (HbO), deoxyhemoglobin (HbR), and total hemoglobin (HbT) for controls (panel A), CI users (panel B), CIs>controls (panel C), and controls>CIs (panel D).

For statistical analysis, we focused on our predefined ROIs, as shown in Figure 6. Figure 3A-C shows unthresholded t-maps, masked by our ROIs. We averaged β values within each ROI, and statistically tested for group differences correcting for multiple comparisons across the three ROIs (Figure 3D). The temporal profile of the hemodynamic response in three selected ROIs also suggests the increased activity in the left PFC in the CI users relative to controls, and a decreased activity in both left and right auditory regions. Two sample t-statistics for the mean β values in each ROI support a statistically significant difference between the control and CI groups in left PFC, t(27) = 2.3, p=0.015 (one-tailed) and right auditory cortex, t(23) = 3.54, p=0.0017 (two-tailed) (Figure 3D). The observed change in the left auditory cortex was not statistically significant, t(36) = 1.46, p=0.15 (two-tailed). The threshold for statistical significance, Bonferroni corrected for multiple comparisons across three ROI analyses, was 0.016.

Figure 3 with 3 supplements see all
Region of interest (ROI)-based statistical analysis for spoken word recognition task.

Unthresholded t-maps in response to the spoken words spatially masked in the three ROIs for (A) 18 controls, (B) 20 right ear cochlear implant (CI) users, and (C) CIs>controls, highlight the group differences in certain brain areas. (D) Temporal profile of the hemodynamic response in three selected ROIs (left prefrontal cortex [PFC], left auditory, and right auditory cortices). The error bars indicate the standard error of the mean over n=20 for CI users and n=18 for controls. Two-sample t-tests for mean β value in each ROI have been calculated between controls and the CI user group, confirming a significant increase in left PFC (p=0.015) and a significant decrease in the right auditory cortex (p=0.0017) in CI users (indicated by an asterisk above their corresponding box plots, corrected for multiple comparisons). The observed change in the left auditory cortex was not significant (p=0.15).

Figure 3—source data 1

Noisy source-detector numbers are provided along with the threshold used for identifying them for each cochlear implant (CI) user.

https://cdn.elifesciences.org/articles/75323/elife-75323-fig3-data1-v2.docx

To further investigate whether the significant decrease in the right auditory cortex might be due to lower light level values around the CI transducer, we first performed a simulation of the HD-DOT sensitivity profile by blocking the optodes around the CI transducer. We found minimal overlap between the CI-related signal loss and the right auditory ROI (Appendix 2). We also performed an additional simulation analysis by first calculating the mean value of optical light power for each measurement. Then, for each source-detector location, we defined the approximate coupling coefficient as the mean over the first nearest neighbor measurements that have that source or detector in common. Figure 3—figure supplements 12 show the coupling coefficients of each source and detector for all CI users and controls in the flat view, corrected by the average sensitivity of the avalanche photodiode detectors (APDs) plotted on a log base 10 scale to create linearly distributed color scale. We defined noisy optodes in the CI cohort as ones that had a coupling coefficient of 30% of maximum or lower based on the values for each participant across their two speech runs. If no optode was identified using this threshold, we lowered the threshold by 10% increments until at least one source or detector was identified (under the logic that the CI transducer must block some light). Figure 3—source data 1 shows the source-detector numbers and the threshold used for identifying them for each CI user. For controls, no source or detector had a coupling coefficient of less than 30% of the maximum.

After identifying the noisy source-detector numbers for all 20 CI users, we pre-processed the data for 18 controls 100 times. For each of these analyses, we blocked the identified sources and detectors in a selected CI user for one control, assigned by a random permutation, effectively simulating the CI-based dropout in our control participants. Then, we replicated the ROI analysis for the right auditory cortex in Figure 3, far right panel, and found that in all 100 shuffles, the right auditory cortex in control subjects still had a significantly larger activation compared to the CI users (Figure 3—figure supplement 3).

Behavioral measures

An important consideration in studying CI users is the variability in their speech perception abilities, hearing thresholds, and the relationship with brain activity. Figure 4 shows exploratory analyses between the magnitude of the activation in the left PFC ROI for the CI cohort with respect to the speech perception score, left ear hearing threshold unaided, left ear hearing threshold at test (aided if the subject used a hearing aid), and right ear hearing threshold (CI-aided).

Figure 4 with 1 supplement see all
Relationship between the magnitude of activation in left prefrontal cortex (PFC) and behavioral scores in cochlear implant (CI) users.

Plots of the Pearson correlation r between the magnitude of the mean β value in the left PFC region of interest (ROI) are shown with respect to speech perception score, left ear hearing threshold unaided, left ear hearing threshold (aided if the subject used a hearing aid), and right ear CI-aided hearing threshold. Hearing threshold was defined as four-frequency pure-tone average (4fPTA) at four frequencies, 500, 1000, 2000, and 4000 Hz.

Using p<0.05 (uncorrected) as a statistical threshold, left PFC activation positively correlated with left ear unaided thresholds (p=0.01, Pearson r=0.55) and negatively correlated with right ear CI-aided thresholds (p=0.02, Pearson r=–0.49). Left PFC activation did not correlate with speech perception score (p=0.4, Pearson r=0.17) and aided hearing threshold for the left ear (p=0.1, Pearson r=0.36). Figure 4—figure supplement 1 shows the hearing thresholds (audiograms) for left and right ears for each control and CI participant. These scores were unavailable for one control participant, who was thus not included in these analyses.

Discussion

Using high-density optical brain imaging, we examined the brain networks supporting spoken word recognition in listeners with CIs relative to a matched group of controls with bilateral normal hearing. We found that relative to controls, when listening to words in quiet, listeners with CIs showed reduced activity in the right auditory cortex and—critically—increased activity in left PFC. We review these two findings in turn below.

Increased PFC activity in CI users

When listening to spoken words in quiet, listeners with normal hearing typically engage the left and right superior temporal cortex, including primary and secondary auditory regions (Binder et al., 2000; Price et al., 1992; Rogers et al., 2020; Wiggins et al., 2016). Our current results for controls show this same pattern. However, when listeners with CIs performed the same task, we found that they also engaged left PFC significantly more than the controls.

Although we only tested a single level of speech difficulty (i.e., speech in quiet), prior studies have parametrically varied speech intelligibility and found intelligibility-dependent responses in the PFC. Use of several types of signal degradation (Davis and Johnsrude, 2003) revealed a classic ‘inverted-U’ shape response in the PFC as a function of speech intelligibility, with activity increasing until the speech became very challenging and then tapering off. A similar pattern was reported in functional near-infrared spectroscopy (fNIRS) (Lawrence et al., 2018).

A pervasive challenge for understanding the role of PFC in speech understanding is the close anatomical relationship of core language processing regions and domain-general regions of PFC (Fedorenko et al., 2012). We attempted to add some degree of functional specificity to our interpretation by including a spatial working memory task presumed to strongly engage domain-general regions with minimal reliance on language processing (Duncan, 2010; Woolgar et al., 2015). Ideally, we would have used functional ROIs individually created for each subject. However, we were not convinced that our data for this task were sufficiently reliable at the single-subject level, as we only had one run per subject. Furthermore, we did not have spatial working memory task data for all subjects. Thus, our functional localization relies on group-average spatial working memory responses. The region we identified—centered in left inferior frontal sulcus—corresponds well to other investigations of non-language tasks (e.g., Duncan, 2010), and supports our preferred interpretation of engagement of domain-general regions of dorsolateral PFC during listening. However, the ROI also extends into the dorsal portion of inferior frontal gyrus (IFG), and we cannot rule out the possibility that this frontal activity relates to increased language (as opposed to domain-general) processing.

Reduced auditory cortical activity in CI users

We found reduced activity in the right auditory cortex in CI users relative to controls, which we attribute to differences in auditory stimulation. We limited our sample to CI listeners with unilateral right-sided implants but did not restrict left ear hearing. Most of our subjects with CIs had poor hearing in their left ears, which would result in reduced auditory information being passed to the contralateral (right) auditory cortex. This was as opposed to controls who had bilateral hearing. Prior fNIRS studies have also shown that activity in the superior temporal cortex corresponds with stimulation and comprehension (Olds et al., 2016; Zhou et al., 2018). We also leave open the possibility that the CI hardware in the right hemisphere interfered with the signal strength, although our simulation studies suggest this cannot fully explain the group differences (Appendix 2 and Figure 3—figure supplements 13).

What is potentially more interesting is the hint at a lower level of activity in the left auditory cortex of the CI users compared to controls, even though all CI listeners were receiving adequate stimulation of their right auditory nerve with a right CI. There are several possible explanations for this finding. First, activity in superior temporal cortex does not reflect only ‘basic’ auditory stimulation, but processing related to speech sounds, word meaning, and other levels of linguistic analysis. Thus, although subjects with CIs were certainly receiving stimulation and speech intelligibility scores were generally good, some variability was still present (mean speech perception score in quiet = 0.88, SD = 0.09). The overall level of speech processing was significantly (p=0.00005) lower for CI users than controls (mean speech perception score = 0.99, SD = 0.01), resulting in decreased activity (indeed, because the depth of HD-DOT includes only about 1 cm of the brain, much of primary auditory cortex is not present in our field of view, and the observed group differences were localized in non-primary regions of STG and MTG).

Perhaps the most provocative explanation is that a reduction in top-down modulatory processes (Davis and Johnsrude, 2007) plays out as reduced activity in the temporal cortex. That is, given that effortful listening depends on attention (Wild et al., 2012), it might be that processes related to top-down prediction (Cope et al., 2017; Sohoglu et al., 2012; Sohoglu et al., 2014) are muted when too much cognitive control is required for perceptual analysis. Reconciling this interpretation with predictive coding accounts of speech perception (Blank and Davis, 2016; Sohoglu and Davis, 2020) will require additional work. We emphasize that the group differences in left auditory regions were not significant, and thus our interpretation is speculative.

Individual differences in PFC activation during spoken word recognition

Because of the variability of outcomes in CI users (Firszt et al., 2004; Holden et al., 2013), one promising thought is that individual differences in brain activation may help explain variability in speech perception ability. Although our study was not powered for individual difference analysis (Yarkoni and Braver, 2010), we conducted exploratory correlations to investigate this avenue of inquiry. Interestingly, we saw a trend such that poorer hearing in the left (non-CI) ear was correlated with increased activity in PFC. Our subjects with CIs had significant variability in left ear hearing. Because the speech task was conducted using loudspeakers, we would expect both ears to contribute to accurate perception. Thus, poorer hearing in the left ear would create a greater acoustic challenge, with a correspondingly greater drain on cognitive resources. This interpretation will need additional data to be properly tested.

Comparison with previous fNIRS studies in CI users

Due to limitations of using fMRI and EEG in studying CI users, fNIRS-based neuroimaging is an attractive method for studying the neural correlates of speech perception in this population (Hassanpour et al., 2015; Lawler et al., 2015; Saliba et al., 2016; Sevy et al., 2010; Zhou et al., 2018). Some prior studies have looked at relationships of behavioral performance and brain activity in CI users. For example, Olds et al., 2016, studied temporal lobe activity and showed that both normal hearing and CI users with good speech perception exhibited greater cortical activations to natural speech than to unintelligible speech. In contrast, CI users with poor speech perception had large, indistinguishable cortical activations to all stimuli. Zhou et al., 2018, found that regions of temporal and frontal cortex had significantly different mean activation levels in response to auditory speech in CI listeners compared with normal-hearing listeners, and these activation levels were negatively correlated with CI users’ auditory speech understanding.

