Spectral Modulation Detection Performance and Speech Perception in Pediatric Cochlear Implant Recipients

Abstract

Purpose:

The aims of this study were, for pediatric cochlear implant (CI) recipients, (a) to determine the effect of age on their spectral modulation detection (SMD) ability and compare their age effect to that of their typically hearing (TH) peers; (b) to identify demographic, cognitive, and audiological factors associated with SMD ability; and (c) to determine the unique contribution of SMD ability to segmental and suprasegmental speech perception performance.

Method:

A total of 104 pediatric CI recipients and 38 TH peers (ages 6–11 years) completed a test of SMD. CI recipients completed tests of segmental (e.g., word recognition in noise and vowels and consonants in quiet) and suprasegmental (e.g., talker discrimination, stress discrimination, and emotion identification) perception, nonverbal intelligence, and working memory. Regressions analyses were used to examine the effects of group and age on percent-correct SMD scores. For the CI group, the effects of demographic, audiological, and cognitive variables on SMD performance and the effects of SMD on speech perception were examined.

Results:

The TH group performed significantly better than the CI group on SMD. Both groups showed better performance with increasing age. Significant predictors of SMD performance for the CI group were age and nonverbal intelligence. SMD performance predicted significant variance in segmental and suprasegmental perception. The variance predicted by SMD performance was nearly double for suprasegmental than for segmental perception.

Conclusions:

Children in the CI group, on average, scored lower than their TH peers. The slopes of improvement in SMD with age did not differ between the groups. The significant effect of nonverbal intelligence on SMD performance in CI recipients indicates that difficulties inherent in the task affect outcomes. SMD ability predicted speech perception scores, with a more prominent role in suprasegmental than in segmental speech perception. SMD ability may provide a useful nonlinguistic tool for predicting speech perception benefit, with cautious interpretation based on age and cognitive function.

There has been increasing interest in determining the relation between spectral resolution abilities and speech understanding for listeners with a cochlear implant (CI). Spectral resolution ability is most often assessed with a spectral ripple discrimination (SRD) test, the spectral-temporally modulated ripple test (SMRT), or a spectral modulation detection (SMD) test. In an SRD test, the SRD threshold is the highest spectral modulation rate, measured in ripples per octave (RPO), at which a listener can discriminate between a rippled noise and its phase-reversed rippled counterpart when the modulation depth remains constant (Henry & Turner, 2003). In the SMRT, spectrally rippled stimuli have a modulation phase that drifts with time (Aronoff & Landsberger, 2013). The SMRT threshold is the highest spectral density (in RPO) that can be discriminated from a reference stimulus with 20 RPO. In an SMD test, the SMD threshold is the smallest modulation depth at which the participant can discriminate between a spectrally modulated noise and one with no spectral modulation, for a fixed spectral modulation rate (Litvak et al., 2007). In general, research from adult CI recipients suggests that performance on SMD, SMRT, and SRD tasks correlates with speech recognition performance (Anderson et al., 2012; Gifford et al., 2018; Henry & Turner, 2003; Henry et al., 2005; Saoji et al., 2009; Spahr et al., 2011). Spectral resolution tasks measure patient performance independent of linguistic ability, making them ideally suited for assessing speech perception potential in young pediatric patients with deficits in spoken language abilities due to hearing loss (Drennan et al., 2014, 2016).

From cross-sectional studies of typically hearing (TH) children, results suggest that spectral resolution abilities develop through ages 7–8 years, with performance at approximately 9 years of age approaching adultlike levels (Allen & Wightman, 1992; Kirby et al., 2015; Landsberger et al., 2018). For children with CIs, however, studies have generally not found improved performance on spectral modulation or spectral ripple tasks with increasing age (Gifford et al., 2018; Horn et al., 2017; Landsberger et al., 2018). One exception is the study by Jeddi et al. (2019) in which SMRT performance improved with the pediatric CI recipient's age. Studies comparing performance of adult CI recipients to prelingually deafened pediatric CI recipients have also yielded equivocal findings; some report that pediatric patients perform more poorly than adults on tasks of spectral resolution (Gifford et al., 2018; Landsberger et al., 2018), whereas others show similar performance (Jeddi et al., 2019; Jung et al., 2012).

