Longitudinal Effect of Deactivating Stimulation Sites Based on Low-rate Thresholds on Speech Recognition in Cochlear Implant Users

Abstract

Objective:

The objective of the current study was to examine the longitudinal effect of deactivating stimulation sites estimated to produce broad neural excitation on speech recognition.

Design:

Spatial patterns of neural excitation were estimated based on a previously established psychophysical measure, i.e., detection threshold for low-rate pulse trains. Stimulation sites with relatively poor thresholds were deactivated in an experimental map. The acute effect was evaluated, in quiet and in noise, immediately after the experimental map was created (baseline), after the subjects practiced with the experimental map for two months (treatment), and after the subjects’ daily map was switched back again to the clinical map for another two months (withdrawal).

Study sample:

Eight Cochlear Nucleus device users participated in the study.

Results:

For both listening in noise and in quiet, the greatest effect of deactivation was observed after the subjects were given time to adapt to the new frequency allocations. The effect was comparable for listening in fluctuating and steady-state noises. All subjects benefited from deactivation for listening in noise, but subjects with greater variability in thresholds were more likely to benefit from deactivation for listening in quiet.

Conclusions:

The benefit of electrode deactivation for speech recognition can increase with practice.

Introduction

Channel interaction is thought to be one of the biggest challenges that limit the speech recognition outcomes with cochlear implants (CIs). Channels can interact for a number of reasons, such as abnormally high impedance, severe nerve degeneration, patchy neural survival patterns along the tonotopic axis, and/or large electrode-neuron distances. The frequency information assigned to a given electrode may be coded, due to channel interaction, at places broader than intended. Channel interaction is thought to have a negative impact on speech recognition. If information is smeared across tonotopic regions, the spectral resolution necessary for discriminating spectral modulation patterns might be insufficient. Previous studies have demonstrated that low spectral resolution within an implanted ear, as assessed by spectral ripple discrimination and electrode discrimination, is associated with poorer speech recognition performance (Holden et al, 2016; Zhou, 2017; Won et al, 2007; Henry et al, 2000).

Several approaches have been used to identify locations that might produce excessive channel interactions. Psychophysical studies have used masking patterns (Boex et al, 2003; Chatterjee & Shannon, 1998; Dingemanse et al, 2006; Kwon & van den Honert, 2006; McKay, 2012; Nelson et al, 2011), detection thresholds in focused stimulation (Bierer & Faulkner, 2010), and the amount of loudness summation (Bierer & Faulkner, 2010) to estimate the degree to which a stimulation site interacts with its neighboring locations. Tripolar detection thresholds have been shown to correlate with the psychophysical spatial tuning curves, suggesting that higher current levels required to achieve threshold in a focused current field predict broader neural excitation thus possible channel interaction (Bierer & Faulkner, 2010). Long et al. (2014) indicated that the higher current required for detection was partly attributed to greater distance to inner wall of the cochlea at those sites. The amount of loudness summation from activating multiple electrodes is also an indication of how likely the electrodes stimulate overlapping neural populations (McKay et al, 2003; Tang et al, 2011). The idea is that loudness would increase if multiple electrodes that are activated simultaneously stimulate independent nerve fibers. A more direct measure of channel interactions, at least those that are due to current spread, was one that was based on the imaging techniques. CT-based 3D reconstructions of implanted cochlea can reveal the electrodes located at the lateral wall positions of the scala tympani and thus are likely to produce a wide current field (Labadie et al, 2016). Recent evidence indicates that broad neural activation patterns might not only reduce spectral resolution, but also reduce sensitivity to amplitude modulations. Zhou et al. (2018) reported that poor amplitude modulation detection thresholds tended to be observed at stimulation sites measured with flat spatial tuning curves. Two approaches have been used in the past to reduce channel interactions within cochlear implants, which involve either deactivating electrodes estimated to produce unusually broad activation or adjusting electrode configuration to confine the activating current to a small field (e.g., Labadie et al, 2016; Berenstein, et al, 2008). Channel deactivation based on psychophysical or imaging measures has resulted in better speech recognition compared to using all functional electrodes, especially in noise (Labadie et al, 2016; Zhou, 2017). Applying focused electrode configuration in attempts to achieve more spatially selective stimulation, however, did not always lead to improved speech recognition performance (Berenstein et al, 2008; Bierer & Litvak, 2016; Srinivasan et al, 2013).