Our current study differs in that we were able to simultaneously measure responses in temporal and frontal lobes with a high-density array. Our finding of increased recruitment of left PFC in CI listeners is broadly consistent with the recruitment of PFC in fNIRS studies in normal-hearing participants listening to simulated distorted speech (Defenderfer et al., 2021; Lawrence et al., 2018). Potential differences in anatomical localization could reflect the difference between the type of the distortion created by an actual CI compared to simulated degraded speech listening scenarios in the previous studies, or in source localization accuracy across different imaging hardware. To our knowledge, our study is the first fNIRS-based study to use a non-verbal spatial working memory task for localizing the PFC ROI in the same population, and it may be that increased use of cross-domain functional localizers will prove to be a useful approach in future work.

Caveats, considerations, and future directions

As with all fNIRS-based functional brain imaging, our ability to image neural activity is limited by the placement of sources and detectors, and in depth to approximately 1 cm of the cortex. Our HD-DOT cap covers the regions of superficial cortex commonly identified in fMRI studies of speech processing, including bilateral superior (STG) and middle temporal gyri (MTG), and pars triangularis of the left IFG (Davis and Johnsrude, 2003; Rodd et al., 2005; Rogers et al., 2020; Wild et al., 2012). We also have good coverage of occipital cortex and middle frontal gyrus. However, in the present study we do not image pars opercularis of the IFG, the cerebellum, or any subcortical structures. Thus, it is certainly possible that additional regions not identified here play different roles in understanding speech for CI listeners, a possibility that will require fNIRS setups with improved cortical coverage and converging evidence from other methods to explore.

We also note that some care should be taken in the degree to which we have identified PFC. As we have pointed out, the functional organization of the frontal cortex is complex and differs from person to person; in particular, fMRI studies have demonstrated that regions responding to language and non-language tasks lie nearby each other (Fedorenko et al., 2012). Although we attempted to improve the interpretation of our results by using a functional localizer, we were not able to define subject-specific ROIs, which would be preferable. Future work with single-subject localizers and functional ROIs is needed to more clearly resolve this issue.

Cognitive demand during speech understanding frequently comes up in discussions of effortful listening (Pichora-Fuller et al., 2016). Although understandable, it is important to keep in mind that the construct of listening effort is not clearly defined (Strand et al., 2021) and different measures of ‘effort’ do not always agree with each other (Strand et al., 2018). Here, we interpret increased activity in PFC as reflecting greater cognitive demand. Although we did not include an independent measure of cognitive challenge, we note that simulated CI speech (i.e., noise-vocoded speech) is associated with delayed word recognition (McMurray et al., 2017) and poorer memory for what has been heard (Ward et al., 2016), consistent with increased dependence on shared cognitive processes (Piquado et al., 2010). Relating activity in PFC to cognitive demand is also consistent with decreased activity in PFC when speech becomes unintelligible (Davis and Johnsrude, 2003).

Finally, we emphasize that in the current study we only measured responses to speech in quiet. Everyday communication frequently occurs in the presence of background noise, which can be particularly challenging for many listeners with CIs (Firszt et al., 2004). Exploring how activity in PFC and other regions fluctuates in response to speech at various levels of background noise would be a highly interesting future extension of this work. Based on parametric modulations of acoustic challenge in listeners with normal hearing, we might expect increasing activity in PFC as speech gets more challenging (Davis and Johnsrude, 2003), followed by cingulo-opercular activity once intelligibility is significantly hampered (Vaden et al., 2013).

In summary, using high-density optical neuroimaging, we found increased activity in PFC in listeners with CIs compared to listeners with normal hearing while listening to words in quiet. Our findings are consistent with a greater reliance on domain-general cognitive processing and provide a potential framework for the cognitive effort that many CI users need to expend during speech perception.

Materials and methods

Subjects

We recruited 21 adult CI patients aged 26–79 years (17 right-handed, 2 left-handed, 2 not available), and 19 age- and sex-matched controls (18 right-handed, 1 left-handed) (demographic information in Table 1). We excluded one CI user due to poor signal quality (evaluated as mean band limited signal-to-noise ratio (SNR) of all source-detectors) and one control due to excessive motion (see Appendix 1 for details). All CI patients had a unilateral right CI (manufacturer: 11 Cochlear, 6 Advanced Bionics, 3 Med-El). All subjects were native speakers of English with no self-reported history of neurological or psychiatric disorders. All aspects of these studies were approved by the Human Research Protection Office (HRPO) of the Washington University School of Medicine. Subjects were recruited from the Washington University campus and the surrounding community (IRB 201101896, IRB 201709126). All subjects gave informed consent and were compensated for their participation in accordance with institutional and national guidelines.

Table 1
Demographic information.
MeasureControlCI users
Number of subjects (# of females)18 (11)20 (11)
Mean age at test in years (std)57.57 (12.74)56.80 (14.09)
Mean years of CI use (std)8.10 (6.51)
Mean speech perception score (AzBio sentences in quiet) (std)0.99 (0.01)0.88 (0.09)
Mean right ear 4fPTA (std)16.02 (6.74)21.85 (5.30) with CI on
Mean left ear 4fPTA (std)16.61 (7.67)91.25 (26.77) unaided
Mean left ear 4fPTA at test*, (std)73.28 (37.72)
Mean duration of deafness right ear12.58 (11.74)
  1. If no response at a given frequency, a value of 120 dB HL was assigned.

  2. *

    With hearing aid, if the subject used amplification. Eight out of 20 CI (cochlear implant) users used hearing aids.

  3. 4fPTA (four-frequency pure-tone average), average pure tone threshold at four frequencies (500, 1000, 2000, 4000 Hz).

HD-DOT system

Request a detailed protocol

Data were collected using a continuous-wave HD-DOT system comprised of 96 sources (LEDs, at both 750 and 850 nm) and 92 detectors (coupled to APDs, Hamamatsu C5460-01) to enable HbO and HbR spectroscopy (Figure 5; Eggebrecht et al., 2014). The design of this HD-DOT system provides more than 1200 usable source-detector measurements per wavelength at a 10 Hz full-field frame rate. This system has been validated for successfully mapping cortical responses to language and naturalistic stimuli with comparable sensitivity and specificity to fMRI (Eggebrecht et al., 2014; Fishell et al., 2019; Hassanpour et al., 2015).

High-density diffuse optical tomography (HD-DOT) system and the experimental design.

(A) Schematic of a subject wearing the HD-DOT cap along with an illustration of the task design. (B) Simplified illustration of the HD-DOT system (far left), regional distribution of source-detector light levels (middle), and source-detector pair measurements (~1200 pairs) as gray solid lines illustrated in a flat view of the HD-DOT cap (far right). (C) An example point spread function (PSF) and the HD-DOT sensitivity profile illustrated on the volume and spatially registered on the cortical view of a standard atlas in lateral and posterior views.

Experimental design

Request a detailed protocol

Subjects were seated on a comfortable chair in an acoustically isolated room facing an LCD screen located 76 cm from them, at approximately eye level. The auditory stimuli were presented through two speakers located approximately 150 cm away at about ±21° from the subjects’ ears at a sound level of approximately 65 dBA. Subjects were instructed to fixate on a white crosshair against a gray background while listening to the auditory stimuli, holding a keyboard on their lap for the stimuli that required their response (Figure 5A, left panel). The HD-DOT cap was fitted to the subject’s head to maximize optode-scalp coupling, assessed via real-time coupling coefficient readouts using an in-house software. The stimuli were presented using Psychophysics Toolbox 3 (Brainard, 1997) (RRID:SCR_002881) in MATLAB 2010b.

The spoken word recognition paradigm consisted of six blocks of spoken words per run. Each block contained 15 s of spoken words (one word per second), followed by 15 s of silence. Two runs were performed in each study session with a total of 180 words in about 6 min (Figure 5A, middle panel). This task was first introduced by Petersen et al., 1988, and subsequently replicated in other studies with the same HD-DOT system used in this study (Eggebrecht et al., 2014; Fishell et al., 2019; Sherafati et al., 2020). All sound files were generated from a list of short simple nouns (one to four syllables) pronounced with a female digital-voice audio using AT&T Natural Voices text-to-speech generator with the ‘Lauren’ digital voice.

To better understand the left PFC activity we observed in our first few CI users, we adopted a spatial working memory task introduced in previous studies (Fedorenko et al., 2011; Fedorenko et al., 2013) in the remaining subjects to aid in functionally localizing domain-general regions of PFC. In this spatial working memory task, subjects were asked to remember four locations (easy condition) or eight locations (hard condition) in a 3×4 grid, appearing one at a time. Following each trial, subjects had to choose the pattern they saw among two choices, one with correct and one with incorrect locations. This task requires keeping sequences of elements in memory for a brief period and has been shown to activate PFC. Each run for the spatial working memory task was about 8 min, with a total of 48 trials in the run (Figure 5A, right panel).

Data processing

Request a detailed protocol

HD-DOT data were pre-processed using the NeuroDOT toolbox (Eggebrecht and Culver, 2019). Source-detector pair light level measurements were converted to log-ratio by calculating the temporal mean of a given source-detector pair measurement as the baseline for that measurement. Noisy measurements were empirically defined as those that had greater than 7.5% temporal standard deviation in the least noisy (lowest mean motion) 60 s of each run (Eggebrecht et al., 2014; Sherafati et al., 2020). The data were next high pass filtered at 0.02 Hz. The global superficial signal was estimated as the average across the first nearest neighbor measurements (13 mm source-detector pair separation) and regressed from all measurement channels (Gregg et al., 2010). The optical density time-traces were then low pass filtered with a cutoff frequency of 0.5 Hz to the physiological brain signal band and temporally downsampled from 10 to 1 Hz. A wavelength-dependent forward model of light propagation was computed using an anatomical atlas including the non-uniform tissue structures: scalp, skull, CSF, gray matter, and white matter (Ferradal et al., 2014; Figure 5C). The sensitivity matrix was inverted to calculate relative changes in absorption at the two wavelengths via reconstruction using Tikhonov regularization and spatially variant regularization (Eggebrecht et al., 2014). Relative changes in the concentrations of oxygenated, deoxygenated, and total hemoglobin (ΔHbO, HbR, ΔHbT) were then obtained from the absorption coefficient changes by the spectral decomposition of the extinction coefficients of oxygenated and deoxygenated hemoglobin at the two wavelengths (750 and 850 nm). After post-processing, we resampled all data to a 3 × 3 × 3 mm3 standard atlas using a linear affine transformation for group analysis. In addition to the standard HD-DOT pre-processing steps used in the NeuroDOT toolbox, we used a comprehensive data quality assessment pipeline (see Appendix 1) to exclude the subjects with low pulse SNR or high motion levels.