Studies examining the effects of CI recipient characteristics, such as age at CI and duration of CI use, on spectral resolution performance have also yielded inconsistencies, with some showing no significant effects (Gifford et al., 2018; Landsberger et al., 2018) and another showing better performance related to earlier receipt of a CI (Jeddi et al., 2019). More recently, Kirby et al. (2019) examined the effects of cognitive abilities of 24 pediatric hearing aid (HA) recipients (ages 6–13 years) on their SMRT performance. The cognitive battery included tests of nonverbal intelligence (nonverbal intelligence quotient [IQ]), visual working memory, auditory working memory, and executive function. Regression analyses revealed that higher nonverbal IQ and visual working memory span were associated with better SMRT performance. The authors suggest that these correlations may reflect the general cognitive demands of the task rather than a direct relationship between cognitive abilities and spectrotemporal processing. They conclude that future studies of other clinical populations, including pediatric CI recipients, may serve to elucidate the role of cognition in spectral resolution tasks.

To date, the research literature on pediatric CI recipients has been equivocal regarding the relation between spectral resolution performance and speech perception. Jung et al. (2012) found that performance on the spectral ripple task was significantly correlated with speech reception thresholds for spondees in noise (r = −.72, p = .018) for 10 pediatric CI recipients (M _age = 12 years). Horn et al. (2017) tested 15 pediatric CI recipients, 7–15 years old, and also found statistically significant correlations between SRD performance, for modulation depths of 10 and 20 dB, and thresholds for 50% recognition of spondees in noise (r = −.64 and −.54, corrected p values = .005 and .038, respectively). Jeddi et al. (2019) administered the SMRT task to 75 prelingually deafened children with CIs (8–12 years of age) and found that SMRT performance was statistically significantly correlated with word and sentence recognition (r = .57, p < .001 and r = .56, p < .001, respectively). By contrast, Gifford et al. (2018) found that for 36 pediatric CI recipients (ages 5.6–17.9 years), performance on the quick SMD (QSMD) task was not significantly correlated with recognition of words in quiet (r = .31, n = 36, p = .06), sentences in quiet (r = .09, n = 36, p = .61), or sentences in noise (r = .12, n = 22, p = .61).

These studies of relations between spectral resolution performance and speech perception have focused primarily on segmental perception (phonemes, words, and sentences) or “what is said.” The role of segmental perception in facilitating spoken language development in deaf or hard of hearing (DHH) children has been well documented (Blamey et al., 2001; Ching et al., 2018; Geers, 2006; Geers et al., 2003).

There is another category of speech perception, however, that has received far less attention, namely, suprasegmental perception. Although segmental perception may be thought of as the ability to understand “what is said,” suprasegmental perception refers to the ability to perceive “how it is said” (Wenrich et al., 2017). Suprasegmental perception includes the ability to discern a talker's vocal emotion, gender, age, and identity, as well as the linguistic cues to syllable and word stress, and intonation cues that differentiate questions from statements. A few studies have demonstrated a unique contribution of suprasegmental perception to language development in children who are DHH with CIs (Davidson et al., 2019; Geers et al., 2013).

Typically developing infants attend to the acoustic cues underpinning suprasegmental perception (intonation, stress, and rhythm) to parse naturally connected speech into words before they encode phoneme and phoneme combinations unique to their language (Jusczyk et al., 1999; Seidl & Cristia, 2008; Seidl & Johnson, 2008; Swingley, 2009). Early perceptual skills, both segmental and suprasegmental, serve as building blocks for vocabulary and, ultimately, for word combinations and sentences (Werker & Yeung, 2005).

Recently, Davidson et al. (2019) examined the unique contributions of both segmental and suprasegmental speech perception to receptive vocabulary and language abilities of 117 pediatric CI recipients (5–9 years old) using bilateral CIs (BCIs) or bimodal devices. The segmental speech perception tasks included open-set word recognition in quiet and in noise as well as tests of phoneme discrimination. Tests of syllable stress discrimination, nonword stress imitation, talker discrimination, and emotion identification comprised the suprasegmental test battery. Using hierarchical multiple regression analyses with both segmental and suprasegmental speech perception scores as predictor variables, they found that segmental and suprasegmental speech perception skills contributed independently to receptive vocabulary performance and that suprasegmental skills contribute uniquely to receptive language performance. Child, audiological, and demographic variables that contributed to better segmental and suprasegmental speech perception skills were older age at test, higher nonverbal IQ, earlier receipt of a CI, and longer acoustic experience with an HA (coupled with CI use) for those with greater residual hearing. Depending on level of hearing, even children who used BCIs at the time of testing benefited from a period of bimodal device use. The potential predictive value of both segmental and suprasegmental speech perception for spoken language development in young DHH children motivates exploration of the unique contribution of SMD performance to each of these speech perception outcomes once the influence of demographic, cognitive, and audiological predictors has been accounted for.