In searching for a clinically applicable time-efficient channel interaction measure that could be used in all implant devices, Zhou (2016; 2017) identified monopolar psychophysical detection thresholds for low-rate stimuli. Low-rate monopolar thresholds were strongly correlated with the slopes of the psychophysical masking patterns, with elevated low-rate thresholds predicting broader estimated spatial spread of neural excitation. The results of Zhou (2016) also showed that the strength of the correlation between the monopolar thresholds and estimated spread of neural excitation systematically weakened as the stimulus used for measuring detection was higher in pulse rate and/or longer in duration. The results were attributed to the possible facilitative effect that broad spatial activation of nerve fibers might have on detecting high-rate or long-duration pulse trains. Broad stimulation of the cochlea may lower the average firing probability of the activated neurons, making it more likely for the neurons to respond to an increase in stimulation density (e.g., an increase in rate) (Zhou & Dong, 2017; Zhou & Pfingst, 2016). In fact, in many cases, high-rate stimuli were detected with a much lower current level at broadly stimulated places than those that were relatively sharply tuned, rendering high-rate detection thresholds insensitive to spatial activation patterns in the nerve fibers (Zhou, 2016). In contrast to the typical small across-site variation in thresholds using monopolar configuration at a clinical default rate (900 pulses per second/electrode), low-rate thresholds were also found to be significantly more variable across the array (Zhou, 2016).

Based on the low-rate thresholds, deactivation strategies were performed to exclude five sites with poor low-rate thresholds, in a manner that they were not immediately next to each other to avoid large new frequency-place mismatch (Zhou, 2016, 2017). These studies showed that subjects’ performance significantly improved after the deactivation relative to using the whole array. This benefit was more consistently found for listening in noise than for listening in quiet (Zhou, 2016). Random deactivation with the same number of electrodes using the same rule often resulted in poorer performance relative to that using the clinical full-array map (Zhou, 2016). In cases where random deactivation did result in an improvement, the magnitude of improvement was smaller than deactivation based on the low-rate thresholds (Zhou, 2016). These results suggested that subjects with excessive channel interactions across the whole array may benefit from random deactivation, but the benefit would not be maximized unless deactivation targeted the places that produced the greatest spread. These results were also in line with previous findings, which showed that turning off every other electrode on the array (either odd- or even-numbered electrodes) did not provide consistent benefits across subjects (Mani et al, 2004). Further, Zhou (2017) showed that deactivation of the stimulation sites with poor low-rate thresholds resulted in improved spectral ripple discrimination thresholds, and more importantly, the magnitude of improvement in the spectral ripple discrimination thresholds predicted the magnitude of improvement in the speech reception thresholds (SRTs). These results suggested that using fewer but more independent channels may produce better outcomes than using the whole array.

The aim of the current study was to examine the longitudinal effect of deactivating electrodes with poor low-rate thresholds. By examining the longitudinal effect of deactivation, we aimed to answer the question of whether the observed acute benefit would be increased by providing subjects with time to adapt to the new frequency allocation. An A-B-A design was used, where the baseline performance with the clinical map was obtained, followed by a two-month practice with the experimental map with deactivated electrodes (treatment), and then back to the subject’s clinical map again for another two months (treatment withdrawal). At all three visits, performances with the clinical and the experimental maps were both measured so that the effect of deactivation could be quantified at any given time point. Comparing results of the two maps at the same time point makes it possible to quantify whether the relative performance of the two maps changes over time, thus evaluating whether there is an effect of adaptation. It also ensures that any internal noise from the subject at a given visit (i.e., fatigue or health issues) was controlled for. As mentioned earlier, deactivation of the interacting electrodes not only improved speech recognition but also resulted in better spectral ripple discrimination. The second focus of the current study was to examine whether the improved spectral resolution was sufficient to offer release of masking from fluctuating noises. In order to test the hypothesis, the effect of deactivation on speech recognition was evaluated in three noise types: a steady-state noise (SSN), an amplitude modulated (AM) noise and a female-talker interferer (FI).

Methods

Subjects and Hardware

Eight subjects (9 ears) who were implanted with the Cochlear Nucleus® (Cochlear Corporation, Englewood, CO) devices participated in the study. The subjects were not selected based on their performance in our earlier studies but because they were able to pay multiple visits to the lab at the time intervals the study required. One subject was bilaterally implanted (S10). For this subject, deactivation was applied in both implants and the ears were evaluated independently.

