Speech recognition as a function of the number of channels for an array with large inter-electrode distances

Abstract

This study investigated the number of channels available to cochlear implant (CI) recipients for maximum speech understanding and sound quality for lateral wall electrode arrays—which result in large electrode-to-modiolus distances—featuring the greatest inter-electrode distances (2.1–2.4 mm), the longest active lengths (23.1–26.4 mm), and the fewest number of electrodes commercially available. Participants included ten post-lingually deafened adult CI recipients with MED-EL electrode arrays (FLEX28 and STANDARD) entirely within scala tympani. Electrode placement and scalar location were determined using computerized tomography. The number of channels was varied from 4 to 12 with equal spatial distribution across the array. A continuous interleaved sampling-based strategy was used. Speech recognition, sound quality ratings, and a closed-set vowel recognition task were measured acutely for each electrode condition. Participants did not demonstrate statistically significant differences beyond eight channels at the group level for almost all measures. However, several listeners showed considerable improvements from 8 to 12 channels for speech and sound quality measures. These results suggest that channel interaction caused by the greater electrode-to-modiolus distances of straight electrode arrays could be partially compensated for by a large inter-electrode distance or spacing.

I. INTRODUCTION

A cochlear implant (CI) is a hearing prosthesis that stimulates the auditory nerve directly through an electrode array, surgically placed into the cochlea. Different electrodes along the array are designed to present lower frequencies to more apical electrodes and higher frequencies to more basal electrodes to mimic the tonotopic organization in the cochlea. The number of active electrodes along the array is manufacturer dependent and typically corresponds to the number of frequency bands provided, often referred to as channels. Channel interaction is a general term used to describe substantial overlap in electric fields from adjacent electrode contacts. In intracochlear electrical stimulation, channel interaction is unavoidable because the electrode contacts rest in a highly conductive fluid and are relatively far from the spiral ganglia in the modiolus, which contributes to poor spectral resolution (e.g., Jones et al., 2013; Won et al., 2014).

Improvements in speech recognition performance with an increasing number of active electrodes are assumed to indicate greater channel independence (Friesen et al., 2001). In the original CI studies of channel independence, adult CI recipients demonstrated asymptotic speech understanding with 4–8 channels. Specifically, Fishman et al. (1997) investigated channel independence in a group of 11 post-lingually deafened adults with Cochlear Nucleus N22 straight electrode arrays. In this study, participants used the spectral peak (SPEAK) coding strategy (Skinner et al., 1994), which is a version of “n-of-m” in which 6 channels with the greatest spectral energy are stimulated in each time frame out of 20 total electrodes (i.e., 6-of-20). They reported asymptotic speech recognition with five channels for consonants, four channels for topic-related sentences, and eight channels for vowels and monosyllables.

Friesen et al. (2001) further investigated channel independence in a group of ten Cochlear Nucleus N22 recipients and nine Advanced Bionics Clarion device recipients. The Cochlear N22 recipients used the SPEAK coding strategy, and the Clarion device recipients used either simultaneous analog stimulation (SAS) or continuous interleaved sampling (CIS). The SAS coding strategy stimulates all active electrodes simultaneously in an analog manner as opposed to CIS, which uses non-overlapping, pulsatile stimulation. The CIS coding strategy is envelope-based and implements stimulation on a fixed number of electrodes (Wilson, 1991). Friesen et al. (2001) reported asymptotic speech recognition scores in quiet and in noise with eight channels for vowels and consonants. They also found marginally significant improvements between 7 and 10 channels for monosyllables and sentences. More recently, Shannon et al. (2011) replicated their seminal report with seven Advanced Bionics CII recipients reporting no further gains beyond eight channels for any of the speech recognition measures (vowels, consonants, monosyllables, and sentences) as well as for subjective sound quality (SQ).

While these previous studies made significant contributions to the field, influencing clinical CI programming, the development of new signal processing strategies, and electrode design, the field has seen several significant changes since, including (a) expansion of CI criteria with many recipients now having significant residual hearing both before and after surgery, (b) improved surgical techniques with less traumatic electrode array insertion using slimmer, more flexible electrode arrays, (c) advances in signal processing strategies, (d) improvement in post-operative computerized tomography (CT) imaging coupled with advancement in image processing allowing specification of post-operative electrode array position for scalar location and distance to the modiolus, and (e) recognition that good CI performers often demonstrate ceiling effects on speech recognition measures in quiet.

Given these changes in the field, several studies have reexamined the long-standing contention that no more than eight independent channels are needed to maximize CI recipients' performance. For example, Croghan et al. (2017) reported significantly higher scores for 22 active electrodes versus 12 active electrodes in a group of ten adult CI recipients, nine with precurved electrode arrays and one with a straight electrode array. Similarly, Schvartz-Leyzac et al. (2017) looked at eight adult precurved electrode recipients and reported a significant improvement on sentences in quiet and on speech recognition thresholds in noise (dB SNR) using sentences with 20 compared to eight active electrodes. However, both studies used eight maxima in an n-of-m strategy, irrespective of the number of active electrodes. This is an important distinction because stimulation was limited to only the eight channels with the greatest spectral energy within each time frame, even with 20 electrodes active across the electrode array (i.e., 8-of-20). Additionally, these two studies did not have image-based confirmation of electrode array location for their precurved electrode array recipients. This is important, given that precurved electrode arrays are associated with scala tympani-scala vestibuli (ST-SV) translocation in up to 42% of cases (Wanna et al., 2014). Insertions of translocated electrode arrays are more traumatic to intracochlear structures and often lead to below average audiologic outcomes, including lower speech recognition scores and hearing preservation rates (O'Connell et al., 2016).