The objectives of this study are as follows:

1. to compare the effect of age on SMD ability for CI recipients and their TH peers (cross-sectional data);

2. to identify demographic, cognitive, and audiological factors that are associated with SMD performance in pediatric CI recipients; and

3. to determine the unique contribution of SMD performance to segmental and suprasegmental speech perception scores after controlling for demographic, cognitive, and audiological characteristics.

Method

Participants

Participants included age-matched groups of children with CIs and TH children. Table 1 summarizes the distributions of self-reported education categories, gender, race, and ethnicity as well as age at test for the CI and TH groups. All children, in both groups, used spoken English as their primary language. None had diagnosed cognitive impairments. Maternal education was used as a sociodemographic variable, categorized by total years of reported education: high school (through 12 years), postsecondary (13–16 years), and postgraduate (> 16 years). No statistical differences in demographic characteristics between the groups were found for age, F(1, 140) = 1.86, p = .175; maternal education, χ²(2) = 2.81, p = .245; race, χ²(2) = 7.85, p = .097; and ethnicity, χ²(2) = 0.39, p = .823.

Table 1.

Demographics of children with cochlear implants (CIs) and typically hearing (TH) children.

Variable	M	SD	Range
Age at test (years), CI group (n = 104)	9.0	1.3	6.8–11.3
Age at test (years), TH group (n = 38)	8.7	1.3	6.9–11.0
	High school	College	Postgraduate
Maternal education, CI group	17%	45%	38%
Maternal education, TH group	16%	32%	53%
	Percentage		Percentage
Gender, CI group	48% Female		52% Male
Gender, TH group	50% Female		50% Male
Race, CI group (8% DNR)	68% Caucasian		24% Minority a
Race, TH group (5% DNR)	82% Caucasian		13% Minority b
Ethnicity, CI group (1% DNR)	89% Not Hispanic		10% Hispanic
Ethnicity, TH group	89% Not Hispanic		11% Hispanic

Note. DNR = did not report.

^a Minority: African American, 8%; Asian, 8%; and Multiracial, 8%.

^b Minority: Multiracial, 13%.

CI Participants

One hundred four children, ranging in age from 6 to 11 years, were recruited from multiple CI centers and schools across the United States as part of a longitudinal study examining the effects of early acoustic hearing on spoken language skills. The original sample included 117 children who were administered a test battery when they were between 5 and 9 years old (Davidson et al., 2019). The current sample consists of those children who returned 2 years later for several speech perception tests and the SMD test. All CI children used two devices, either bimodal devices or BCIs, received sequentially or simultaneously.

TH Participants

Thirty-eight age-matched TH peers served as controls for the SMD test.

Procedure

Test Protocols

The spectral modulation and speech perception tests described below were administered with an additional set of tests (e.g., standardized language and reading tests). Tests were grouped and arranged into 30-min sessions. At each test site, an individual participant performed the 30-min sessions in a different order than the other participants. A total of 3–3.5 hr of testing was given over a 5-hr span; two 30-min breaks and lunch filled the remaining time.

Spectral Modulation

Spectral modulation ability was assessed in both the CI and TH groups. Spectrally modulated noise stimuli were obtained from colleagues at Arizona State University (R.M.U. Spahr, personal communication, September 9, 2011) and are now clinically available as the EasyQSMD test (Gifford et al., 2014). The stimuli consisted of a four-octave-wide (350–5600 Hz) white noise with the application of a desired spectral modulation depth and frequency, and with a random starting phase in the spectral modulation. Inverse Fourier transform of the noise spectrum produced a waveform with the specified spectral shape. Stimuli had spectral modulation frequencies of 0.5 or 1.0 cycles/octave and modulation depths of 10, 11, 13, 14, or 16 dB. Reference stimuli were four-octave-wide noises with no spectral modulation. Each stimulus was 350 ms in length, and the interstimulus interval was 300 ms. Each trial was a three-interval, three-alternative forced-choice task, with one of the stimuli being a noise with spectral modulation whereas the other two were noises with no spectral modulation (i.e., spectrally unmodulated white noise). Both spectral modulation frequencies and all five modulation depths were mixed in a set of 60 trials. The SMD stimuli used in the test session were received prior to both the Gifford et al. (2014) publication and the availability of the EasyQSMD 2018 publication (with download). The spectral modulation frequencies (0.5 and 1.0 cycles/octave) and the spectral modulation depths (10, 11, 13, 14, and 16 dB) are the same as the QSMD. In addition, the number of stimuli (60) is the same as in the three-interval, three-alternative forced-choice method of constant stimuli that was employed. The stimuli in this study had interstimulus intervals of 300 ms, whereas the EasyQSMD interstimulus interval is 400 ms. However, the noise signals (both unmodulated and spectrally modulated) are 350 ms for both the stimuli reported here and the EasyQSMD stimuli.