The demographic characteristics of the subjects are shown in Table 1. All subjects used the ACE speech processing strategy and a TSPL set at 25 dB SPL. Subjects all gave written informed consent before participating in the study. The use of human subjects in this study was approved by the Institutional Review Board at East Carolina University.

Table 1.

Subject demographics.

Subject number	Gender	Age (yrs)	Duration of implant use (yrs)	Duration of deafness (yrs)	Etiology	Device type	Processor
S4L	F	57.6	5.2	5.2	Hereditary	CI24RE (CA)	CP810
S6R	F	84.9	4.7	69.9	Hereditary	CI24RE (CA)	CP900
S7R	F	71.1	5.9	33.7	Unknown	CI24RE (CA)	CP810
S10L	F	66.5	15.6	16.3	Hereditary	CI24RE (CS)	CP900
S10R	F	66.5	15.6	16.3	Hereditary	CI24RE (CA)	CP900
S14L	M	81.7	6.0	20.7	Unknown	CI512	CP910
S18L	F	66.1	3.3	3.6	Hereditary	CI422	CP910
S19L	F	71.2	10.6	4.3	Unknown	CI24RE(CA)	CP1000
S22R	F	73.1	5.5	0.4	Unknown	CI24RE(CA)	CP920

Psychophysical Tests and Programming

Psychophysical testing was carried out using a freedom processor controlled by the NIC II research interface and MATLAB. Psychophysical detection thresholds were measured using biphasic pulse trains with an interphase gap of 8 μs, phase duration of 25 μs, stimulation rate of 80 pulses per second (pps) and stimulus duration of 250 ms. Thresholds were measured in random order across the electrode array using the method of adjustment. For each electrode, the low-rate stimulus was played at a presentation rate of 2 Hz (pulse trains separated by 500 ms) and the subject could listen to the stimulus for as long as he/she needed. The subject was instructed to adjust the current level of the stimulus using buttons on a graphic user interface on the computer screen that allowed level adjustments in increments of 1, 5, or 25 clinical units (CU). The subject was instructed to increase the level until the stimulation produced a comfortably loud percept then decrease and bracket the stimulus above and below threshold to find the lowest level that was just detectable. This routine was to help subjects differentiate tinnitus from the stimulus. The threshold measurements were repeated twice and the average was taken for each electrode. Repeats were conducted if the difference between the two runs was larger than 2 dB and the two closest measurements were averaged. The across-site variation in the 80-pps thresholds was examined and the five sites with the highest thresholds that were not immediately adjacent to each other were identified (See Figure 1). A portion of these data has been published in Zhou (2016). The number of electrodes that were to be deactivated was based on previous studies (e.g., Zhou, 2016; 2017), although it is likely that the ideal number is different for each subject.

Fig. 1.

Across-site variation in the 80-pps thresholds. The grey symbols indicate the stimulation sites that were deactivated. Stimulation sites 1–22 correspond to electrodes from the basal to apical end on the electrode array.

The subjects were encouraged to update their clinical map, if the annual audiology appointment was due within four months, and take the necessary time (at least one month) to adjust to this new clinical map before they started the study. Otherwise, they were asked not to have the map adjusted until the study was completed. An experimental map was created for each subject, where sites with the high 80-pps thresholds were deactivated and frequencies reallocated globally. Note that no two consecutive electrodes could be turned off. The experimental map was identical to the subject’s clinical map in all other settings including speech processing strategies, stimulation rate, and smart sound settings. The map was sent to the subject’s audiologist to be loaded on the subject’s daily-use processor. Clear instructions were given to the audiologist regarding the smart sound settings. Some subjects requested that they keep their clinical map on a backup processor. Others requested that they keep the clinical map on one of the four slots on the daily-use processor. The flexibility was given in case any adverse effects were to occur with the experimental map in a real-life situation and an immediate appointment with the audiologist was not possible. No subjects however used the clinical map during the treatment phase, even though the flexibility was given.