Our group recently conducted two studies investigating speech recognition as a function of the number of channels. First, we studied 11 precurved electrode array recipients—verified by imaging to be entirely in scala tympani—and demonstrated significant performance improvements with 22 active electrodes and 16 maxima using an n-of-m strategy (i.e., 16-of-22), as compared to programs with the same lower and upper stimulation level settings using 4–10 electrodes active with the CIS strategy (Berg et al., 2019b). For a CIS signal coding strategy, we observed significantly higher performance for 16 versus 8 active electrodes for monosyllable recognition and sentence recognition in noise at +5 dB signal-to-noise ratio (SNR). We hypothesized that this performance improvement occurs as a consequence of lower electrode-to-modiolus distance afforded by precurved electrode arrays because closer electrode-to-modiolus distances require less charge for upper stimulation levels (e.g., Saunders et al., 2002; Cohen, 2009; Davis et al., 2016). Indeed, electrode-to-modiolus distance and charge are essential factors, as a lower charge has been shown to yield less channel interaction (e.g., Chatterjee and Shannon, 1998). Studies have also shown a correlation between electrode arrays with lower electrode-to-modiolus distances and higher word recognition scores (Holden et al., 2013; Chakravorti et al., 2019).

In a second study, our group studied 18 post-lingually deafened adult CI recipients, including 11 recipients with straight electrode arrays entirely in scala tympani and seven recipients with precurved but translocated electrode arrays (Berg et al., 2020). Again, electrode placement was verified by imaging. Both groups of participants demonstrated asymptotic speech recognition scores with 8–10 active electrodes, consistent with previous literature (e.g., Fishman et al., 1997; Friesen et al., 2001; Shannon et al., 2011) and in contrast with our work with precurved electrode arrays well-placed in scala tympani. Given these results, we hypothesized that straight and translocated precurved electrode arrays have less channel independence because of their lateral wall placement as they are further from the spiral ganglia in the modiolus.

In addition to electrode-to-modiolus distance, another largely unexplored factor affecting channel independence is the spacing between neighboring electrodes on the array, referred to as inter-electrode distance. Adjacent electrodes can stimulate the same nerve fiber bundles if placed too close together. Even for non-simultaneous stimulation, spiral ganglia could be stimulated by an incorrect electrode if neighboring electrodes stimulate the same nerve fiber bundles. Increasing the inter-electrode distance could reduce this source of channel interaction and increase channel independence. Fu and Shannon (1999) examined how phoneme identification was affected by the spacing of electrode contacts in three adult Cochlear N22 recipients. Participants used the SPEAK coding strategy and four active electrodes at a time. Spatial separation between the electrodes was systematically increased from 1.5 to 4.5 mm by varying active electrodes while keeping the most apical active electrode fixed. Speech recognition performance peaked when electrodes were separated by 3–3.75 mm. While these results can provide insight into the effects of inter-electrode distance on speech recognition, this study used bipolar stimulation. Clinical processors today use monopolar stimulation, which has larger bandwidths than bipolar stimulation (Nelson et al., 2011).

The channel independence studies for adult CI recipients previously described examined electrode arrays with relatively small inter-electrode distances (∼0.7–1.1 mm). Only two studies, to our knowledge, have investigated channel independence with MED-EL CI systems. First, Garnham et al. (2002) tested 11 CI listeners with the COMBI 40+ CI system; however, it is unknown which electrode arrays were included in their sample. These implant systems' potential choices were the COMPRESSED, MEDIUM, or STANDARD electrode arrays, ranging in active length from 12.1 to 26.4 mm and in inter-electrode spacing from 1.1 to 2.4 mm, respectively. Electrode placement information was also not provided. Closed-set vowels presented at –10 dB SNR, closed-set consonants presented at +5 dB SNR, and sentences presented at +10 dB SNR were tested in 2-, 3-, 4-, 6-, 8-, and 10-channel conditions. Monosyllables presented both in quiet and at +10 dB SNR were also tested in 6-, 8-, and 11-channel conditions. Pink noise was used on all tasks. Results showed asymptotic performance with 6 channels for vowels, 4 channels for consonants, and 8–10 channels for sentences. Significant increases from 6 to 11 channels for monosyllables were also observed both in quiet and in noise. However, these results should be interpreted with caution because 7 of the 11 participants in this study had less than 6 months of experience with their CI, making distinguishing effects of CI duration of use and effects of channel conditions difficult (Wilson and Dorman, 2008). Second, Goupell et al. (2008) replicated the findings of Garnham et al. (2002) in seven CI listeners with either the MED-EL COMBI 40+ or Pulsar series implants. However, similarly to Garnham et al. (2002), it is unknown which electrode arrays were included and whether the electrode arrays were well-placed in scala tympani. Goupell et al. (2008) demonstrated that 8–10 channels were sufficient for achieving asymptotic performance on sentences in quiet and noise. Findings from both studies are grossly consistent with previous data investigating channel independence with straight electrodes (Fishman et al., 1997; Friesen et al., 2001; Shannon et al., 2011; Berg et al., 2020).