The participant was instructed to choose the noise that sounded different from the other two (an “odd-man-out” scenario). Lists were presented in the sound field at 60 dBA, with the participant seated 1 m from the speaker. After as many as five practice trials, participants listened to the three noises in each trial (three horizontally situated blocks were labeled “1,” “2,” and “3” on the computer screen) and were instructed to point to the block representing the noise that sounded “different.” For each listener, SMD performance was determined by calculating the percent correct in 60 trials. Scores of < 44% correct are likely due to chance or guessing.

Audiological and Device Characteristics

Means and standard deviations of audiological characteristics (age at first HA, age at first CI, unaided pure-tone average [PTA], duration of acoustic experience, and age at second CI) of the CI recipients are shown in Table 2. Duration of acoustic experience was calculated for each child in the CI group: (a) for children using simultaneously implanted BCIs, age at BCIs minus age at first HAs; (b) for children using sequential BCIs, age at second CI minus age at first HAs; and (c) for children using bimodal devices, age at test minus age at first HAs. The unaided PTA associated with acoustic experience for each child was (a) the better ear pre-implant PTA for those receiving simultaneous BCIs, (b) pre-implant PTA for the second CI ear for those receiving sequential BCIs, and (c) PTA for the HA ear for those using bimodal devices.

Table 2.

Audiological characteristics of children (n = 104) with cochlear implants (CIs).

Variable	M	SD	Range
Age at first HA (months)	10.3	8.3	1–29
Age at first CI (years)	2.1	1.1	0.7–4.5
Unaided PTA a , b , c (dB HL)	92.4	19.6	30–125
Duration of acoustic experience (years), i.e., duration of HA use b , d	3.1	3.0	.03–11.0
Age at second CI (years), n = 83	2.9	1.7	0.7–7.7
Device configuration	Bimodal (21), sequential bilateral (64), simultaneous bilateral (19)
CI manufacturer	Cochlear (79), Advanced Bionics (22), MED-EL (3)

Note. HA = hearing aid; PTA = pure-tone average; BCI = bilateral cochlear implant.

^a PTA of 0.5, 1, and 2 kHz (dB HL) for the acoustic ear. The unaided PTA associated with acoustic experience for each child was the better ear pre-implant PTA for those receiving simultaneous BCIs, pre-implant PTA for the second CI ear for those receiving sequential BCIs, and PTA for the HA ear for those using bimodal devices.

^b For one participant who acquired meningitis at 20 months of age, the unaided PTA, 30 dB HL, is a time-weighted average of thresholds: within normal limits through 20 months and profound thresholds for the 1 month prior to receipt of BCIs. This participant's duration of acoustic experience represents the time prior to receipt of simultaneous BCIs (21 months).

^c The average unaided PTA for the bimodal group is 75 dB HL (SD = 16.6 dB, range: 40–100 dB HL).

^d Duration of acoustic experience was calculated from age at first HAs to age at BCIs for children using simultaneously implanted BCIs, age at second CI for children using sequential BCIs, and age at test for children using bimodal devices.

Speech Perception Assessment Battery (CI Group Only)

Segmental perception. The Lexical Neighborhood Test (LNT; Kirk et al., 1995), a monosyllabic open-set word identification test, was administered using prerecorded stimuli presented from a laptop computer to a loudspeaker placed at 0° azimuth approximately 1 m from the child. One 50-word list was presented at 60 dBA in multitalker noise (speech and noise at 0° azimuth) at a signal-to-noise ratio (SNR) of +8 dB. The 8 dB SNR was chosen to avoid floor and ceiling effects. The child repeated each word presented, and the percentage of correctly repeated words was calculated.