Test Time Intervals

Speech recognition, as detailed below, was measured for the clinical map and the experimental map at three visits. At the first visit (V1), the baseline of the subjects’ performance with the clinical map was established and the acute effect of the experimental map was measured. The experimental map was then loaded and activated on the same day for some subjects. For others, the subjects informed us the start date of the experimental map after they returned home to visit their audiologist. Since the activation of the experimental map, the subjects had used the experimental map only, as instructed, for two months. By the end of the two-month period, the subjects visited the lab the second time (V2) and speech recognition was once again measured with the two maps. In the treatment withdrawal phase, the subjects switched back to their clinical map and were asked to use that map only for another two months. Two subjects contacted the lab during this withdrawal phase and informed us that they used the experimental map once for a short period of time (~ 10 mins and 25 mins). Due to the fact that these two incidents were very short events, their effect, if any, on treatment withdrawal was considered negligible. By the end of the two-month treatment-withdrawal phase, subjects came back for a third visit (V3), when both maps were evaluated again. Upon the completion of the study, the subjects were free to choose to use any map that they preferred.

Speech Recognition Tests

Speech recognition was evaluated in both quiet and in noise. SRTs were measured using CUNY (City University of New York) sentences. Three types of noises were used: a speech-shaped white noise (SSN) using the long-term spectrum of the CUNY sentences, an AM noise (4 Hz amplitude modulation) with a 100% modulation depth, and a female-talker interferer (fundamental frequency of 253 Hz), which was a voice recording of HINT (Hearing in Noise Test) sentences with the silences removed between sentences. The fundamental frequency of the CUNY male speaker was 130 Hz. The noise was played alone for 1.5 seconds before the sentence, during the sentence, and 0.5 seconds alone after the sentence. The onset and offset of the noise were cosine ramped. The sentences were kept at a level of 65 dB (A) SPL, while the level of the noise was adapted. Signal to noise ratio (SNR) was calculated for the timeframe that the sentence overlapped with the noise. The subjects were instructed to repeat the sentence that they heard. SNR started at 20 dB and adapted with a step size of 2 dB using a one-down one-up rule based on the subject’s response. Performance converged at 50% of the psychometric function (% correct versus SNR) after 12 reversals. SRT was taken as the average of the SNRs at the last 6 reversal points.

Speech recognition in quiet was evaluated using the TIMIT (Texas Instruments and Massachusetts Institute of Technology) sentences (Garofolo et al, 1993), which contain recordings of 630 speakers of eight major dialects of phonetically rich American English. Many of the sentences were semantically incoherent thus providing very limited contextual cues. Because the performance-intensity functions of the TIMIT sentences for cochlear implant listeners are unknown, the sentences were normalized to have equal root mean square values. They were calibrated to be delivered at 65 dB (A) SPL. For each condition, two lists of TIMIT sentences were randomly selected without replacement. Again, the subjects were instructed to repeat back what they heard and were encouraged to make their best guesses when they did not understand every word in the sentence. The number of words correctly identified was used to calculate a percent correct score.

At the first visit, for each subject, the two maps were evaluated in random order. SRTs were measured twice for each noise type under each map condition and the thresholds were averaged. The noise conditions were randomized. At the second and third visits, the map that the subjects used for the past two months was evaluated first followed by the relatively less experienced map within that time period. The reason for that was if any learning of test material or format occurred at those visits, it would have only favored the less experienced map, which was tested the second, to avoid overestimation for performance of the more experienced map. Therefore, the testing order was kept blind to the subject, but could not be kept blind to the tester. There is a possibility that the tester was biased. Testing was always performed with a laboratory-owned processor.

Statistics

Repeated-measures ANOVA was used to examine whether the overall effect of deactivation was significant (main factor of map), whether subjects were learning the test procedure (main factor of visit), and whether performance was better in some noise types than others (main factor of noise, if applicable). The interaction terms were examined to determine whether the effect of map was greater for some noise types than others (map × noise). Most importantly, to determine whether the effect of deactivation increased with the practice of the experimental map and decreased with treatment withdrawal, the interaction term map × visit was examined for a quadratic (increasing then decreasing) trend. To control for Type I errors, p value was adjusted to determine statistical significance.

Results

Figure 1 shows the across-site variation pattern of the 80-pps thresholds for each subject. The stimulation sites selected for deactivation are indicated by grey symbols. Some of these results have been published (Zhou, 2016; 2017). Missing data from S10L and S18L were due to non-functional electrodes identified during clinical mapping.