The current study aimed to investigate the number of independent channels available to MED-EL recipients with electrode arrays verified to be well-placed in scala tympani by imaging data. We narrowed our focus to the FLEX28 (28 mm) or STANDARD (31.5 mm) electrode arrays because they feature the largest inter-electrode distances (2.1 or 2.4 mm, respectively) commercially available and present the best opportunity to decrease channel overlap. We proposed two potential hypotheses for the number of channels available to MED-EL recipients for asymptotic speech recognition and SQ measures. On the one hand, the current MED-EL recipients could perform similarly to recipients with Cochlear straight electrode arrays and the previously tested MED-EL COMBI 40+ and Pulsar recipients, limited to 8–10 independent channels, because of more distal electrode-to-modiolus placement near the lateral wall of the cochlea. On the other hand, the current MED-EL recipients could perform similarly to recipients with well-placed Cochlear precurved electrode arrays because of decreased channel overlap caused by the large inter-electrode distances featured on the STANDARD and FLEX28 electrode arrays. This potential reduction in channel interaction could afford these recipients access to more than 8–10 independent channels.

II. METHODS

A. Study participants

All ten participants were post-lingually deafened adult CI recipients with MED-EL CI systems. Four participants had Concert CI systems, five participants had Synchrony CI systems, and one participant had a Synchrony2 CI system. Three participants had a MED-EL STANDARD electrode array, which is 31.5 mm in length (measured from the base where the electrode array joins to the receiver simulator to the apical tip) and 26.4 mm in active length (distance from electrode 1 to 12) and has an inter-electrode distance of 2.4 mm (measured between the centers of adjacent electrodes). Eight participants had a MED-EL FLEX28 electrode array, which is 28 mm in length and 23.1 mm in active length and has an inter-electrode distance of 2.1 mm. These two electrode arrays were selected among all Food and Drug Administration (FDA)-approved devices because they have the longest length and largest inter-electrode distance, which could allow less channel interaction from neighboring electrodes than shorter, more compact electrode arrays. All participants had all 12 electrodes within scala tympani as confirmed by post-operative CT scans and image analysis (e.g., Noble et al., 2012). Mean electrode-to-modiolus ($\overline{m}$) distances for each electrode contact were calculated for each participant and each condition during image analyses. For example, the four-channel condition $\overline{m}$ for participant 1 was calculated by averaging the $\overline{m}$ distances for electrodes 1, 4, 8, and 12. Inclusion criteria required at least 6 months of CI experience and at least 11 active electrodes in their clinical program. Participants were also required to score 20% correct or better on AzBio sentences at +5 dB SNR with their clinical program to avoid floor effects. Table I provides demographic information for all participants.

TABLE I.

Participant demographics, including electrode array type, channel stimulation rate of CIS channels, CI experience in months, active electrodes in clinical program, mean electrode-to-modiolus distance ($\overline{m}$) in mm, angle of insertion depth in degrees, age in years, and AzBio sentence score at +5 dB SNR with their clinical program. Electrodes (E) deactivated in the clinical program are notated in parentheses. The $\overline{m}$ listed includes all 12 electrode contacts. N/A, not applicable.

Participant No.	Electrode array	Channel stimulation rate (pps)	CI usage experience (months)	Active electrodes in clinical program (deactivated electrode)	Mean electrode-to-modiolus distance ($\overline{m}$) (mm)	Angle of insertion depth (deg)	Age (yrs)	AzBio score at +5 dB SNR with clinic program (%)
1	FLEX28	1436	19	11 (E11)	1.19	394	58	22
2	STANDARD	1293	68	11 (E8)	1.17	701	73	48
3	STANDARD	1207	29	12	1.26	447	54	39
4	FLEX28	1237	47	12	1.12	500	52	22
5	STANDARD	1600	80	12	1.37	529	45	58
6	FLEX28	1389	34	12	1.37	510	35	43
7	FLEX28	1093	68	11 (E2)	1.30	513	68	27
8	FLEX28	1207	70	12	1.24	581	71	55
9	FLEX28	1351	7	11 (E12)	1.04	532	55	65
10	FLEX28	1224	21	11 (E2)	1.25	458	75	20
Mean	N/A	N/A	44.3	N/A	1.23	516	58.6	40