After pre-processing, the response for the speech task was estimated using a standard general linear model (GLM) framework. The design matrix was constructed using onsets and durations of each stimulus presentation convolved with a canonical hemodynamic response function (HRF). This HRF was created using a two-gamma function (2 s delay time, 7 s time to peak, and 17 s undershoot) fitted to the HD-DOT data described in a previous study (Hassanpour et al., 2015). We included both runs for each subject in one design matrix using custom MATLAB scripts (Appendix 1—figure 2).

For modeling the spatial working memory task, we used a design matrix with two columns representing easy and hard conditions. The duration of each easy or hard trial was modeled as the total time of stimulus presentation and evaluation. Events were convolved with the same canonical HRF described in the spoken word perception task to model hemodynamic responses (Hassanpour et al., 2014). We used the easy+hard response maps as a reference for defining the PFC ROI, which was more robust than the hard>easy contrast previously used for younger populations.

After estimating the response (β map) for each subject for each task, we performed a second-level analysis in SPM12 (Wellcome Trust Centre for Neuroimaging) version 7487 (RRID:SCR_007037). Extracted time-traces for each subject were then calculated using a finite impulse response model.

We only present the ΔHbO results in the main figures as we have found that the ΔHbO signal exhibits a higher contrast-to-noise ratio compared to ΔHbR or ΔHbT (Eggebrecht et al., 2014; Hassanpour et al., 2014).

Functionally defined ROIs

Request a detailed protocol

To perform a more focused comparison between controls and CI users, we defined three ROIs, independent from our spoken word recognition dataset, for statistical analysis.

To accurately localize the elevated PFC activation in the CI group, we collected HD-DOT data from nine subjects (four controls and five CI users in 13 sessions) using a spatial working memory task (Fedorenko et al., 2012). The visual spatial working memory task robustly activates PFC due to its working memory demands (and visual cortex because of its visual aspect). We chose this task to localize the PFC ROI for performing an ROI-based statistical analysis between controls and CI users. Our logic is that in prior work this task shows activity that dissociates from nearby language-related activity (Fedorenko et al., 2012), and thus the region of PFC localized by this task is more likely to reflect domain-general processing (as opposed to language-specific processing). Our results show strong bilateral visual and PFC activations in response to this task (Figure 6A left). We then defined the left PFC ROI as the cluster of activation in the left PFC (Figure 6A right). This region was centered in inferior frontal sulcus, extending into both dorsal IFG and inferior MFG. The location is broadly consistent with domain-general (‘multiple demand’) activation (e.g., meta-analysis in Duncan, 2010).

Defining functional regions of interest (ROIs).

(A) Spatial working memory activation for five CI users and four controls over 13 sessions. The prefrontal cortex (PFC) ROI was defined as the cluster of activation in the PFC region, after p<0.05 (uncorrected) voxelwise thresholding. (B) Seed-based correlation map for a seed located in the left auditory cortex (left map). Right auditory ROI defined by masking the correlation map to include only the right hemisphere (right map). (C) Seed-based correlation map for a seed located in the right auditory cortex (left map). Left auditory ROI defined by masking the correlation map to include only the left hemisphere (right map).

To define the left and right auditory ROIs, we used fMRI resting state data from a previously published paper (Sherafati et al., 2020) that was masked using the field of view of our HD-DOT system. We defined the left and right auditory ROIs by selecting a 5 mm radius seed in the contralateral hemisphere and finding the Pearson correlation between the time-series of the seed region with all other voxels in the field of view. Correlation maps in individuals were Fisher z-transformed and averaged across subjects (Figure 6B–C, left). Right/left auditory ROIs were defined by masking the correlation map to include only the right/left hemisphere (Figure 6B–C right). These ROIs extend well beyond primary auditory cortex, which is important to capture responses to words often observed in lateral regions of STG and MTG.

Speech perception score

Request a detailed protocol

To measure auditory-only speech perception accuracy, we presented each participant with two lists of AzBio sentences in quiet (Spahr et al., 2012) (except one participant who only heard one set). We calculated speech perception accuracy as the proportion of correctly repeated words across all sentences.

Hearing threshold

Request a detailed protocol

We summarized hearing thresholds using a four-frequency pure-tone average (4fPTA), averaging thresholds at 500, 1000, 2000, and 4000 Hz. For unaided testing, a value of 120 dB was assigned if there was no response at a given frequency; for aided or CI testing, a value of 75 dB was used. For CI users, hearing was tested with the right ear CI alone, left ear alone unaided, and left ear with a hearing aid, if worn at the time of testing. Figure 4—figure supplement 1 shows the audiograms for both controls and CI users.

Quantification and statistical analysis

Request a detailed protocol

Based on prior optical neuroimaging studies with similar speech-related tasks (Defenderfer et al., 2021; Eggebrecht et al., 2014; Hassanpour et al., 2015; Pollonini et al., 2014; Zhou et al., 2018), we anticipated 15–20 subjects per group would be sufficient to detect a moderate effect size while also within our resource limitations. We therefore aimed to have at least 15 subjects in each group. We performed a one-tailed two-sample t-test for our main hypothesis (increased recruitment of PFC in CI users), for which we had a strong directional hypothesis (greater activity in listeners with CIs than controls), and a two-tailed t-test for the left and right auditory cortex changes for which we did not have directional hypotheses. We adjusted for unequal variances for left PFC and right auditory ROIs based on the significance of Levene’s test.

Appendix 1

Data quality assessment steps

Motion artifact detection

Motion artifacts were detected in a time-point by time-point manner on the 10 Hz bandpass filtered first nearest neighbor (13 mm) measurements, using a previously described motion detection method in HD-DOT, the global variance of the temporal derivatives (GVTD) (Sherafati et al., 2018; Sherafati et al., 2017; Sherafati et al., 2020).

GVTD, similar to DVARS (derivative of variance) in fMRI (Power et al., 2012; Smyser et al., 2011), is a vector g that is defined as the RMS of the temporal derivatives (yjiyji1) across a set of optical density measurements.

(1) g=[g1gM],gi=1Nj=1N(yjiyji1)2

N is the number of first nearest neighbor measurements, and M is the number of time-points. GVTD indexes the global instantaneous change in the time-traces. Higher GVTD values indicate high motion levels. We defined the motion criterion (gthresh) (red line in Appendix 1—figure 1 and 4) based on the GVTD distribution mode (κ) plus a constant c = 10 times the standard deviation computed on the left (low) side of the mode (σL).

gthresh=κ+cσL,

Appendix 1—figure 1 shows two example GVTD time-traces for a low-motion and a high-motion run. The GVTD threshold was calculated as gthresh=κ+10σL and is shown as a red line.

Appendix 1—figure 1
Examples of global variance of the temporal derivative (GVTD) time-traces for a low-motion and a high-motion spoken word recognition high-density diffuse optical tomography (HD-DOT) data.

The red lines indicate the GVTD threshold of κ+10σL of each run.

Including motion regressors in the design matrix

After determining the time-points that passed the GVTD threshold, we then included a motion regressor as a column for each time-point that passed the motion threshold with one for the noisy time-point, and zeros for every other point in the design matrix, a method commonly known as one-hot encoding. An example design matrix including two runs in a session of spoken word recognition task along with their motion regressors are shown in Appendix 1—figure 2.

Appendix 1—figure 2
Including motion regressors in the design matrix.

(A) An example design matrix for the general linear model (GLM) for the spoken word recognition task, including a constant column (cst), the task times for run1, one-hot encoding columns for time-points passing the global variance of the temporal derivative (GVTD) threshold for run 1, task times for run2, and one-hot encoding columns for time-points passing the GVTD threshold for run 2. (B) The GVTD time-traces and the temporal masks for excluding the time-points passing the GVTD threshold are shown in (B) for run 1 and (C) for run 2.

Heartbeat (pulse) SNR across HD-DOT field of view

One indicator of detecting physiological signal in optical neuroimaging is the detection of heartbeat frequency (pulse) (Sherafati, 2020). If the data has a very low heartbeat SNR it indicates that the lower frequencies emergent from the hemodynamic fluctuations might also have a low SNR. Therefore, we excluded the subjects that showed a very low mean pulse SNR across the field of view, as it indicates that their brain signal is also less reliable.

The underlying reason behind low pulse SNR is either a poor optode-scalp coupling or high levels of motion. If the optode-scalp coupling is good, normally pulse SNR is only low during high-motion (identified as high GVTD) epochs of data. However, if the optode-scalp coupling is poor, we will see low pulse SNR throughout the run. Note that there are not well-studied definitions and cutoffs for the calculation of pulse SNR in fNIRS-based methods. Therefore, since we already regressed time-points contaminated with motion using the GVTD index, here we only exclude the subjects that had a very low mean pulse SNR across the cap compared to other subjects.

We calculated the pulse SNR using the NeuroDOT function PlotPhysiologyPower.m (https://github.com/WUSTL-ORL/NeuroDOT_Beta; Eggebrecht and Culver, 2019). This function first calculates the fast Fourier transform (FFT) of the optical density signal (log mean ratio). Then, it calculates the pulse SNR based on the ratio of the band-limited pulse power ~1 Hz (P) and the noise floor (N).

The band-limited pulse power (P) is defined as the sum of the squares of the FFT magnitudes of a small frequency band around the peak of FFT magnitudes in the 0.5–2 Hz frequency bandwidth. The noise floor (N) is then defined as the median of the squares of the FFT magnitudes of the FFT indices in the same 0.5–1 Hz frequency band, excluding the indices defined as the peak frequency window (used for the heartbeat pulse power). Finally, the pulse SNR was defined as 10×log10(P/N).

For pulse SNR calculation, we only used the second nearest neighbor measurements which penetrate through the cortex (30 mm source-detector separation), second wavelength (850 nm), and good measurements (ones with <7.5% std). Appendix 1—figure 3 shows an example for high SNR values across the field of view (top) and its signature in a subset of the measurements for the same run (bottom) (Appendix 1—figure 3A) and an example for low SNR values across the field of view (top) and a subset of measurements for that run (bottom) (Appendix 1—figure 3B).

Appendix 1—figure 3
Examples of a good and a bad pulse signal-to-noise ratio (SNR) in high-density diffuse optical tomography (HD-DOT) data.

Example pulse SNR plot and a selection of the measurements from the HD-DOT array for (A) a high-quality pulse SNR, and (B) a low-quality pulse SNR. Note the heartbeat frequency (~1 Hz) that appears as around 10 peaks in 10 s.

Exclusion of subjects based on low pulse and high motion

We rank ordered all subjects based on their mean band limited SNR across the field of view and their mean GVTD values (averaged over two spoken word recognition runs for each subject). We then excluded one of the CI users due to a very low band limited SNR (Appendix 1—figure 4A) and one control due to a very high motion level (Appendix 1—figure 4B).

Appendix 1—figure 4
Exclusion of subjects with very low band limited signal-to-noise ratio (SNR) and very high motion levels.

Sorting all subjects based on (A) mean band limited SNR value across the high-density diffuse optical tomography (HD-DOT field of view, and (B)) mean global variance of the temporal derivative (GVTD) values across spoken word recognition runs. The red boxes indicate the subjects excluded based on each quality score.