SRTs measured for the two maps (C: Clinical; E: Experimental) on three visits (V1, V2 and V3) for each individual subject are shown in Figure 2. A three-way repeated-measures ANOVA was used to examine the main effect of visit, map, and noise type on the SRT performance. The main effect of visit was not significant [F (2, 16) = 0.82, p = 0.46], which suggested that the overall performance did not significantly fluctuate across visits. The main effect of noise type was significant [F (2, 16) = 21.17, p < 0.001]. Tests of within-subject contrasts showed that there was a significant linear trend for the effect of noise [F (1, 8) = 28.75, p = 0.001], indicating that the SRTs were progressively worse with the steady-state noise, the amplitude-modulated noise, and the female interferer (Fig. 2). A main effect of map was significant [F (1, 8) = 41.29, p < 0.001]. Overall SRTs measured with the experimental map were lower (better) compared to those with the clinical map. The interaction term map × noise was not significant [F (2, 16) = 0.26, p = 0.77], which suggested that the effect of map was not greater for some noise conditions than others. Results of contrasts also showed a significant quadratic trend for the interaction term map × visit [F (1, 8) = 7.01, p = 0.02]. While the effect of map was significant at each of the three visits [V1: F (1, 8) = 19.30, p = 0.002; V2: F (1, 8) = 77.13, p < 0.001; V3: F (1, 8) = 37.2, p <0.001], the quadratic trend indicated that the effect of map increased at the second visit after acclimatization to the map and decreased at the third visit after treatment withdrawal. The magnitude of benefit provided by the experimental map is shown in Figure 3 for each individual under the various test conditions. Correlational analysis indicated across subjects, the magnitude of SRT improvement averaged across conditions was not related to the extent to which the low-rate thresholds varied across the array (across-site variance) [r = 0.38, p > 0.025].

Fig. 2.

Speech reception thresholds (SRTs). Grey scale of the bars indicates noise conditions (SSN: steady-state noise; AM: amplitude-modulated noise; FI: female interferer). “C” stands for “clinical”, “E” stands for “experimental”, and “V” stands for “visit.” Error bars represent standard deviations. Asterisks indicate statistically significantly better performance in the AM noise relative to the SSN noise.

Fig. 3.

Magnitude of improvement in SRTs as a result of deactivation (performance with the experimental map minus that with the clinical map). Grey scale of the bars indicates noise conditions (SSN: steady-state noise; AM: amplitude-modulated noise; FI: female interferer). Error bars represent standard deviations. “V” stands for “visit.” Asterisks indicate statistically significant benefit.

Given the small sample size, the effects were analyzed again for each individual. A critical difference value was used to determine whether performance under one condition was significantly different than that under another condition using a 95% confidence interval limit (±1.12 dB; 4 lists of 12 sentences for each SRT) (Nilsson et al., 1994). Significant benefit of deactivation is indicated by asterisks shown in Figure 3. Of the 54 comparisons between the clinical map and the experimental map, 47 were statistically significant. For each visit, performance was also compared between the AM and SSN condition to examine whether there was release of masking from fluctuating noises. The conditions under which significant effects were found are indicated by asterisks in Figure 2. It should be noted that, although there is no published data, the confidence interval for the CUNY sentences should be larger than that estimated for the HINT sentences (Nilsson et al., 1994). We therefore acknowledge that the analysis at the individual subject level may have overestimated the effect.

Individual performances of sentence recognition in quiet (TIMITs) are shown in Figure 4. A two-way repeated-measures ANOVA was performed to test the main effect of visit and map on TIMIT sentence recognition in quiet. The main effect of visit was not significant [F (1, 8) = 1.01, p = 0.34], suggesting that performance was stable across visits. The main effect of map was significant [F (1, 8) = 9.80, p = 0.01], indicating better performance with the experimental map than that with the clinical map. The effect of map was significant at visit 2 but not significant at visit 1 and only marginally significant at visit 3 [V1: F (1, 8) = 2.69, p = 0.13; V2: F (1, 8) = 27.49, p = 0.01; V3: F (1, 8) = 4.99, p = 0.05], suggesting that practice with the experimental map helped the subjects benefit from the deactivation. The magnitude of benefit provided by the experimental map is shown in Figure 5 for each individual. Correlational analysis indicated that across subjects, the magnitude of improvement in sentence recognition in quiet was significantly correlated with the subjects’ across-site variance in the low-rate thresholds [r = 0.80, p < 0.025], with greater threshold variation associated with larger benefit of deactivation.