An adapted (and shortened) version of the Children's Test of Nonword Repetition (Dillon et al., 2004) was also used to assess segmental perception. The test list consists of 20 nonwords, five each with two, three, four, and five syllables. Digital recordings of nonwords were played in random order in the sound field. Children were told that they would hear a “funny word” and should repeat it back as best as they could. Their imitation responses were recorded via a head-mounted microphone connected to a digital recorder. Graduate students/clinicians in speech-language pathology transcribed each nonword production using the International Phonetic Alphabet. The production of each nonword was scored for vowels and consonants correct.

Suprasegmental perception. Emotion identification was assessed using prerecorded sentences spoken by a female talker with four emotional intentions: angry, scared, happy, and sad (Davidson et al., 2019; Geers et al., 2013). Thirty-six semantically neutral sentences with simple vocabulary were presented, and the child was asked to choose among the four emotions pictured on the computer screen. A percent-correct emotion identification score was calculated.

Talker discrimination of female voices was assessed using a subset of prerecorded sentences produced by eight talkers from the Indiana multitalker speech database (Bradlow et al., 1996; Karl & Pisoni, 1994). Two sentences, with different scripts, were presented, and the child was asked to determine whether the sentences were spoken by the same woman or by different women. A percent-correct score was calculated from a total of 32 sentence‐pair presentations.

Stress discrimination was assessed using pairs of bisyllabic nonwords (consonant–vowel–consonant–vowel; Thiessen & Saffran, 2007; Wenrich et al., 2017) with the same phonetic sequence (e.g., “diga”–“diga”) and the same total duration. The stress pattern of a nonword was either spondaic (equal stress on each syllable), trochaic (stress on the first syllable), or iambic (stress on the second syllable). Pairs of nonwords were presented with either the same stress pattern (spondee–spondee, trochee–trochee, and iamb–iamb) or different stress patterns (e.g., trochee–spondee and iamb–trochee). Syllabic stress was represented acoustically by elongated duration, increased intensity, and increased fundamental frequency (F0) relative to the unstressed syllable. After presentation of each pair of bisyllabic nonwords, the child reported whether the two nonwords had the “same” or “different” stress patterns. An overall percent-correct score was calculated for 30 trials. APEX3 software was used for administering the SMD task and the three suprasegmental tasks (Francart et al., 2008).

Cognitive Skills (CI Group Only)

Nonverbal cognition. A block design task, administered 2 years prior to the current test session (Davidson et al., 2019), was used to estimate nonverbal intelligence. Depending on the age of the child at the time of testing, the administered Block Design subtest was from either the Wechsler Preschool and Primary Scale of Intelligence–Fourth Edition (WPPSI-IV; Wechsler, 2012) or the Wechsler Abbreviated Scale of Intelligence–Second Edition (WASI-II; Wechsler, 2011). Children 6 years and older received the WASI-II, and those under 6 years old received the WPPSI-IV.

Memory. A digit span task, administered at the current test session, assessed verbal short-term memory (digits forward) and working memory (digits backward) using the Number Repetition subtest of the Clinical Evaluation of Language Fundamentals–Fourth Edition (Semel et al., 2004).

Regression models were used to examine the effects of various predictor variables on SMD performance and speech perception performance. For each regression model, predictors were added in stages so that the contribution of predictors above and beyond previously entered predictors could be determined. For SMD performance, the effects of group (TH vs. CI) and age at test were first added as main effects. Linear and quadratic effects were examined to support any curvilinear trends in the data. Interactions between group and age were added next to determine if the effects of age were similar for both groups. Regressions for SMD scores were conducted to identify demographic, cognitive, and audiological factors that are associated with SMD performance in CI recipients. Finally, regressions were used to determine the unique contribution of SMD scores to segmental and suprasegmental speech perception performance after controlling for demographic, cognitive, and audiological variables. This research was approved by the Human Studies Committee at Washington University School of Medicine (IRB ID 201305136).

Results

Objective 1. Compare the Effects of Age on SMD Ability in Children With CI and TH Children

From cross-sectional data, compare the effects of age on SMD ability in children with CI and TH children. The mean score for the CI group was 67% correct and ranged from 22% to 100%. The mean score for the TH group was 77% correct and ranged from 32% to 97%. The distributions of scores on the SMD test for the CI and TH groups are shown by the box plots in Figure 1.

Figure 1.

Box plots show spectral modulation detection (SMD) scores in percent correct for the cochlear implant (CI) and typically hearing (TH) groups. The limits of the box represent the lower and upper quartiles of the distribution (the difference is the interquartile range [IQR]), and the horizontal line through the box represents the median (median values are displayed next to the line). Whiskers represent the minimum and maximum scores in the distribution, excluding any outliers. Open-circle outliers are values between 1.5 and 3 IQRs from the end of a box.