Appendix 2

Simulating the effect of the CI transducer blocking the optodes

We identified the approximate area of the HD-DOT field of view that fell under the transducer across people. In most subjects, the transducer only blocked three to five optodes. Due to the high density of the optodes in HD-DOT systems (13 mm nearest source-detector distance), even when an optode is blocked, the light still penetrates through that region from the neighboring sources. Therefore, in most cases, light level plots reveal sensitivity reduction in only two or three optodes. Appendix 2—figure 1A shows the location of the CI transducer in one CI user and the optode numbers. Appendix 2—figure 1B shows the sensitivity of the HD-DOT cap if all the optodes are fully coupled with the scalp. Appendix 2—figure 1C and D demonstrate that the sensitivity drop due to the CI transducer is mainly affecting the IFG and MTG when four and six source-detector pairs are simulated to be completely blocked. Note that the right auditory ROI (Figure 6B) mainly includes the STG.

Appendix 2—figure 1
Effects of the cochlear implant (CI) transducer on high-density diffuse optical tomography (HD-DOT) sensitivity.

(A) A right ear CI user wearing the HD-DOT cap. The yellow contours around the CI transducer with respect to the ear fiducials illustrate the source numbers (encircled in red) and detector numbers (encircled in blue) around the transducer. (B) The left panel shows the HD-DOT source-detector grid, overlaid on the mesh used in the study for image reconstruction. The right panel shows the sensitivity of the HD-DOT cap around the cortex including all optodes. (C) The left panel shows four sources (in red) and four detectors (in blue) excluded from the sensitivity calculation. The right panel shows the sensitivity of the HD-DOT cap excluding those sources and detectors. (D) The left panel shows six sources (in red) and six detectors (in blue) excluded from the sensitivity calculation. The right panel shows the sensitivity of the HD-DOT cap excluding those sources and detectors. Note that the exclusion of sources and detectors in (C) and (D) only resulted in sensitivity drop off in the inferior and middle temporal gyri.

Data availability

Stimuli, data, and analysis scripts are available from https://osf.io/nkb5v/.

The following data sets were generated
    1. Sherafati A
    2. Bajracharya A
    3. Peelle JE
    (2022) Open Science Framework
    ID https://osf.io/nkb5v/. Prefrontal cortex supports speech perception in listeners with cochlear implants.

References

    1. Davis MH
    2. Johnsrude IS
    (2003)
    Hierarchical processing in spoken language comprehension
    The Journal of Neuroscience 23:3423–3431.
  1. Conference
    1. Sherafati A
    2. Eggebrecht AT
    3. Burns-Yocum TM
    4. Culver JP
    (2017)
    A global metric to detect motion artifacts in optical neuroimaging data (Conference Presentation
    In Neural Imaging and Sensing.
  2. Book
    1. Sherafati A
    (2020)
    Separating Signal from Noise in High-Density Diffuse Optical Tomography
    Washington University in St. Louis.

Decision letter

  1. Timothy D Griffiths
    Reviewing Editor; University of Newcastle, United Kingdom
  2. Barbara G Shinn-Cunningham
    Senior Editor; Carnegie Mellon University, United States
  3. Timothy D Griffiths
    Reviewer; University of Newcastle, United Kingdom
  4. Phillip E Gander
    Reviewer; University of Iowa, United States

Our editorial process produces two outputs: i) public reviews designed to be posted alongside the preprint for the benefit of readers; ii) feedback on the manuscript for the authors, including requests for revisions, shown below. We also include an acceptance summary that explains what the editors found interesting or important about the work.

Decision letter after peer review:

Thank you for submitting your article "Dorsolateral prefrontal cortex supports speech perception in listeners with cochlear implants" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, including Timothy D Griffiths as Reviewing Editor and Reviewer #1, and the evaluation has been overseen by Barbara Shinn-Cunningham as the Senior Editor. The following individual involved in the review of your submission has agreed to reveal their identity: Phillip E Gander (Reviewer #2).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential revisions:

The authors should address the following major points in a revision that are highlighted in detail in the individual referee reports.

1. Acknowledgement of and discussion of technical limitations. These are related to the depth of signal and field of view which means that the whole of the network for speech perception is not analysed.

2. Discussion of task effects. There is an effect shown in the frontal cortex listening to the world in quiet but no measurement of correlates of more challenging listening that might show this with the greatest power or demonstrate additional areas. And although listening effort is invoked to explain the data there are no physiological measurements of this.

3. Clearer definition and justification of the ROI within which extra activation occurs. The result in the frontal ROI is just significant without correction for multiple comparisons so it is important that the ROI is clearly justified and explained

Reviewer #1 (Recommendations for the authors):

I would like the authors to consider these comments. Basically, I think the demonstration of the DLPFC signal increase during speech perception by the CI group supports models based on the use of additional resources beyond the traditional speech network. An issue for me is that the technique does not allow the examination of other candidate areas.

1. The technique is an advance on fNIRS used previously in terms of spatial resolution but suffers from the same issue with respect to the demonstration of deep sources. This is helpfully shown by the authors in the volumetric field of view diagram in figure 1. The demonstration of the DLPFC involvement in the listening task is robust in the group data including the ROI analysis: this area can be optimally imaged with the technique. The field of view excludes deep parts of the superior temporal plane including the medial parts of the primary auditory cortex and planum temporale behind it, and the deep opercular part of IFG. This prevents the comparison of the signal between CI users and controls in all of the conventional parts of the language network (and other parts of the multiple demand network discussed below).

2. With respect to the decrease in signal in the left auditory cortex in the CI users I think the interpretation is limited because the medial primary cortex is not imaged. I agree with the authors' suggestion that the change here likely reflects a decrease in high-level processing of speech sounds (in the cortex near the convexity) but the discussion might include acknowledgement of the technical limitation due to field of view.

3. The demonstration of increased signal in the left DLPFC is the most robust finding here and is interesting in and of itself. This is consistent with the idea advanced by the senior author that more difficult listening engages anterior cingulate, frontal operculum, premotor cortex, DLPFC and parietal cortex (eg PMID: 28938250 figure 3). Of these areas, only the DLPFC is adequately imaged by this technique (although the authors might comment on whether the inferior parietal cortex is within FOV and demonstrated by contrast in figure 4C). The discussion here focuses more on the multiple-demand network idea and uses a WM task similar to Duncan and others to define this but from this perspective the imaging of the deep operculum, frontal lobe and medial frontal lobe is still missing.

4. I found it hard to interpret the between-subject effect of behavioural scores in figure 6. The effect of left ear hearing (non-CI side) goes in a direction I would predict if the DLPFC signal reflected listening effect and the right ear hearing (CI side) goes in the opposite direction. I guess the range of hearing deficit is broader on the left and this has been emphasised more in the discussion. The authors might comment.

5. Do the authors have any data on varying behaviour within subjects like presenting the sound in noise? Again, any effect on DLPFC would be interesting. But I think the main point here is the use of a different network as a function of CI use even in quiet.

Reviewer #2 (Recommendations for the authors):

1. Hypothesis/results interpretation:

Can the justification for the hypothesis be unpacked a little more, for this particular case? If challenging listening requires more processing in a wider array of brain regions to take over for greater cognitive demand, for example in DLPFC. And CI users experience greater cognitive demand during listening, one might hypothesize that listening to single words without background noise would be the hardest condition in which to find an effect (despite the finding of Dwyer et al. 2014). Alternatively, one might compare words in quiet with words in noise, since speech in noise understanding is the main complaint of people with hearing loss (especially CI users). And the brain mechanisms required under these conditions might be different to what is a less challenging task. Perhaps the authors wanted to start in the most "trivial" case and test more conditions later.

Figure 4. The difference t-map has a somewhat loose threshold. Presumably this is because the differences in the frontal cortex would be lost. The authors have previously used more stringent thresholding, but perhaps there is justification here.

The CI>Control t-map shows the strongest differences in the visual/visual association cortex, auditory cortex. On the left it appears that a relevant frontal activation occurs in the inferior frontal gyrus, as might be predicted from the literature.

Reduced activation in the left auditory cortex between CI and control could still be due to threshold differences ~5dB (may be significant). Peripheral fidelity is much lower in the CI group, with also poor contribution ipsilaterally.

The authors propose that lower left auditory cortex activation could also be related to lower speech perception scores, however these scores are drawn from the AzBio test (sentences in noise) which has the potential for indexing very different processes than a word in the quiet test as was the case in this study.

2. ROI definition:

Choosing an ROI functionally defined from a spatial WM task to look for overlap in brain regions sharing a putative cognitive process is unclear if the comparison is a verbal task. If the goal was to find common cognitive resources the referenced paper by Federenko and colleagues achieved such a goal, providing an ROI with which to test against a common set of cognitive processes. Further, the claim that the task activates a brain region, DLFPC, which does seem to include dorsal aspects of inferior frontal gyrus on pars triangularis. Yet the activation found include this region in this study, while this region of pars triangularis does not appear in the work by Federenko and colleagues leave the reader confused about why these differences would exist, unless it was for other reasons, like localization differences between HD-DOT and fMRI.

The results might change considerably if anatomical ROIs were used instead, i.e., IFG or a DLPFC ROI.

It is fine to use the logic that one task (A) primarily indexes cognitive function X (hopefully) and let's see if another task (B) also shares some of that presumed function by masking B with A. That shows (hopefully) that the functions are shared across tasks. In this case, the function of (spatial) working memory is shared between a spatial working memory and speech task in CI users.

DLPFC is an anatomical label, not a functional one. In this case, the authors have used a functional ROI and so need to name it as such, e.g., 'WM', while reference to the term DLPFC is not helpful and should be removed from the manuscript (and title) except perhaps in the discussion.

Line 194. It is slightly confusing to define auditory ROIs from a fMRI resting state paradigm. The referenced paper used a 'word listening' activation for both fMRI and HD-DOT. So are the authors using the acoustic word task rather than rest? And further, if there is data using HD-DOT with a similar paradigm then that might be a more sensible ROI definition?

Line 195-199. The procedure for the definition of the ROI is not clearly motivated, in particular, because the data are obtained from a previous publication which displays thresholded values (at least for the auditory task) that clearly defines the auditory cortex. A correlational approach might (appropriately) be very broad and include areas outside of 'canonical' auditory temporal cortex. Was this intended? What was the threshold for the correlation?

Line 240-245. This text implies the ROI was defined with the current data set rather than another as written in the methods. Please clarify.

Figure 3A. The 'DLPFC' ROI seems to have a large proportion of the cluster within pars triangularis, which is not quite the same as the fMRI overlap representation in Figure 2 of Federnko et al. 2013, in which the bulk of activation is within the middle frontal gyrus.

3. HD-DOT sensitivity:

As stated in the referenced White and Culver paper: "The high-density array has an effective resolution of about 1.3 cm, which should prevent identification to the wrong gyrus."

This means that localization errors to middle frontal gyrus, as opposed to inferior frontal gyrus, could be a concern in this case in which the activation appears to be across a sulcal boundary.