Fig. 4.

Sentence recognition in quiet (TIMITs). Grey scale of the bars indicates map conditions. Error bars represent standard deviations. “V” stands for “visit.”

Fig. 5.

Magnitude of improvement in TIMIT sentence recognition as a result of deactivation (performance with the experimental map minus that with the clinical map). Error bars represent standard deviations. “V” stands for “visit.” Asterisks indicate statistically significant benefit.

The effect of activation was similarly analyzed again for each individual. The 95% confidence interval critical difference values were calculated based on the number of words in the TIMIT sentence lists (Thornton & Raffin, 1978). Note that the critical value was different for different baseline performance (i.e., performance with the clinical map). Significant improvement was only seen for S10R at visit 3 and for S22R at visit 2 (asterisks in Fig. 5).

Discussion

The current study extended previous research that identified benefit for speech recognition outcomes with deactivation of stimulation sites estimated to produce unusually broad neural excitation and examined its longitudinal effect. The current study tested the hypothesis that experience with the experimental map with deactivation would provide subjects with greater benefit compared to that seen in the acute experiments. This was achieved by using an A-B-A design, where the effect was measured immediately after the fitting of the experimental map with deactivation, after two months of adjustment to the experimental map, and finally two months after the subjects’ daily map was switched back to the original one without deactivation. The effect was also evaluated for various noise conditions to examine whether deactivation would help with listening in some noise types more than others.

Performance as a Function of Noise

Results of the present study showed a performance pattern in various noisy listening conditions that was consistent with the literature. That is, the overall performance was better with a steady-state noise (spectrally-shaped white noise) than with fluctuating maskers (Fu & Nogaki, 2005; Qin & Oxenham, 2003; Stickney et al, 2004; Turner et al, 2004; Li & Loizou, 2010; Rader et al, 2013; Zirn et al, 2016). The present data suggested that performance worsened when a steady-state noise was replaced with a noise that was amplitude modulated, and reduced furthermore when speech was presented in a single talker. This pattern was opposite to what is typically observed in normal-hearing listeners where fluctuating noises offer a release of masking relative to steady-state noises. The reversed effect of noise for CI users could be due to the low spectral resolution in the modern implant devices which provides limited pitch cue that is important for segregating signals from competing sound sources (Qin & Oxenham, 2003; Turner et al, 2004). Spectral smearing resulting from channel interaction, or the lack of temporal fine structure information could also hinder CI listeners from taking advantage of the temporal and spectral dips in a fluctuating noise where the signal to noise ratio is favorable (Fu & Nogaki, 2005; Lorenzi et al, 2006). Recently, Oxenham and Kreft (2016) suggested that poor frequency selectivity of CIs may smooth the inherent temporal-envelope fluctuations in a white noise rendering it less effective in masking.

Further, current results showed that the effect of deactivation was not dependent on noise conditions. Overall, relative to the performance using the whole array (clinical map), deactivation resulted in better SRT performance, but the magnitude of the benefit was comparable across noise conditions. The lack of interaction between the main factors of map and noise can also be interpreted as that the relative difference in performance across noise types (i.e., best in SSN followed by AM then FI) did not change as a result of deactivation. Inspecting individual data, there were instances where performance was better with the AM noise than SSN indicating masking release (Fig. 2, asterisks), but this effect was not always just seen with the deactivated map and the effect was not consistent across subjects or visits. Previous results showed that deactivation of the stimulation sites estimated to produce broad excitation enabled the listeners to better discriminate frequency modulation patterns between signal spectra, as indicated by improved spectral ripple discrimination thresholds (Zhou, 2017). However, results of the present study suggested that the possible improvement in spectral resolution was not sufficient to provide consistent release of masking from fluctuating noises across subjects and visits.

Effect of Learning on Deactivation Benefit for SRT

The overall SRT performance, regardless of the map setting, was stable across visits, as revealed by a nonsignificant effect of visit. This means that the subjects were familiar with the test and no learning occurred in terms of the test format or material. Group results indicated that performance with the experimental map on average was better relative to the clinical map at each visit, producing a mean SRT improvement of 3.77, 4.47, and 3.53 dB respectively. The effect of deactivation was re-evaluated for each individual using the critical difference values. The results indicated improvement with deactivation for almost all comparisons between the clinical and experimental maps.