B. Materials and conditions

All experimental activities were completed in accordance with Institutional Review Board (IRB) approved protocols at Vanderbilt University and Vanderbilt University Medical Center. Prior to testing with the experimental programs, participants were tested using their clinical program settings with 12 active electrodes and the Fine Structure 4 (FS4) speech coding strategy, the clinical default strategy used by all participants at the time of enrollment, to establish baseline performance. The FS4 speech coding strategy functions similarly to a CIS-based strategy but is designed to also transmit temporal fine structure (TFS) information via variable rate stimulation on the four most apical electrodes. The channel stimulation rate for the remaining eight electrodes is designed to be as high as possible but is dependent on each individual's measured stimulation levels as well as the number of active electrodes. The channel stimulation rates of the non-TFS electrodes are listed for each participant in Table I. To limit these potential differences in TFS information provided and channel stimulation rates, the four experimental CI programs using 4, 8, 10, and 12 active electrodes were created using the high definition continuous interleaved sampling (HDCIS) speech coding strategy, which is the CIS-based strategy available to MED-EL recipients. The channel stimulation rates listed in Table I were used in the experimental HDCIS programs and kept constant regardless of the number of active electrodes. Figure 1 displays the specific electrodes activated to achieve the spatially selective programs. It should be noted that this study did not use patient customized image-guided electrode deactivation techniques used by the authors in other investigations (Noble et al., 2014).

FIG. 1.

Channel deactivation methods and associated frequency allocations for all conditions. In cases for which the participant had an electrode(s) deactivated in their clinical program, we chose to activate the closest available electrode to maintain the greatest spatial separation between activated electrodes. For example, if electrode 1 elicited a non-auditory percept (e.g., facial stimulation), electrode 2 would be activated instead for the four-channel condition.

The spatially selective programs were based on the deactivation methods of Friesen et al. (2001); however, this is not a direct replication, as they did not use MED-EL electrode arrays and they had maintained the participants' everyday frequency allocation map throughout all conditions. Instead, the present study—as well as our previous study investigating musical SQ as a function of the number of channels (Berg et al., 2019a)—automatically reallocated the frequency map based on the number of active electrodes as would occur in a clinical manipulation. In other words, though the input frequency range was held constant across the different experimental conditions (250–8500 Hz), as the number of active electrodes was reduced, the frequency bands allocated for each channel were broader for each active electrode as compared to the clinical frequency allocations.

The same Sonnet behind-the-ear processor was used for all ten participants during the study. Parameters were left as programmed in the participants' own programs. Participants with electrodes deactivated in their clinical program were left deactivated during the study. Specifically, participants 1, 2, 7, 9, and 10 all had one electrode deactivated clinically. The remaining five participants had all 12 electrodes active in their clinical program (see Table I). Upper stimulation levels were globally adjusted on each experimental program using a loudness scaling chart to achieve an equivalent “most comfortable” rating from the participant. Loudness ratings were completed in “live speech mode” for all experimental programs. Threshold levels were not adjusted from the participant's clinical program; however, aided detection thresholds with the CI were verified to be within 20–30 dB hearing level (HL) from 250 to 6000 Hz before the participant began the study.

Channel condition and assessment measure order were both randomized using a Latin square design. All testing was completed acutely; that is, there was no acclimatization period provided for the different channel conditions, consistent with past studies (e.g., Fishman et al., 1997; Friesen et al., 2001; Shannon et al., 2011; Croghan et al., 2017; Schvartz-Leyzac et al., 2017). All testing was performed partially blinded. Participants were blinded to the program they were listening with during the study. However, the same research assistant programmed and tested all participants and was aware of which program the participant was listening with during testing. The four CI experimental programs and the participant's clinical program were tested using a loudspeaker at 0° azimuth and 1 m from the participant in a single walled sound booth using consonant-nucleus-consonant (CNC) words, AzBio sentences in quiet and in noise at +5 dB SNR using 20-talker babble noise, and vowels (closed set). Participants completed one 50-word list of CNC words and two 20-sentence lists of AzBio sentences, one in quiet and one at +5 dB SNR, for each channel condition. Speech recognition lists were randomized, and lists were not repeated for any given participant. Subjective SQ judgments were assessed using a visually presented 10-point scale (1 = very poor; 10 = very good), in which the participant rated the overall SQ of the list of CNC words and AzBio sentences in quiet and +5 dB SNR for each condition. Subjective SQ ratings were assessed immediately following the corresponding speech recognition measure. For example, participants completed a list of CNC words and then made a SQ judgment based on the list of CNC words. Participants were allowed to rate the SQ in between two numbers. For example, if a participant noted the SQ between a 3 and a 4, a rating of 3.5 was recorded. Vowel stimuli consisted of 13 synthetic vowels in /bVt/ format (“bait, Bart, bat, beet, Bert, bet, bit, bite, boat, boot, bought, bout, but”). Vowel phonemes (e.g., /ai/) were equal duration (90 ms) so that vowel length could not serve as a cue. Target stimuli were presented at a calibrated level of 60 dB sound pressure level (SPL) 1 m from the loudspeaker and 0° azimuth from the seated participant in free field.