Figure 1. The sensitivity profile appears to show that greater sensitivity exists for the superior aspect of inferior frontal gyrus and inferior aspect of the middle frontal gyrus (and visual cortex). Is that a simple misinterpretation from the figure? Or do the authors know of issues with respect to the HD-DOT technique weighting responses to the regions of greater sensitivity?

Figure 1 supplement. Is it possible there may be some concern about the drop in signal on the right due to variability in the location of the CI transducer across participants? Is it possible the drop in signal in right middle/superior temporal and inferior frontal is also due to loss of measurement here across the group of CI users?

Specific comments:

1. Lines 56-57: Please clarify. Are you saying that listening to degraded input reduces the ability for certain other processing, for instance, performance on a second task that involves memory encoding? Or is it that memory encoding on the speech is limited, or both? It is not entirely clear from the text.

2. What kind of CI device? Since it is not specified, I assume this is a standard length implant and not one for hearing preservation.

3. What was the distribution of CI duration? Is it possible there may be some people under 1 year of use? If so this may be difficult with respect to interpretation for those users, since their performance with the device may not be their best, i.e., they may still yet improve. This process of adaptation to the implant could easily involve different brain regions other than measured in this study, or a different weighting depending on adaptation.

4. Similarly, does the range of unaided hearing in the CI group get to the level of some "functional" residual hearing, or does the skew go to higher thresholds? If there are people with PTAs of ~60 they may perform differently than people without functional residual hearing? Perhaps residual hearing could be added as a covariate to the analyses.

There may also be clues based on performance on the AzBio task. Is there a relationship between residual hearing and AzBio score? The thrust of these questions is that residual hearing likely has an impact on performance, which similarly may influence brain mechanisms.

5. Figure 1. The Spatial WM task appears to have multiple components within it that would be beyond just WM, for example, reward (from the feedback), decision making, response execution, which would all be tied into the level of temporal resolution of the imaging technique. This might temper conclusions about WM per se.

6. Visual activation appears only on the left, which is unclear (compare to Figure 3A). Was the display directly in-front of the participant, or something about their vision?

7. Figure 1. In the spatial working memory task the wait time range of 4000-250ms is unclear.

8. Experimental design. Please provide more information about the following: presentation level for acoustic stimuli, types of words spoken, e.g., monosyllabic, nouns, single talker, gender, etc.

9. Line 149. Confusing text: "…regressed from the all measurement…"

10. Line 176. Was the 'easy+hard' contrast used to provide better SNR?

Alternatively, why not use the ROI defined previously? Did they vary, because if they did that is instructive, perhaps as implied by aging effects?

11. Line 191. Which group of subjects? Is this not the study group as implied on line 176?

12. Line 260. Why does the supplemental analysis not include a contrast of control > CI?

13. Line 304-305. Incomplete sentence.

14. Line 359. The cited literature of similar findings in normal listeners, mostly highlights the inferior frontal gyrus, however, middle frontal gyrus is included in Defenderfer et al.

Reviewer #3 (Recommendations for the authors):

Introduction

Line 81: Why predict effects in DLPFC only? There could be a range of 'non-core' brain regions that support speech perception. Were the effects lateralized to left DLPFC predicted in advance of the study?

Methods

Line 91: How were the sample sizes of the CI group and the control group determined? I note that page 7 explains that 15-20 subjects per group was thought to be sufficient, based on similar studies. Did these previous studies include a cross-sectional design?

Page 4. Table 1 reports left ear 4fPTA (unaided). The footnote for Table 1 reports that 8/20 CI users had a hearing aid but did they use them during the speech task?

Line 120. Where were the 2 speakers located relative to the participants? +/- 45 degrees? Which sound level were the words presented at?

Lines 127-130. More details about the spoken word task is needed. Is this a standardized test or a novel test that the authors developed themselves?

Line 131. Which 'preliminary results'?

Line 191. "…based on the response … in a group of subjects…". Which subjects? Merge with relevant information on page 8.

Line 216. Is "Pollonini et al. 2014" cited correctly?

Line 238. Why define only left DLPFC as an ROI?

Page 9. The visual spatial working memory task also activated 'auditory' regions in the right hemisphere that look like they overlap with the 'right auditory ROI' created by the authors. Do the authors consider this overlap to be problematic?

Page 9. Legend for Figure 3. "…in the DLPFC region, survived after…". A word is missing.

Page 10. Figure 4C suggests that CI > controls contrast identified quite a few group differences. So why then focus on left DLPFC?

Page 11. Results of two-sample t-tests are reported in the legend of Figure 5. This information should be taken out of the legend and reported in full. Why were t-tests more appropriate than an ANOVA? Please report results of statistical analyses fully i.e. not just p-values. Were reported p-vals corrected for the number of t-tests that were used?

Page 11. The authors report the results of correlation analyses. Please add further details e.g. are Pearson's correlation coefficients reported?

Page 11. The authors report that the threshold used for statistical significance in the correlation analyses was 'uncorrected'. Presumably this means that reported p-vals were not corrected for the 4 correlation analyses shown?

Page 12. Correlation analyses for CI users. The rationale for these analyses is missing. Why carry out simple correlation analyses on data for CI users only? Potential predictors of speech processing (4PTA) could have been used as a covariate in the main analyses (e.g. Figure 5).

https://doi.org/10.7554/eLife.75323.sa1

Author response

Essential revisions:

The authors should address the following major points in a revision that are highlighted in detail in the individual referee reports.

1. Acknowledgement of and discussion of technical limitations. These are related to the depth of signal and field of view which means that the whole of the network for speech perception is not analysed.

Thank you for encouraging us to clarify the limitations in our approach. Of course, we completely agree that though we have good coverage relative to other optical neuroimaging approaches, we do not cover the full brain, and there are indeed brain regions we might expect to be involved in speech perception that we do not image. We have added the following to the Discussion:

“As with all fNIRS-based functional neuroimaging, our ability to image neural activity is limited by the placement of sources and detectors, and in depth to approximately 1 cm of the cortex. Our HD-DOT cap covers the regions of superficial cortex commonly identified in fMRI studies of speech processing, including bilateral superior and middle temporal gyri, and pars triangularis of the left inferior frontal gyrus (Davis and Johnsrude 2003; Rodd, Davis, and Johnsrude 2005; Rogers et al. 2020; Wild et al. 2012). We also have good coverage of occipital cortex and middle frontal gyrus. However, in the present study we do not image pars opercularis of the inferior frontal gyrus, the cerebellum, or any subcortical structures. Thus, it is certainly possible that additional regions not identified here play different roles in understanding speech for CI listeners, a possibility that will require fNIRS setups with improved cortical coverage and converging evidence from other methods to explore.”

2. Discussion of task effects. There is an effect shown in the frontal cortex listening to the world in quiet but no measurement of correlates of more challenging listening that might show this with the greatest power or demonstrate additional areas. And although listening effort is invoked to explain the data there are no physiological measurements of this.

Thank you for raising these two important points.

Regarding the first point, it is important to note that both CI simulations (i.e., noise vocoding) and data from CI listeners show difficulty understanding speech in quiet (Willis, Helen; 2018; The feasibility of the dual-task paradigm as a framework for a clinical test of listening effort in cochlear implant users. Doctoral thesis, University College London). Thus, our perspective is that speech may be “in quiet” in the room, but to the auditory system of a listener with a cochlear implant, the acoustic signal is degraded, and it is precisely the brain’s response to that degraded signal we are measuring. We readily agree that further exploring responses to speech in noise (which, of course, is common in everyday listening, and can be particularly challenging for many listeners with cochlear implants) is highly interesting, and we hope a future extension of this work. We now explicitly note these points in the introduction (to clarify the logic behind studying speech in quiet) and in the Discussion:

“Finally, we emphasize that we report here only responses to speech in quiet. Everyday communication frequently occurs in the presence of background noise, which can be particularly challenging for many listeners with cochlear implants (Firszt et al. 2004). Exploring how activity in PFC and other regions fluctuates in response to speech at various levels of background noise would be a highly interesting future extension of this work. Based on parametric modulations of acoustic challenge in listeners with normal hearing, we might expect increasing activity in PFC as speech gets more challenging (Davis and Johnsrude 2003), followed by cingulo-opercular activity once intelligibility is significantly hampered (Vaden et al. 2013).”

With respect to the second point, it is entirely correct that we do not have a non-imaging measure of cognitive demand during listening. That being said, we think there are good reasons to think that differences here are driven by a response to acoustic challenge, which we believe falls comfortably under the “listening effort” umbrella. However, there are two important clarifications we should make. First, we readily acknowledge that listening effort is not clearly defined (Strand et al., 2020), and in its fullest definition encompasses more than just cognitive challenge (Pichora-Fuller et al., 2016). In fact, in the manuscript we never use the term “listening effort” to describe our paradigm for precisely this reason. We have also done our best to explain where our cognitive-demand-centered interpretation comes from. In short, from imaging studies showing parametric responses to acoustic challenge (including noise vocoding) (e.g., Davis and Johnsrude, 2003), behavioral studies showing that acoustic challenge hinders memory for words (Rabbitt, 1968) and stories (Ward et al., 2016); and the many reports of listeners with CIs having difficulty understanding speech (Firszt et al., 2004, McMurray et al. 2017, and many others). We have added these points to the Discussion:

“Cognitive demand during speech understanding frequently comes up in discussions of effortful listening (Pichora-Fuller et al. 2016). Although understandable, it is important to keep in mind that the construct of listening effort is not clearly defined (Strand et al. 2020) and different measures of “effort” do not always agree with each other (Strand et al. 2018). Here, we interpret increased activity in PFC as reflecting greater cognitive demand. Although we did not include an independent measure of cognitive challenge, we note that simulated CI speech (i.e., noise-vocoded speech) is associated with delayed word recognition (McMurray, Farris-Trimble, and Rigler 2017) and poorer memory for what has been heard (Ward et al., 2016), consistent with increased dependence on shared cognitive processes (Piquado et al. 2010). Relating activity in PFC to cognitive demand is also consistent with decreased activity in PFC when speech becomes unintelligible (Davis and Johnsrude 2003).”

3. Clearer definition and justification of the ROI within which extra activation occurs. The result in the frontal ROI is just significant without correction for multiple comparisons so it is important that the ROI is clearly justified and explained

Thank you for giving us the opportunity to clarify these important points. We start by discussing the definition and justification for the PFC region of interest, followed by a description of a revised analysis.

We have added some additional detail to the justification of the prefrontal cortex ROI. The full revised text reads:

“To accurately localize the elevated prefrontal cortex activation in the CI group, we collected HD-DOT data from 9 subjects (4 controls and 5 CI users in 13 sessions) using a spatial working memory task (Fedorenko, Duncan, and Kanwisher 2012). The visual spatial working memory task robustly activates PFC due to its working memory demands (and visual cortex because of its visual aspect). We chose this task to localize the PFC ROI for performing an ROI-based statistical analysis between controls and CI users. Our logic is that in prior work this task shows activity that dissociates from that in inferior frontal gyrus (Fedorenko, Duncan, and Kanwisher 2012), and thus the region of PFC localized by this task is more likely to reflect domain-general processing (as opposed to language-specific processing). Our results show strong bilateral visual and PFC activations in response to this task (Figure 6A left). We then defined the left PFC ROI as the cluster of activation in the left PFC region (Figure 6A right). This region was centered in inferior frontal sulcus, extending into both dorsal IFG and inferior MFG. The location is broadly consistent with domain-general (“multiple demand”) activation (e.g., meta-analysis in Duncan, 2010).”