Recall that at visits 1 and 3, the subjects had been using their clinical map for a long period of time before the experimental map was evaluated. The significant acute effect however was not surprising given that a consistent benefit of deactivation without learning was identified for listening in noise in the previous studies (Zhou, 2016; 2017). After the stimulation sites were deactivated, the frequencies were reallocated globally to the remaining active electrodes. The acute benefit could suggest that the detrimental effect of frequency reallocation was outweighed by the positive effect of reducing channel interaction. It is important to note that although the effect of deactivation was significant at all visits, the greatest effect was seen at visit 2 after the subjects had been given two months of experience with the new frequency allocation assigned to the reduced number of electrodes. Although the effect of learning was small, the fact that the benefit increased suggested that the subjects experienced some forms of adaptation, but complete adaptation likely would take longer to occur (Reiss et al, 2014). The fact that the benefit of deactivation decreased after the subjects were switched back to their clinical maps indicated that the increased benefit of deactivation at visit 2 was attributed to learning the new map settings and not other confounding factors.

Effect of Deactivation on Sentence Recognition in Quiet

Similar to the SRT data, there was no evidence of learning the test procedure or material for sentence recognition in quiet. Performance collapsed across map conditions was comparable across visits. The effect of deactivation was overall smaller for listening in quiet than in noise. This was evidenced first by the fact that unlike the SRT data, the benefit of deactivation for sentence recognition in quiet was not statistically significant for all three visits. There was no acute benefit at visit 1. Analyses of the individual data also indicated that only a few comparisons between the clinical and experimental maps were significant using the critical difference values (Fig. 5, S10R and S22R), while a majority of the comparisons for individual data was significant for sentence recognition in noise (Fig. 3). Further, the benefit of deactivation for sentence recognition in quiet depended on subjects’ threshold variation patterns. A strong correlation between the across-site variation in the low-rate thresholds and the effect of deactivation suggested that subjects who had highly variable thresholds across the array were more likely to improve on sentence recognition in quiet with deactivation. The results possibly suggest that deactivation would only improve speech recognition in quiet if the reduction in channel interaction was accompanied also by a less variable excitation pattern across the array. The results aligned with previous findings that showed negative correlation between variance in the maximum comfortable loudness levels (C levels) and speech recognition performance (Pfingst & Xu, 2005). For sentence recognition in noise, deactivation of electrodes that had only slightly higher thresholds than others, in subjects with small variability in thresholds, still resulted in an improvement in tolerance of noise. Recall that for the SRT data, all subjects showed significant benefit from deactivation, regardless of their threshold variation pattern. These results suggested that an improvement in channel independence is more important for speech recognition in noise than for speech recognition in quiet.

The TIMIT and SRT data did share a similar pattern. That is, the effect of deactivation for both listening in noise and in quiet became greater at visit 2, after the subjects were given time to acclimate to the new frequency allocation. For sentence recognition in quiet, an adaptation process similar to that discussed earlier for the listening in noise tasks has occurred. The effect of this adaptation however seemed to have sustained over the time, after the treatment was withdrawn. The benefit only decreased by less than one percentage point at visit 3, and it remained marginally statistically significant. The average increase in performance for sentence recognition in quiet was 3.88%, 10.78% and 10.05%, at the three visits respectively.

Conclusions

Overall, deactivation of stimulation sites estimated to produce broad neural excitation led to improved speech recognition performance in noise and in quiet. For listening in noise, the effect of deactivation was comparable across the three noise conditions tested. With or without deactivation, performance was better in steady-state noises than in fluctuating noises, consistent with the previous reports. Deactivation resulted in significantly improved SRTs at all three visits, but the effect was the greatest after subjects were given time to acclimate to the new map. The quadratic pattern in the data suggests that performance systemically improved and worsened as the deactivation was applied and then withdrawn. With deactivation, all subjects improved on speech recognition in noise, even for those with small variability in their thresholds. The benefit of deactivation for speech recognition in quiet was only significant after the subjects were given time to adjust to the new map. There was no acute benefit. The benefit reduced again, although only slightly, after the deactivation was withdrawn. For speech recognition in quiet, it seemed that deactivation was more likely to help with subjects with a greater variability in thresholds.