III. RESULTS

A one-way multivariate analysis of covariance (MANCOVA) was completed with number of channels as the independent variable, mean electrode-to-modiolus distance ($\overline{m}$) as a co-variate, and speech/auditory perception scores and SQ ratings as the dependent variables. In an attempt to minimize the influence of floor and ceiling effects, CNC words and AzBio sentence recognition scores in quiet and in noise were converted from percent correct to rationalized arcsine units (RAU) (Studebaker, 1985) prior to all analyses. Sound quality ratings for CNC words and AzBio sentences in quiet and in noise were converted to standardized z-scores prior to analysis to mitigate inter-rater variability. Post hoc analyses were completed with all-pairwise, multiple comparisons using a Holm–Sidak statistic. There was a statistically significant difference between the number of channels on the combined dependent speech recognition and SQ rating variables, collapsed across measures, after accounting for mean electrode-to-modiolus distance (F_{(21, 84)} = 2.78, p = 0.001, Wilk's Λ = 0.22 η_p² = 0.40). A summary of the results is displayed in Table II and described in further detail below.

TABLE II.

Results of MANCOVA for speech recognition performance and SQ ratings. The corrected model was used to interpret the results. Post hoc analyses were pairwise Holm–Sidak comparisons. d.f., degrees of freedom.

Test	Channels				$\overline{m}$
	d.f.	F ratio	p	Post hoc (p < 0.05)	d.f.	F ratio	p
CNC words (RAU)	3, 35	14.23	<0.001	4 < 8, 10, 12	1, 35	1.60	0.214
CNC SQ	3, 35	9.21	<0.001	4 < 8, 10, 12	1, 35	1.29	0.263
AzBio quiet (RAU)	3, 35	12.41	<0.001	4 < 8, 10, 12	1, 35	3.40	0.074
AzBio quiet SQ	3, 35	14.58	<0.001	4 < 8, 10, 12 8 < 12	1, 35	2.94	0.095
AzBio noise (RAU)	3, 35	9.64	<0.001	4 < 8, 10, 12	1, 35	19.26	<0.001
AzBio noise SQ	3, 35	8.12	<0.001	4 < 8, 10, 12	1, 35	13.30	0.001
Vowels	3, 35	4.07	0.008	4 < 12	1, 35	10.06	0.003

Mean CNC words, AzBio sentence recognition scores in quiet and in noise, and vowels are displayed in percent correct in Fig. 2. Mean SQ ratings for CNC words and AzBio sentences in quiet and in noise are shown in Fig. 3. Error bars represent ±1 standard error measure (SEM). Horizontal lines across the 12-channel condition represent mean scores for the participants with their clinical program using 11 or 12 channels and the FS4 speech coding strategy. The condition using the participant's clinical program was used for comparison purposes only to estimate a potential familiarity effect and thus was not included in statistical analyses.

FIG. 2.

Mean speech recognition outcomes, in percent correct, for the ten participants across all channel conditions for CNC words (a), AzBio sentences in quiet (b) and at +5 dB SNR (c), and vowels (d). Error bars are ±1 SEM. Horizontal dotted lines across each measure represent mean scores for the ten participants with their clinical program using 12 channels and the FS4 speech coding strategy.

FIG. 3.

Mean SQ ratings for the ten participants across all channel conditions for CNC words (a) and AzBio sentences in quiet (b) and at +5 dB SNR (c). Error bars are ±1 SEM. Horizontal dotted lines across each measure represent mean scores for the ten participants with their clinical program using 12 channels and the FS4 speech coding strategy.

A. CNC word recognition and SQ

Mean CNC word recognition was 32%, 60%, 65%, and 74% for the 4-, 8-, 10-, and 12-channel conditions, respectively. There was a significant main effect of number of channels (F_{(3, 35)} = 14.23, p < 0.001, η_p² = 0.62) for CNC word recognition. There was not a significant effect of the co-variate $\overline{m}$ on CNC word recognition (F_{(1, 35)} = 1.60, p = 0.214, η_p² = 0.04). Post hoc analyses revealed significant performance differences between four electrodes and all other electrode conditions (p < 0.001 for all comparisons). No other comparisons were statistically significant for this measure. Mean CNC word SQ ratings were 2.7, 5.0, 6.1, and 6.9 for the 4-, 8-, 10-, and 12-channel conditions, respectively. There was a significant main effect of number of channels (F_{(3, 35)} = 9.21, p < 0.001, η_p² = 0.51) for CNC word SQ ratings. There was not a significant effect of the co-variate $\overline{m}$ on CNC word SQ ratings (F_{(1, 35)} = 1.29, p = 0.263, η_p² = 0.04). Post hoc analyses revealed significant performance differences between 4 and 8 channels (p = 0.027), 4 and 10 channels (p < 0.001), and 4 and 12 channels (p < 0.001). No other comparisons were statistically significant for this measure.