In terms of anatomy, our frontal cortex ROI is centered in the inferior frontal sulcus (IFS), which also corresponds to peak activity in meta analyses of “multiple demand” tasks—for example, in this meta-analysis from Duncan (2010, Trends in Cognitive Science).

Author response image 1

Being centered in IFS, there is some overlap with both IFG and MFG. If we trust our source localization, we don’t think the IFG overlap is inferior enough to suggest language-related processing, although we concede we do not have definitive evidence for this. The ROI is unquestionably in the prefrontal cortex. In our view, centered in IFS is sufficient for considering this dorsolateral prefrontal cortex. However, we have gone with the more general “prefrontal cortex” in the title and text of the manuscript because we don’t think this point affects our main findings or interpretation.Regarding correction for multiple comparisons, we have revisited this section and clarified that the critical comparison in PFC is corrected for the ROIs looked at.

Reviewer #1 (Recommendations for the authors):

I would like the authors to consider these comments. Basically, I think the demonstration of the DLPFC signal increase during speech perception by the CI group supports models based on the use of additional resources beyond the traditional speech network. An issue for me is that the technique does not allow the examination of other candidate areas.

1. The technique is an advance on fNIRS used previously in terms of spatial resolution but suffers from the same issue with respect to the demonstration of deep sources. This is helpfully shown by the authors in the volumetric field of view diagram in figure 1. The demonstration of the DLPFC involvement in the listening task is robust in the group data including the ROI analysis: this area can be optimally imaged with the technique. The field of view excludes deep parts of the superior temporal plane including the medial parts of the primary auditory cortex and planum temporale behind it, and the deep opercular part of IFG. This prevents the comparison of the signal between CI users and controls in all of the conventional parts of the language network (and other parts of the multiple demand network discussed below).

We acknowledge the limitation of our system field of view in covering the entirety of brain networks involved in speech processing, particularly deep structures. We agree this is an important point, and now note this explicitly in the Discussion.

2. With respect to the decrease in signal in the left auditory cortex in the CI users I think the interpretation is limited because the medial primary cortex is not imaged. I agree with the authors' suggestion that the change here likely reflects a decrease in high-level processing of speech sounds (in the cortex near the convexity) but the discussion might include acknowledgement of the technical limitation due to field of view.

This comment is very important and we thank the reviewer for reminding us of this fact. We have added a sentence in the Discussion section to address this point.

3. The demonstration of increased signal in the left DLPFC is the most robust finding here and is interesting in and of itself. This is consistent with the idea advanced by the senior author that more difficult listening engages anterior cingulate, frontal operculum, premotor cortex, DLPFC and parietal cortex (eg PMID: 28938250 figure 3). Of these areas, only the DLPFC is adequately imaged by this technique (although the authors might comment on whether the inferior parietal cortex is within FOV and demonstrated by contrast in figure 4C). The discussion here focuses more on the multiple-demand network idea and uses a WM task similar to Duncan and others to define this but from this perspective the imaging of the deep operculum, frontal lobe and medial frontal lobe is still missing.

We focused on the PFC ROI for two reasons. First, our reading of prior fMRI literature (and work from our own group) suggests that when speech is challenging to understand, but still intelligible, increases are seen in many regions of the dorsolateral prefrontal cortex. However, when intelligibility begins to suffer (i.e., errors are introduced), the cingulo-opercular network is engaged (see compelling fMRI work from Eckert, Vaden, and colleagues). Because we were presenting speech in quiet, to CI users with fairly good auditory-only perception, we focused on the PFC activity. As you point out, we also do not have the ability to image deep areas—and so we are not particularly well-suited for studying the cingulo-opercular network.

We had considered including an inferior parietal ROI, which as you say, is also often implicated in multiple-demand tasks. However, we did not have a clear ROI defined based on our spatial working memory task. Thus, we did not feel justified in including one. (It is also unclear whether our parietal coverage adequately covers IPS—it may, but the absence of strong functional activity here led us to abandon this idea for now.)

4. I found it hard to interpret the between-subject effect of behavioural scores in figure 6. The effect of left ear hearing (non-CI side) goes in a direction I would predict if the DLPFC signal reflected listening effect and the right-ear hearing (CI side) goes in the opposite direction. I guess the range of hearing deficit is broader on the left and this has been emphasised more in the discussion. The authors might comment.

We agree with this comment. We also think that our rather small sample size and the small variability in the right ear hearing levels, makes our PFC correlation with right ear 4fPTA harder to interpret. We added a sentence to state this more clearly in the discussion.

5. Do the authors have any data on varying behaviour within subjects like presenting the sound in noise? Again, any effect on DLPFC would be interesting. But I think the main point here is the use of a different network as a function of CI use even in quiet.

Because we have brain imaging data for speech in quiet, we felt that the strongest justification for brain-behavior correlations would be to use behavioral scores from a similar condition.

Reviewer #2 (Recommendations for the authors):

1. Hypothesis/results interpretation:

Can the justification for the hypothesis be unpacked a little more, for this particular case? If challenging listening requires more processing in a wider array of brain regions to take over for greater cognitive demand, for example in DLPFC. And CI users experience greater cognitive demand during listening, one might hypothesize that listening to single words without background noise would be the hardest condition in which to find an effect (despite the finding of Dwyer et al. 2014). Alternatively, one might compare words in quiet with words in noise, since speech in noise understanding is the main complaint of people with hearing loss (especially CI users). And the brain mechanisms required under these conditions might be different to what is a less challenging task. Perhaps the authors wanted to start in the most "trivial" case and test more conditions later.

As we note above, speech in quiet is still degraded from the perspective of the auditory system of a CI user. The brain responses to degraded speech are not uniform, but depend (among other things) on the overall challenge and intelligibility. We focused on speech in quiet to (a) demonstrate differences in processing for listeners with CIs “even” in quiet, and (b) to maximize intelligibility, which in our view helps interpretation (i.e., we answer the question: what are listeners doing to understand speech when they are being successful?).

Figure 4. The difference t-map has a somewhat loose threshold. Presumably this is because the differences in the frontal cortex would be lost. The authors have previously used more stringent thresholding, but perhaps there is justification here.

Because we had rather specific anatomically-constrained hypotheses, we constructed the study first and foremost around an ROI-based analysis. However, we also thought it was important to show non-ROI results. Hence, in Figure 2 we include unthresholded maps of parameter estimates, and thresholded (but not corrected) t maps. We do this to give a sense of the extent of the data and to be transparent about what is, and isn’t showing up at the whole-brain level. It is correct that the result in PFC does not survive correcting for multiple comparisons at the whole brain level.

The CI>Control t-map shows the strongest differences in the visual/visual association cortex, auditory cortex. On the left it appears that a relevant frontal activation occurs in the inferior frontal gyrus, as might be predicted from the literature.

Reduced activation in the left auditory cortex between CI and control could still be due to threshold differences ~5dB (may be significant). Peripheral fidelity is much lower in the CI group, with also poor contribution ipsilaterally.

An additional reason we provide the whole-brain maps is to help readers think about additional hypotheses and guide future research. We agree that activity in IFG is also of interest.

The authors propose that lower left auditory cortex activation could also be related to lower speech perception scores, however these scores are drawn from the AzBio test (sentences in noise) which has the potential for indexing very different processes than a word in the quiet test as was the case in this study.

We used the AzBio sentences, but these were presented in quiet to get a measure of speech-in-quiet understanding. We have clarified this in the revised manuscript.

2. ROI definition:

Choosing an ROI functionally defined from a spatial WM task to look for overlap in brain regions sharing a putative cognitive process is unclear if the comparison is a verbal task. If the goal was to find common cognitive resources the referenced paper by Federenko and colleagues achieved such a goal, providing an ROI with which to test against a common set of cognitive processes. Further, the claim that the task activates a brain region, DLFPC, which does seem to include dorsal aspects of inferior frontal gyrus on pars triangularis. Yet the activation found include this region in this study, while this region of pars triangularis does not appear in the work by Federenko and colleagues leave the reader confused about why these differences would exist, unless it was for other reasons, like localization differences between HD-DOT and fMRI.

The results might change considerably if anatomical ROIs were used instead, i.e., IFG or a DLPFC ROI.

It is fine to use the logic that one task (A) primarily indexes cognitive function X (hopefully) and let's see if another task (B) also shares some of that presumed function by masking B with A. That shows (hopefully) that the functions are shared across tasks. In this case, the function of (spatial) working memory is shared between a spatial working memory and speech task in CI users.

DLPFC is an anatomical label, not a functional one. In this case, the authors have used a functional ROI and so need to name it as such, e.g., 'WM', while reference to the term DLPFC is not helpful and should be removed from the manuscript (and title) except perhaps in the discussion.

We have updated the term to “prefrontal cortex” because this region is unambiguously in the prefrontal cortex. In our view, it is in the dorsal and lateral aspect of the prefrontal cortex, which motivated our use of DLPFC. However we acknowledge there is no clear consensus on DLPFC boundaries and we don’t find this label critical to our argument.

Line 194. It is slightly confusing to define auditory ROIs from a fMRI resting state paradigm. The referenced paper used a 'word listening' activation for both fMRI and HD-DOT. So are the authors using the acoustic word task rather than rest? And further, if there is data using HD-DOT with a similar paradigm then that might be a more sensible ROI definition?

We appreciate the reviewer's attention to this point. In the referenced paper, both word listening and resting state data were presented. We made Author response image 2 to show that there is a large overlap between the left and right auditory ROIs found using the seed based resting state maps and the auditory maps in response to the hearing words task. The dice coefficient between our proposed resting state based ROI and the hearing words based ROI (thresholded at p = 0.05) is 0.7 and the voxelwise similarity of the two is also 0.7. We chose the resting state based ROIs based on fMRI data, because it is independent of any particular task, although we see the alternative logic of using the same task, as well. In practice regions defined both ways are similar, and do not affect our conclusions.

Author response image 2

Line 195-199. The procedure for the definition of the ROI is not clearly motivated, in particular, because the data are obtained from a previous publication which displays thresholded values (at least for the auditory task) that clearly defines the auditory cortex. A correlational approach might (appropriately) be very broad and include areas outside of 'canonical' auditory temporal cortex. Was this intended? What was the threshold for the correlation?

It is not uncommon for speech-related activity to extend beyond the superior temporal plane, through STG and into STS (and sometimes MTG). Our aim was to define an ROI that covered this span of auditory responses and corresponded to our prior work using this particular DOT system (see, e.g., Eggebrecht et al. 2014). We found the ROIs defined this way showed good general correspondence to the activation in our speech tasks, which was our specific use case. For thresholding the ROIs we used a threshold of z(r) = 0.1.

Line 240-245. This text implies the ROI was defined with the current data set rather than another as written in the methods. Please clarify.

Thanks for bringing this ambiguity to our attention. To make this point clear, we consolidated “Region of interest analysis” in Methods and “Functionally defined ROIs” in Results to ““Functionally defined regions of interest”” in Methods.