Regression models estimated the effects of group (CI and TH) and age on SMD score. The coefficients, standard errors, and significance values appear in Table 3. At the first step, the main effects of age and group are statistically significant and account for 22% of the variability in SMD scores. As age increases, SMD performance increases (p < .001), and the CI group performed more poorly than the TH group (p < .001). At the second step, the interaction between the linear effects of age and group was not significant (variance of approximately 23%). In Step 3, when age quadratic effects are included, the predicted variance increases to 26%. The results are significant (p < .05), indicating that SMD performance increases with age, reaches a maximum, and plateaus. The interaction between group and quadratic effects of age was not significant. The lack of group-and-age interactions (linear or quadratic) indicates the effects of age are similar for both CI and TH groups. Figure 2 illustrates predictions from the parametric model for the CI and TH groups. The regression model produces predictions for the two groups that are close to their outcome means when age is near the average for each group. A 95% confidence band is also included for each group.

Table 3.

Regression coefficients for predicting spectral modulation detection scores.

Step	Variable	Coefficient
		Estimate	SE	t	Pr(> |t|)
1	(Intercept)	21.4	10.3	2.08	.04
1	Age	6.4	1.1	5.62	.001
1	Group (CI, TH)	−11.6	3.3	−3.54	.001
2	Age × Group (CI, TH)	1.3	2.6	0.49	.62
3	Age2	−2.5	1.0	−2.44	.02
4	Age2 × Group (CI, TH)	2.2	2.5	0.87	.38

Note. Step 1: F(2, 139) = 20.1, p < .001, multiple R ² = .22. Step 2: F(3, 138) = 13.4, p < .001, multiple R ² = .23. Step 3: F(4, 137) = 11.9, p < .001, multiple R ² = .26. Step 4: F(5, 136) = 9.65, p < .001, multiple R ² = .26. Bolded text indicates statistical significance at the .05 level or better. CI = cochlear implant; TH = typically hearing.

Figure 2.

Regression model predictions of spectral modulation detection percent correct as a function of age (at time of test) for the cochlear implant (CI; dashed line) and typically hearing (TH; solid line) groups. Shaded area represents 95% confidence intervals. Solid and dashed vertical lines indicate age at peak performance for the TH and CI groups, respectively.

The regression equations can be used to calculate the location of maximum performance for each curve (i.e., for each group) by finding the first derivative of the regression equation and then solving for a value of 0. Maximum performance is indicated for each group by a vertical line projecting to the age axis. The difference in location of maximum performance for the CI and TH groups was tested via bootstrapping and found not to be significant. The confidence intervals around these values are very wide, reflecting weak quadratic effects for each group. Consequently, the confidence interval for the location difference is also wide and includes the null value of 0. This is consistent with the regression models, which did not show a significant interaction between quadratic effects of age and group. The plateau in the curve for the CI group occurs at an age of 10.6 years; the plateau for the TH group occurs at an age of 9.6 years (see Figure 2). Thus, it appears that both groups reach a maximum in SMD performance with age, with maximum performance occurring for the CI group at a somewhat older age than the TH group. The difference between these age values, however, is not statistically significant.

Objective 2. Identify Demographic, Cognitive, and Audiological Factors That Are Associated With SMD Performance in CI Recipients

Cognitive test results are provided in Table 4. A multiple regression analysis assessed contributions to SMD performance from demographic, cognitive, and audiological factors. Age at test, nonverbal IQ (block design), forward and backward digit span, age at first CI, duration of acoustic experience, and unaided PTA in the acoustically aided ear accounted for 28% of the variance in SMD scores. The coefficients, standard errors, and significance values are shown in Table 5. Only age at test and nonverbal IQ (block design) were significant predictors of SMD performance. These results suggest that maturational and cognitive ability affect performance on the SMD task. Memory and audiological variables did not contribute significantly to SMD performance once age and cognitive level were controlled.

Table 4.

Results of cognitive tests for children with cochlear implants (n = 104).

Measure	M	SD	Range
CELF-4 Number Repetition, forward digit span a	8.2	2.3	0–15
CELF-4 Number Repetition, backward digit span a	3.8	1.8	0–9
CELF-4 Number Repetition, total score a	12.0	3.7	0–23
Block design (nonverbal IQ) b	99.7	16.3	60–145

Note. CELF‐4 = Clinical Evaluation of Language Fundamentals–Fourth Edition; IQ = intelligence quotient.