B. AzBio sentence recognition in quiet and SQ

Mean AzBio sentence recognition in quiet was 36%, 73%, 79%, and 87% for the 4-, 8-, 10-, and 12-channel conditions, respectively. There was a significant main effect of number of channels (F_{(3, 35)} = 12.41, p < 0.001, η_p² = 0.59) for AzBio sentence recognition in quiet. There was not a significant effect of the co-variate $\overline{m}$ on AzBio sentence recognition in quiet (F_{(1, 35)} = 3.40, p = 0.074, η_p² = 0.09). Post hoc analyses revealed significant performance differences between four electrodes and all other electrode conditions (p < 0.001 for all comparisons). No other comparisons were statistically significant for this measure. Mean AzBio SQ in quiet ratings was 2.1, 5.2, 6.2, and 7.5 for the 4-, 8-, 10-, and 12-channel conditions, respectively. There was a significant main effect of number of channels (F_{(3, 35)} = 14.58, p < 0.001, η_p² = 0.63) for AzBio SQ in quiet ratings. There was not a significant effect of the co-variate $\overline{m}$ on AzBio SQ in quiet ratings (F_{(1, 35)} = 2.94, p = 0.095, η_p² = 0.08). Post hoc analyses revealed significant performance differences between 4 and 8 channels (p = 0.001), 4 and 10 channels (p < 0.001), and 4 and 12 channels (p < 0.001). There was also a significant difference between 8 and 12 channels (p = 0.014). No other comparisons were statistically significant for this measure.

C. AzBio sentence recognition at +5 dB SNR and SQ

Mean AzBio sentence recognition at +5 dB SNR was 7%, 27%, 28%, and 37% for the 4-, 8-, 10-, and 12-channel conditions, respectively. There was a significant main effect of number of channels (F_{(3, 35)} = 9.64, p < 0.001, η_p² = 0.52) for AzBio sentence recognition at +5 dB SNR. There was also a significant effect of the co-variate $\overline{m}$ on AzBio sentence recognition at +5 dB SNR (F_{(1, 35)} = 19.26, p < 0.001, η_p² = 0.36). Post hoc analyses revealed significant performance differences between 4 and 8 channels (p = 0.030), 4 and 10 channels (p = 0.018), and 4 and 12 channels (p < 0.001). No other comparisons were statistically significant for this measure. Mean AzBio SQ at +5 dB SNR ratings was 0.6, 2.7, 2.9, and 3.4 for the 4-, 8-, 10-, and 12-channel conditions, respectively. There was a significant main effect of number of channels (F_{(3, 35)} = 8.12, p < 0.001, η_p² = 0.48) for AzBio SQ at +5 dB SNR ratings. There was also a significant effect of the co-variate $\overline{m}$ on AzBio SQ at +5 dB SNR ratings (F_{(1, 35)} = 13.30, p = 0.001, η_p² = 0.28). Post hoc analyses revealed significant performance differences between 4 and 8 channels (p = 0.020), 4 and 10 channels (p = 0.006), and 4 and 12 channels (p < 0.001). No other comparisons were statistically significant for this measure.

D. Vowel recognition

Mean vowel recognition was 23%, 37%, 43%, and 47% for the 4-, 8-, 10-, and 12-channel conditions, respectively. There was a significant main effect of number of channels (F_{(3, 35)} = 4.07, p = 0.008, η_p² = 0.32) for vowel recognition. There was also a significant effect of the co-variate $\overline{m}$ on vowel recognition (F_{(1, 35)} = 10.06, p = 0.003, η_p² = 0.22). Post hoc analyses revealed significant performance differences between 4 and 12 channels (p = 0.040). No other comparisons were statistically significant for this measure.

E. Electrode-to-modiolus distance

Pearson correlations (two-tailed) were used to examine the relationship between $\overline{m}$ and speech recognition scores and SQ ratings in the 12-channel HDCIS condition. Results from these comparisons are displayed in Fig. 4. The $\overline{m}$ across all 12 electrodes, displayed in Table I, was used for these comparisons. There was no significant relationship between CNC word recognition and $\overline{m}$ (r = –0.43, p = 0.214), CNC SQ ratings and $\overline{m}$ (r = –0.14, p = 0.703), AzBio sentence recognition in quiet and $\overline{m}$ (r = –0.40, p = 0.247), AzBio sentences in quiet SQ ratings and $\overline{m}$ (r = –0.26, p = 0.461), or AzBio sentences at +5 dB SNR SQ ratings and $\overline{m}$ (r = –0.61, p = 0.060). However, there was a significant negative relationship between AzBio sentence recognition at +5 dB SNR and $\overline{m}$ (r = –0.70, p = 0.023) and between vowel recognition and $\overline{m}$ (r = –0.71, p = 0.020).

FIG. 4.

Mean electrode-to-modiolus distance across all 12 electrodes compared to performance in the 12-channel HDCIS condition for CNC words (a), CNC SQ ratings (b), AzBio sentences in quiet (c), AzBio sentences in quiet SQ ratings (d), AzBio sentences at +5 dB SNR (e), AzBio sentences at +5 dB SNR SQ ratings (f), and vowels (g).