There, we clarified that:

“To accurately localize the elevated PFC activation in the CI group, we collected HD-DOT data from 9 subjects (4 controls and 5 CI users in 13 sessions) using a spatial working memory task (Fedorenko et al., 2012)” …. “To define the left and right auditory ROIs, we used a previously published fMRI resting state dataset (Sherafati et al., 2020)”

Figure 3A. The 'DLPFC' ROI seems to have a large proportion of the cluster within pars triangularis, which is not quite the same as the fMRI overlap representation in Figure 2 of Federnko et al. 2013, in which the bulk of activation is within the middle frontal gyrus.

One strength of using a functional localizer is that our inferences are most strongly driven by the overlap in activity, rather than putative anatomy. Though, we also note that our ROI is centered on IFS, which is often the center of overlap in multiple-demand network studies. However, there certainly could be some anatomical overlap with portions of the IFG.

3. HD-DOT sensitivity:

As stated in the referenced White and Culver paper: "The high-density array has an effective resolution of about 1.3 cm, which should prevent identification to the wrong gyrus."

This means that localization errors to middle frontal gyrus, as opposed to inferior frontal gyrus, could be a concern in this case in which the activation appears to be across a sulcal boundary.

Our interpretation of the prefrontal activity in CI users is not based primarily on the anatomical landmark, but by the functional definition of the region (which we interpret in an anatomical context).

Figure 1. The sensitivity profile appears to show that greater sensitivity exists for the superior aspect of inferior frontal gyrus and inferior aspect of the middle frontal gyrus (and visual cortex). Is that a simple misinterpretation from the figure? Or do the authors know of issues with respect to the HD-DOT technique weighting responses to the regions of greater sensitivity?

The variation in sensitivity profile proceeds from the arrangement of the sources and detectors. Near the edge of the field of view, sensitivity begins to drop. With our current cap configuration, this does result in small sensitivity differences (Appendix 2– figure 1B). However, we apply a more stringent field of view mask which was created using the intersection of sensitivity across multiple participants to all our data to exclude any voxels with a sensitivity less than around 10% of the maximum sensitivity (Figure 5C). This avoids including any voxels that might be less reliable due to inhomogeneous sensitivity.

Figure 1 supplement. Is it possible there may be some concern about the drop in signal on the right due to variability in the location of the CI transducer across participants? Is it possible the drop in signal in right middle/superior temporal and inferior frontal is also due to loss of measurement here across the group of CI users?

Thank you for raising this important point, which is something we’ve worried about a lot. Although we cannot rule out the effects of lower sensitivity in the magnitude of the activation in certain areas, we were fairly confident that the ROIs we explored are not overlapping with low sensitivity areas.

To quantify the expected change in signal resulting from the CI hardware, in addition to our simulation of two levels of optode blocking (moderate case 4 sources and 4 detectors, and extreme case 6 sources and 6 detectors, now in Appendix 2 – figure 1), we added another simulation study (Figure 3—figure supplements 1-3). Briefly, for each CI user, we calculated how many sources and detectors were affected by the hardware. We then excluded this number of sources/detectors from a control participant. We repeated this process 100 times with different random assignments of excluded channels to the control data to get a sense of what the time course would be for control data (but missing the same number of channels as our CI data). We found a small drop in auditory activation that was still significantly above what we observed in the CI listeners. We note that we cannot completely rule out the effects of CI hardware on signals in the right temporal and parietal regions; however, we do not think it can fully account for our findings in the right auditory cortex. (Because we only studied CI users with a right unilateral implant, regions in the left hemisphere are unaffected.)

Specific comments:

1. Lines 56-57: Please clarify. Are you saying that listening to degraded input reduces the ability for certain other processing, for instance, performance on a second task that involves memory encoding? Or is it that memory encoding on the speech is limited, or both? It is not entirely clear from the text.

When speech is degraded, listeners have trouble remembering it. For example, in Ward et al. (2016) we played normal-hearing participants short stories that were either normal speech, or degraded (but intelligible) due to noise vocoding (16 or 20 channels). We asked participants to repeat back as much of the stories as they could, and found poorer memory for the degraded stories (despite an intelligibility test indicating the stories were intelligible). Perhaps even more compelling evidence comes from studies of word list memory, in which not only the degraded word, but subsequent words are remembered more poorly. Because the subsequent words are not degraded, it is difficult to come up with an acoustic explanation for the memory difficulty—thus pointing towards some shared cognitive resource needed to understand degraded speech, that is thus less available for memory encoding.

2. What kind of CI device? Since it is not specified, I assume this is a standard length implant and not one for hearing preservation.

11 used Cochlear, 6 used Advanced Bionics, 3 used Med EI. We have added this information to the methods.

3. What was the distribution of CI duration? Is it possible there may be some people under 1 year of use? If so this may be difficult with respect to interpretation for those users, since their performance with the device may not be their best, i.e., they may still yet improve. This process of adaptation to the implant could easily involve different brain regions other than measured in this study, or a different weighting depending on adaptation.

We agreed the duration of use is important to consider. We only included CI users if they had at least 1 year of experience with their implant. The years of experience ranged between 1-27.

Author response image 3

We performed some exploratory analyses using this but did not see anything compelling. In general we don’t think this study is adequately powered for individual difference analysis, but experience is certainly a variable we are keeping squarely in mind for future studies.

4. Similarly, does the range of unaided hearing in the CI group get to the level of some "functional" residual hearing, or does the skew go to higher thresholds? If there are people with PTAs of ~60 they may perform differently than people without functional residual hearing? Perhaps residual hearing could be added as a covariate to the analyses.

There may also be clues based on performance on the AzBio task. Is there a relationship between residual hearing and AzBio score? The thrust of these questions is that residual hearing likely has an impact on performance, which similarly may influence brain mechanisms.

We agree regarding the importance of residual hearing and its variability.

We found no significant correlation between the speech perception score (average AzBio score in quiet) and the residual hearing, either defined as left ear 4fPTA unaided, or left ear 4fPTA (aided).

Author response image 4

5. Figure 1. The Spatial WM task appears to have multiple components within it that would be beyond just WM, for example, reward (from the feedback), decision making, response execution, which would all be tied into the level of temporal resolution of the imaging technique. This might temper conclusions about WM per se.

We agree the spatial working memory task encompasses multiple processes. In interpreting this task, and the areas involved, we are relying on the fairly extensive multiple-demand literature tied to these regions, and the work of Fedorenko and colleagues showing that—in single-subject applications—regions activated by this task dissociate from those activated by language tasks. But, also, we have tried to remain somewhat agnostic about the specific cognitive processes engaged. We have made some educated guesses based on the broader literature, but we agree that our own current data have to be understood in terms of what we actually tested.

6. Visual activation appears only on the left, which is unclear (compare to Figure 3A). Was the display directly in-front of the participant, or something about their vision?

The display was directly in front of the participant. The crosshair at the center of the monitor has been aligned to the middle point between the eyes of the participant. We do see a stronger visual activation on the left side, however, we did not collect any vision data and just have verbal confirmation of the participant’s normal vision.

However, Author response image 5 is the same spatial working memory activation (same map depicted in Figure 1C) with a lower t threshold (p = 0.05, t=1.6) showing some activation on the right side which did not survive the more stringent thresholding. Thus we are reluctant to make too much of this qualitative asymmetry.

Author response image 5

7. Figure 1. In the spatial working memory task the wait time range of 4000-250ms is unclear.

We apologize for the confusion. The time is 3750 ms (i.e., 4000 – 250), further subtracting the response time. In this way, the “response + feedback + fixation” portion of each trial is kept to 4000 ms on each trial. We modified the figure text to clarify this.

8. Experimental design. Please provide more information about the following: presentation level for acoustic stimuli, types of words spoken, e.g., monosyllabic, nouns, single talker, gender, etc.

Now added to the revised manuscript.

9. Line 149. Confusing text: "…regressed from the all measurement…"

Thank you for catching this, now repaired.

10. Line 176. Was the 'easy+hard' contrast used to provide better SNR?

That is correct.

Alternatively, why not use the ROI defined previously? Did they vary, because if they did that is instructive, perhaps as implied by aging effects?

If you mean the ROI defined in past studies using fMRI, we had three reasons for using our own ROI. First, we are moving towards subject-specific localizers, and as such, need to collect these data in all of our own participants. Second, our participants indeed differ in age from most prior reports with this task, and this may affect the extent of activation. Third, given the intricate anatomy of the prefrontal cortex we wanted to use a within-modality definition for this ROI.

11. Line 191. Which group of subjects? Is this not the study group as implied on line 176?

As mentioned in lines 359-362, we did not plan on collecting data for the spatial working memory task in the beginning of this study. “To better understand the left PFC activity we observed in our first few CI users, we adopted a spatial working memory task introduced in previous studies (Fedorenko et al., 2011; Fedorenko et al., 2013) in the remaining subjects to aid in functionally localizing domain-general regions of prefrontal cortex.” Since in the literature the variability across individuals has been reported for this task, we aimed at defining an ROI based on at least a subset of our own demographic cohort of interest. Second, we wanted to use the same imaging modality to localize this activation.

12. Line 260. Why does the supplemental analysis not include a contrast of control > CI?

Added.

13. Line 304-305. Incomplete sentence.

Thank you for catching this, now fixed.

14. Line 359. The cited literature of similar findings in normal listeners, mostly highlights the inferior frontal gyrus, however, middle frontal gyrus is included in Defenderfer et al.

We categorize this activity as “prefrontal cortex” to cover both IFG and MFG here.

Reviewer #3 (Recommendations for the authors):

Introduction

Line 81: Why predict effects in DLPFC only? There could be a range of 'non-core' brain regions that support speech perception. Were the effects lateralized to left DLPFC predicted in advance of the study?

Indeed, as expanded upon in greater detail in Peelle (2018) and elsewhere, we agree there are many “non-core” regions that support speech perception in a variety of contexts. Briefly, our focus on DLPFC is motivated by (a) many fMRI studies, including from our group (not all published), showing DLPFC activity for degraded speech; (b) the work of other groups in the frameworks of cognitive control, multiple-demand, or frontoparietal attention networks implicating DLPFC in a variety of demanding tasks; (c) behavioral evidence for domain-general demands when processing acoustically-degraded speech, which we view as consistent with a role for DLPFC (as frequently characterized).

The lateralized DLPFC activity was predicted prior to starting data collection.

Methods

Line 91: How were the sample sizes of the CI group and the control group determined? I note that page 7 explains that 15-20 subjects per group was thought to be sufficient, based on similar studies. Did these previous studies include a cross-sectional design?

In these referenced studies (Defenderfer et al. 2021; Pollonini et al. 2014; Zhou et al. 2018b), the normal hearing cohort in an effortful listening/unintelligible speech situation was considered to be representing the CI user cohort.

Page 4. Table 1 reports left ear 4fPTA (unaided). The footnote for Table 1 reports that 8/20 CI users had a hearing aid but did they use them during the speech task?

Yes, all listeners performed the test in their everyday listening configuration.