^a Raw scores.

^b M = 100, SD = 15.

Table 5.

Regression coefficients for predicting spectral modulation detection scores with demographic and cognitive variables, F(7, 95) = 5.21, p < .001, multiple R ² = .28.

Variable	Coefficient
	Estimate	SE	t	Pr(> |t|)
(Intercept)	−18.9	22.2	−0.85	.397
Age at test	6.63	1.48	4.48	.001
Block design (nonverbal IQ)	0.25	0.11	2.23	.028
CELF-4 Number Repetition, forward digit span	0.82	1.00	0.82	.416
CELF-4 Number Repetition, backward digit span	0.24	1.29	0.18	.855
Age at first CI	−0.25	2.14	−0.12	.907
Duration of acoustic experience	−1.04	0.76	−1.37	.174
Unaided PTA	−0.03	0.11	−0.24	.815

Note. Bolded text indicates statistical significance at the .05 level or better. IQ = intelligence quotient; CELF-4 = Clinical Evaluation of Language Fundamentals–Fourth Edition; CI = cochlear implant; PTA = pure-tone average.

Objective 3. Determine the Unique Contribution of SMD Scores to Segmental and Suprasegmental Speech Perception Performance After Controlling for Demographic, Cognitive, and Audiological Variables

Table 6 shows the CI group averages, standard deviations, and ranges for each of the speech perception tests (segmental: LNT in noise, nonword repetition—vowel, nonword repetition—consonant; suprasegmental: talker discrimination, stress discrimination, and emotion identification). Group average percent-correct scores ranged from 66% to 87% correct; notably, though, chance performance is 33% correct for emotion identification and 50% correct for both talker and stress discrimination.

Table 6.

Percent-correct scores for speech perception tests for children with cochlear implants (n = 104).

Variable	M	SD	Range
Segmental tests
LNT in noise	66.2	13.3	26–86
NWR vowels	78.8	12.3	37–93
NWR consonants	75.1	13.1	40–96
Suprasegmental tests
Talker discrimination	69.2	11.4	40–90
Stress discrimination	87.2	13.8	40–100
Emotion identification	77.3	19.5	19–100

Note. LNT = Lexical Neighborhood Test; NWR = nonword repetition.

Two composite scores, one each for segmental and suprasegmental speech perception, were created using principal components analysis (PCA). The use of composite scores reduces the number of potentially redundant statistical tests, reduces the probability of Type I errors, and increases the reliability of measures used in subsequent analyses. PCA scores are weighted combinations of outcome measures, with the weights being the eigenvectors from the orthogonal decomposition of the correlation matrix. Rescaling by the square root of the eigenvalues converts principal component scores to standard scores for use in further analyses. The segmental PCA score is a weighted combination of the LNT word recognition score in noise and vowel recognition and consonant recognition in nonwords. The loadings of these three measures on the first principal component were .75, .92, and .95, respectively, and accounted for 77% of the variance. The suprasegmental PCA score is a weighted combination of percent-correct scores for the talker discrimination, stress discrimination, and emotion identification tasks. The loadings for these three measures, on the first principal component, were .81, .84, and .75, respectively, and accounted for 64% of the variance. These two standardized PCA scores, with a mean of 0 and an SD of 1, are used to represent segmental and suprasegmental perception skills, respectively. The segmental and suprasegmental standard scores, for each CI participant, are shown in dot plots in Figure 3.

Figure 3.

Segmental and suprasegmental speech perception principal components analysis standard scores for participant children with cochlear implants. Segmental perception principal component (PC) is composed of Lexical Neighborhood Test percent correct in noise, nonword repetition consonants correct, and nonword repetition vowels correct. The suprasegmental perception PC is composed of talker discrimination, stress discrimination, and emotion identification.

The contribution of SMD performance to speech perception performance was examined using multiple regression after controlling for specific audiological and demographic variables that were identified previously (Davidson et al., 2019) as predicting speech perception performance. Regression analyses for the segmental and suprasegmental PCA scores were conducted in two steps. The first step entered those specific audiological and demographic variables found previously to contribute to speech perception ability. SMD performance was then entered at the second step to determine the amount of unique variance it contributed to speech perception. Tables 7 and 8 provide the coefficients, standard errors, and significance values for the regression models at each step for segmental and suprasegmental speech perception, respectively.