Pearson correlations (two-tailed) were also used to examine the relationship between $\overline{m}$ and the change in speech recognition scores and SQ ratings from the 8-channel to 12-channel HDCIS conditions, though not displayed in the figures. However, there was no significant relationship between CNC word recognition and $\overline{m}$ (r = –0.45, p = 0.192), CNC SQ ratings and $\overline{m}$ (r = 0.21, p = 0.565), AzBio sentence recognition in quiet and $\overline{m}$ (r = 0.19, p = 0.609), AzBio sentences in quiet SQ ratings and $\overline{m}$ (r = 0.22, p = 0.537), AzBio sentence recognition at +5 dB SNR and $\overline{m}$ (r = 0.11, p = 0.758), AzBio sentences at +5 dB SNR SQ ratings and $\overline{m}$ (r = 0.12, p = 0.733), or vowel recognition and $\overline{m}$ (r = –0.18, p = 0.611).

IV. DISCUSSION

A. Summary of results

MED-EL recipients, with electrode arrays verified to be completely within scala tympani, displayed a positive trend toward performance improvements beyond 4–8 spatially selective channels. Participants demonstrated significant improvements in SQ ratings between the 8- and 12-channel conditions for sentences in quiet, suggesting noticeable improvements in SQ beyond eight channels. Given these results, having a large (2.1–2.4 mm) inter-electrode distance could partially compensate for the adverse effects of channel interaction caused by the greater electrode-to-modiolus distance of straight electrode arrays compared to precurved arrays. This is supported by the significant negative relationship found between both mean electrode-to-modiolus distance $\overline{m}$ and sentence recognition in noise, and $\overline{m}$ and vowel recognition. However, at the group level, word recognition and SQ ratings, sentence recognition in quiet and SQ ratings, sentence recognition in noise, and vowel recognition measures were not significant beyond eight channels.

While only SQ ratings for sentences in quiet showed significant improvements beyond eight channels at the group level, it is important to distinguish statistical significance from clinical significance. A clinically significant difference is considered to be 10% or greater on speech recognition measures (Bierer et al., 2016). Comparing the 8-channel condition with the 12-channel condition, five participants clinically improved on words by 20%–40%, five participants clinically improved on sentences in quiet by 12%–36%, four participants clinically improved on sentences in noise by 17%–25%, and six participants clinically improved on vowels by 13%–31%. For these participants, even deactivating as few as two electrodes could be detrimental to their speech recognition performance. This holds clinical application, given that a number of studies have investigated potential benefits of channel deactivation in an attempt to improve channel interaction and subsequent auditory performance (e.g., Noble et al., 2014; Bierer and Litvak, 2016; Zhou, 2017).

In contrast, two participants demonstrated a clinically significant decrement on sentences in noise by 12%–18% when the number of channels increased beyond eight channels. For these two participants, deactivating 2–4 electrodes in a dedicated speech-in-noise program could help improve their speech-in-noise scores; however, it is unclear based on these results which electrodes would be optimal to deactivate. Importantly, these same two participants also demonstrated stable or significant improvements with more than eight channels on other measures, so a program with electrodes deactivated may not be optimal as an “everyday” program. No participants demonstrated a clinically significant decrement on any other measure between the 8-channel and 12-channel conditions.

For SQ ratings, there are no guidelines for what is considered a clinically significant improvement or decrement. Comparing the 8-channel condition with the 12-channel condition, six participants rated SQ better for words, eight participants rated SQ better for sentences in quiet, and five participants rated SQ better for sentences in noise. In contrast, two participants rated SQ worse for words, one participant rated SQ worse for sentences in quiet, and two participants rated SQ worse for sentences in noise as the number of channels was increased from 8 to 12. The authors recommend asking the CI recipient about their SQ preferences in addition to comparing their speech recognition scores after changes are made to the number of active electrodes. Due to this individual variability seen in the current study, it is not recommended that clinicians deactivate electrodes for MED-EL recipients simply based on the current study's results.

Individual participant performance was analyzed to identify those who systematically showed improvements beyond eight channels and those who systematically showed decrements beyond eight channels for all speech recognition and SQ measures. This analysis was done to explore potential characteristics that might help identify those CI recipients who would benefit from leaving all 12 electrodes active versus those who would benefit from electrode deactivation. Three participants demonstrated improvements beyond eight channels on every speech recognition and SQ measure tested. These three participants had an average $\overline{m}$ of 1.29 mm and angular insertion depth of 503°, compared to the group average $\overline{m}$ of 1.23 mm and angular insertion depth of 516° of all participants. Of note, the participant who demonstrated the greatest improvements beyond eight channels across all of the measures had the smallest angular insertion depth (394°) of all of the participants. Six participants demonstrated improvements beyond eight channels on the majority of tasks, either five or six of the seven measures. For these participants, it would be most beneficial to leave all 12 electrodes active.

Only one participant demonstrated decreases beyond eight channels on all measures. This participant had an angular insertion depth of 701°, which is much deeper compared to the other participants (range: 394°–581°). For this participant, it would be most beneficial to implement an electrode deactivation plan, though it is unclear which electrodes would be optimal to deactivate based on the current results. Prior research has shown that straight electrode arrays inserted beyond 650° are more likely to disrupt the basilar membrane and/or translocate as well as result in poorer hearing preservation and speech recognition outcomes (O'Connell et al., 2017; Helbig et al., 2018; Morrel et al., 2020). This is likely due to a significant reduction in scala tympani height at the lateral wall of the cochlea beyond 450° (Avci et al., 2014). As such, with a deep angular insertion depth, we hypothesize that at least the most apical electrode of this participant could be causing significant channel interaction and might be a prime candidate for deactivation.