Line 120. Where were the 2 speakers located relative to the participants? +/- 45 degrees? Which sound level were the words presented at?

We have added this to the text.

Lines 127-130. More details about the spoken word task is needed. Is this a standardized test or a novel test that the authors developed themselves?

This task was first introduced by Peterson et al. 1988 (Petersen, S. E., Fox, P. T., Posner, M. I., Mintun, M. and Raichle, M. E. Positron emission tomographic studies of the cortical anatomy of single-word processing. Nature 331, 585–589 (1988)), It was later replicated using HD-DOT by Eggebrecht et al. 2014, Sherafati et al. 2020, Fishel et al., 2020, in multiple studies to correctly localize the auditory cortex activation using both fMRI and HD-DOT.

Line 131. Which 'preliminary results'?

Clarified and changed to “To better understand the left PFC activity we observed in our first few CI users, we adopted a spatial working memory task introduced in previous studies”.

Line 191. "…based on the response … in a group of subjects…". Which subjects? Merge with relevant information on page 8.

Implemented.

Line 216. Is "Pollonini et al. 2014" cited correctly?

Yes.

Line 238. Why define only left DLPFC as an ROI?

In our experience with fMRI studies of acoustically-challenging speech, frontal activity seems qualitatively stronger and/or more common in the left hemisphere (see e.g., Davis and Johnsrude, 2003, Journal of Neuroscience).

Page 9. The visual spatial working memory task also activated 'auditory' regions in the right hemisphere that look like they overlap with the 'right auditory ROI' created by the authors. Do the authors consider this overlap to be problematic?

Our focus is on the left PFC region involved in spatial working memory. The observation of left temporal lobe activity is somewhat puzzling but does not affect our ROI definition.

Page 9. Legend for Figure 3. "…in the DLPFC region, survived after…". A word is missing.

Now fixed.

Page 10. Figure 4C suggests that CI > controls contrast identified quite a few group differences. So why then focus on left DLPFC?

Based on prior literature and our work in the area we had a specific hypothesis regarding left DLPFC. We certainly allow that there may be other areas involved in speech processing for CI users not engaged in controls, but that was not the focus of the current paper.

Page 11. Results of two-sample t-tests are reported in the legend of Figure 5. This information should be taken out of the legend and reported in full. Why were t-tests more appropriate than an ANOVA? Please report results of statistical analyses fully i.e. not just p-values. Were reported p-vals corrected for the number of t-tests that were used?

We have ensured the t statistics and degrees of freedom are in the revised manuscript. The ROI results in Figure 3 (main results) are Bonferroni corrected for multiple comparisons across the three ROIs tested, now also clarified in the text.

Page 11. The authors report the results of correlation analyses. Please add further details e.g. are Pearson's correlation coefficients reported?

Yes, Pearson correlations—now specified.

Page 11. The authors report that the threshold used for statistical significance in the correlation analyses was 'uncorrected'. Presumably this means that reported p-vals were not corrected for the 4 correlation analyses shown?

Correct.

Page 12. Correlation analyses for CI users. The rationale for these analyses is missing. Why carry out simple correlation analyses on data for CI users only? Potential predictors of speech processing (4PTA) could have been used as a covariate in the main analyses (e.g. Figure 5).

The CI listeners were our primary group of interest, and we also assumed they would show more variability in these covariate measures (supported by the data). We completely agree that some of these covariates might also be informative in the control listeners, but we worried there might be a confound (if controls and CI users differ both in group membership and 4PTA, for example, it would cause problems with interpretation). So, we opted to restrict these exploratory analyses to the patient group of interest.

https://doi.org/10.7554/eLife.75323.sa2

Article and author information

Author details

  1. Arefeh Sherafati

    Department of Radiology, Washington University in St. Louis, St. Louis, United States
    Contribution
    Conceptualization, Data curation, Formal analysis, Methodology, Software, Validation, Visualization, Writing - original draft, Writing – review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0003-2543-0851
  2. Noel Dwyer

    Department of Otolaryngology, Washington University in St. Louis, St. Louis, United States
    Contribution
    Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Project administration, Visualization, Writing – review and editing
    Competing interests
    No competing interests declared
  3. Aahana Bajracharya

    Department of Otolaryngology, Washington University in St. Louis, St. Louis, United States
    Contribution
    Investigation, Methodology, Software, Visualization, Writing – review and editing
    Competing interests
    No competing interests declared
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0002-7361-6020
  4. Mahlega Samira Hassanpour

    Moran Eye Center, University of Utah, Salt Lake City, United States
    Contribution
    Conceptualization, Investigation, Software, Writing – review and editing
    Competing interests
    No competing interests declared
  5. Adam T Eggebrecht

    1. Department of Radiology, Washington University in St. Louis, St. Louis, United States
    2. Department of Electrical & Systems Engineering, Washington University in St. Louis, St. Louis, United States
    3. Department of Biomedical Engineering, Washington University in St. Louis, St. Louis, United States
    4. Division of Biology and Biomedical Sciences, Washington University in St. Louis, St. Louis, United States
    Contribution
    Conceptualization, Investigation, Methodology, Software, Writing – review and editing
    Competing interests
    No competing interests declared
  6. Jill B Firszt

    Department of Otolaryngology, Washington University in St. Louis, St. Louis, United States
    Contribution
    Conceptualization, Formal analysis, Investigation, Methodology, Resources, Supervision, Writing – review and editing
    Competing interests
    No competing interests declared
  7. Joseph P Culver

    1. Department of Radiology, Washington University in St. Louis, St. Louis, United States
    2. Department of Biomedical Engineering, Washington University in St. Louis, St. Louis, United States
    3. Division of Biology and Biomedical Sciences, Washington University in St. Louis, St. Louis, United States
    4. Department of Physics, Washington University in St. Louis, St. Louis, United States
    Contribution
    Conceptualization, Formal analysis, Funding acquisition, Investigation, Methodology, Resources, Supervision, Writing – review and editing
    Competing interests
    No competing interests declared
  8. Jonathan E Peelle

    Department of Otolaryngology, Washington University in St. Louis, St. Louis, United States
    Contribution
    Conceptualization, Formal analysis, Funding acquisition, Investigation, Methodology, Resources, Supervision, Writing – review and editing
    For correspondence
    j.peelle@northeastern.edu
    Competing interests
    Reviewing editor, eLife
    ORCID icon "This ORCID iD identifies the author of this article:" 0000-0001-9194-854X

Funding

National Institutes of Health (R21DC015884)

  • Jonathan E Peelle

National Institutes of Health (R21DC016086)

  • Jonathan E Peelle
  • Joseph P Culver

National Institutes of Health (K01MH103594)

  • Adam T Eggebrecht

National Institutes of Health (R21MH109775)

  • Adam T Eggebrecht

National Institutes of Health (R01NS090874)

  • Joseph P Culver

National Institutes of Health (R01NS109487)

  • Joseph P Culver

National Institutes of Health (R01DC019507)

  • Jonathan E Peelle

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Acknowledgements

AS would like to thank Abraham Z Snyder, Andrew K Fishell, Kalyan Tripathy, Karla M Bergonzi, Zachary E Markow, Tracy M Burns-Yocum, Mariel M Schroeder, Monalisa Munsi, Emily Miller, Timothy Holden, and Sarah McConkey for helpful discussions. We also want to thank our participants for their time and interest in our study.

Ethics

Human subjects: All subjects were native speakers of English with no self-reported history of neurological or psychiatric disorders. All aspects of these studies were approved by the Human Research Protection Office (HRPO) of the Washington University School of Medicine. Subjects were recruited from the Washington University campus and the surrounding community (IRB 201101896, IRB 201709126). All subjects gave informed consent and were compensated for their participation in accordance with institutional and national guidelines.

Senior Editor

  1. Barbara G Shinn-Cunningham, Carnegie Mellon University, United States

Reviewing Editor

  1. Timothy D Griffiths, University of Newcastle, United Kingdom

Reviewers

  1. Timothy D Griffiths, University of Newcastle, United Kingdom
  2. Phillip E Gander, University of Iowa, United States

Publication history

  1. Preprint posted: October 16, 2021 (view preprint)
  2. Received: November 5, 2021
  3. Accepted: June 4, 2022
  4. Accepted Manuscript published: June 6, 2022 (version 1)
  5. Version of Record published: June 23, 2022 (version 2)

Copyright

© 2022, Sherafati et al.

This article is distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use and redistribution provided that the original author and source are credited.

Metrics

  • 822
    Page views
  • 249
    Downloads
  • 0
    Citations

Article citation count generated by polling the highest count across the following sources: Crossref, PubMed Central, Scopus.

Download links

A two-part list of links to download the article, or parts of the article, in various formats.

Downloads (link to download the article as PDF)

Open citations (links to open the citations from this article in various online reference manager services)

Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)

  1. Arefeh Sherafati
  2. Noel Dwyer
  3. Aahana Bajracharya
  4. Mahlega Samira Hassanpour
  5. Adam T Eggebrecht
  6. Jill B Firszt
  7. Joseph P Culver
  8. Jonathan E Peelle
(2022)
Prefrontal cortex supports speech perception in listeners with cochlear implants
eLife 11:e75323.
https://doi.org/10.7554/eLife.75323
  1. Further reading

Further reading

    1. Neuroscience
    Maria Cecilia Martinez, Camila Lidia Zold ... Mariano Andrés Belluscio
    Research Article

    The automatic initiation of actions can be highly functional. But occasionally these actions cannot be withheld and are released at inappropriate times, impulsively. Striatal activity has been shown to participate in the timing of action sequence initiation and it has been linked to impulsivity. Using a self-initiated task, we trained adult male rats to withhold a rewarded action sequence until a waiting time interval has elapsed. By analyzing neuronal activity we show that the striatal response preceding the initiation of the learned sequence is strongly modulated by the time subjects wait before eliciting the sequence. Interestingly, the modulation is steeper in adolescent rats, which show a strong prevalence of impulsive responses compared to adults. We hypothesize this anticipatory striatal activity reflects the animals’ subjective reward expectation, based on the elapsed waiting time, while the steeper waiting modulation in adolescence reflects age-related differences in temporal discounting, internal urgency states, or explore–exploit balance.

    1. Computational and Systems Biology
    2. Neuroscience
    Sergio Oscar Verduzco-Flores, Erik De Schutter
    Research Article Updated

    How dynamic interactions between nervous system regions in mammals performs online motor control remains an unsolved problem. In this paper, we show that feedback control is a simple, yet powerful way to understand the neural dynamics of sensorimotor control. We make our case using a minimal model comprising spinal cord, sensory and motor cortex, coupled by long connections that are plastic. It succeeds in learning how to perform reaching movements of a planar arm with 6 muscles in several directions from scratch. The model satisfies biological plausibility constraints, like neural implementation, transmission delays, local synaptic learning and continuous online learning. Using differential Hebbian plasticity the model can go from motor babbling to reaching arbitrary targets in less than 10 min of in silico time. Moreover, independently of the learning mechanism, properly configured feedback control has many emergent properties: neural populations in motor cortex show directional tuning and oscillatory dynamics, the spinal cord creates convergent force fields that add linearly, and movements are ataxic (as in a motor system without a cerebellum).