B. Limitations and future directions

In this study, we chose $\overline{m}$ a priori as one measure of spatial selectivity. However, through our exploratory analysis of individual performance across all measures, we noted that angle of insertion depth may also contribute to the individual variability noted in this study. Further work with a larger sample size of participants with electrode arrays inserted beyond 650° is needed to investigate whether angle of insertion depth could be used to predict whether MED-EL CI recipients would benefit from electrode deactivation or if they would perform optimally with all 12 electrodes active. Another potential related variable, unexplored in the current study, is cochlea size, which is likely related to spatial selectivity from a CI. Future studies should investigate the individual contributions of $\overline{m}$, angle of insertion depth, and cochlea size toward spatial selectivity and channel independence.

In contrast to our previous investigation with Cochlear straight electrode array recipients (Berg et al., 2020), spectral modulation detection using the acoustic version of the quick spectral modulation detection (QSMD) test (Gifford et al., 2014) was not assessed in the current study. This measure was removed from the current study's protocol in an effort to reduce participation time after results from Berg et al. (2020) demonstrated no significant main effect of channels on the QSMD. However, future studies investigating channel independence for MED-EL electrode array recipients should include a sensitive measure of spectral resolution to examine the mechanism underlying the individual variability observed in the current study. We hypothesize that CI recipients with the longest available electrode arrays and the largest available inter-electrode spacing are able to achieve better spectral resolution than those with shorter arrays and closer inter-electrode spacing. This supposition is supported by the current results demonstrating a significant effect of $\overline{m}$ for measures most dependent upon spectral resolution. However, there may be limits to the potential benefits of longer electrode arrays with insertions beyond 650°.

In this study, there was potential bias for the “all on” condition, which utilized the same number of active electrodes and channel-specific frequency allocations as the participant's clinical program (Skinner et al., 2002). However, participants listened with the FS4 speech coding strategy with their clinical program and not the traditional CIS-based coding strategy as used in the current study. The “all on” condition had the potential to be more familiar to participants than the other tested conditions because it is most similar to their clinical program. Despite this potential, participants demonstrated equal or even better performance with the “all on” condition using HDCIS versus their clinical program with FS4. This is illustrated in Figs. 2 and 3 using the horizontal bars to indicate performance with the clinical program overlaying performance with the “all on” program. Importantly, all conditions were tested acutely. As such, the contribution of this familiarity effect is thought to be less critical in this study, though further work is needed to investigate the potential of such familiarity effects on performance.

The MED-EL CI system features a large portfolio of electrode arrays, and the current study only tested participants with the two longest electrode arrays available, the STANDARD and the FLEX28. These electrode arrays were chosen for the current study because they feature the greatest inter-electrode distance and therefore the best potential for minimizing this source of channel interaction. Given this, we would hypothesize that channel independence would not be better with other MED-EL electrode arrays not tested in the current study (e.g., FLEXsoft, FLEX24, FLEX20) because the inter-electrode distance is either equivalent to or smaller than those in the current study. However, further work is needed to fully understand channel independence with MED-EL electrode arrays and the potential to mitigate adverse effects of channel interaction on speech recognition and SQ measures by using a large (2.1–2.4 mm) inter-electrode distance between adjacent electrodes.

In keeping with the protocols of our previous studies (Berg et al., 2019b, 2020), the current study did not completely randomize the SQ judgments within each condition. Participants were asked to judge the SQ of the program they were listening with after completing the corresponding speech recognition measure. For example, participants completed a list of CNC words and then was asked to make a SQ judgment based on the CNC words they just heard. We did not share speech recognition scores with the participants until after they completed the study, but we do recognize that participants may have been influenced by how well they perceived they performed on the speech recognition measure when they were making their SQ judgments. We found that completely randomizing the SQ judgments caused confusion for the participants because they had difficulty remembering how they perceived the SQ unless they were asked immediately after completing the speech recognition of interest. Further research should investigate this potential bias by limiting the number of programs participants are evaluating in one session to help alleviate potential confusion or using entirely separate stimuli, independent of speech recognition assessment, for gauging SQ.

C. Summary

• •Well-placed MED-EL STANDARD and FLEX28 electrode array recipients did not demonstrate statistically significant differences beyond eight channels on measures of speech recognition at the group level.
• •Participants demonstrated significant, noticeable improvements in SQ ratings with more than eight channels.
• •Nine of the ten participants demonstrated clinically significant improvements (≥10%) on at least one measure with more than eight channels.
• •A large (2.1–2.4 mm) inter-electrode distance could partially compensate for the adverse effects on channel independence caused by the greater electrode-to-modiolus distance of straight electrode arrays.
• •Mean electrode-to-modiolus distance had the greatest impact on channel independence for sentence recognition in noise and vowel recognition, tasks that are assumed to heavily rely on spectral resolution, which suggests that spatial selectivity is more important for these types of tasks.