Table 7.

Linear regression models with segmental speech perception principal component score as the dependent variable.

Step	Variable	Coefficient
		Estimate	SE	t	Pr(> |t|)
1	(Intercept)	−2.01	1.16	−1.72	.088
1	Age	0.18	0.07	2.41	.018
1	Age at first CI	−0.28	0.12	−2.42	.018
1	Nonverbal IQ	0.02	0.01	2.77	.007
1	Duration of acoustic experience	0.00	0.04	0.01	.992
1	Unaided PTA	−0.01	0.01	−1.09	.280
2	(Intercept)	−1.70	1.08	−1.58	.118
2	Age	0.03	0.08	0.39	.696
2	Age at first CI	−0.26	0.11	−2.48	.015
2	Nonverbal IQ	0.01	0.01	1.85	.068
2	Duration of acoustic experience	0.02	0.04	0.50	.621
2	Unaided PTA	−0.01	0.01	−0.99	.326
2	SMD percent correct	0.02	0.01	4.15	.001

Note. Step 1: F(5, 94) = 4.71, p < .001, multiple R ² = .20. Step 2: F(6, 93) = 7.47, p < .001, multiple R ² = .33. CI = cochlear implant; IQ = intelligence quotient; PTA = pure-tone average; SMD = spectral modulation detection.

Table 8.

Linear regression models with suprasegmental speech perception principal component score as the dependent variable.

Step	Variable	Coefficient
		Estimate	SE	t	Pr(> |t|)
1	(Intercept)	−2.94	1.05	−2.80	.006
1	Age	0.34	0.07	4.99	.001
1	Age at first CI	−0.24	0.11	−2.23	.028
1	Nonverbal IQ	0.01	0.01	2.21	.030
1	Duration of acoustic experience	0.06	0.04	1.55	.124
1	Unaided PTA	−0.01	0.01	−1.96	.053
2	(Intercept)	−2.48	0.88	−2.83	.006
2	Age	0.14	0.06	2.21	.029
2	Age at first CI	−0.22	0.09	−2.49	.015
2	Nonverbal IQ	.004	.005	0.84	.402
2	Duration of acoustic experience	0.08	0.03	2.65	.010
2	Unaided PTA	−0.01	0.01	−2.10	.038
2	SMD percent correct	0.03	.004	6.69	.001

Note. Bolded text indicates statistical significance at the .05 level or better. Step 1: F(5, 97) = 8.48, p < .001, multiple R ² = .30. Step 2: F(6, 96) = 17.7, p < .001, multiple R ² = .53. CI = cochlear implant; IQ = intelligence quotient; PTA = pure-tone average; SMD = spectral modulation detection.

For segmental perception (see Table 7), age (at test), age at first CI, nonverbal IQ, duration of acoustic experience, and unaided PTA were entered at the first step. Together they accounted for 20% of total variance, with age (at test), age at first CI, and nonverbal IQ emerging as significant predictors. SMD score was added in the second step and contributed 13% of additional variance (33% total predicted variance).

The regression for suprasegmental perception (see Table 8) proceeded in the same manner with the same predictor variables. At the first step, the total explained variance was 30%. Age (at test), age at first CI, nonverbal IQ, and unaided PTA all predicted significant independent variance in the suprasegmental score. The addition of SMD scores, in the second step, contributed an independent additional 23% to the total variance for a total predicted variance of 53%.

Discussion

There has been increasing interest in using tasks of spectral resolution to estimate perceptual abilities of CI recipients. The fact that many of these tasks can be administered quickly and without linguistic stimuli holds appeal for both adult and pediatric CI recipients (and possibly pediatric HA recipients). These proposed attributes may be particularly beneficial for evaluating device benefit, monitoring progress, and determining whether changes to CI programs or HA adjustments improve performance. The clinical utility of these tasks for pediatric CI patients is dependent, in part, on understanding the developmental and cognitive mechanisms that affect performance. This study examined SMD performance by 6- to 11-year-old children with CIs to (a) compare their performance to that of their TH peers, (b) identify factors associated with their SMD performance, and (c) identify the unique contribution of SMD performance to segmental and suprasegmental speech perception.

Objective 1. Compare the Effects of Age on SMD Ability in Children With CI and TH Children

SMD performance of children in both the CI and TH groups improved with the child's age, over the 6- to 11-year-old range examined (cross-sectional results). Similar results, that is, spectral resolution performance is better for older children, are reported by others for TH children (Allen & Wightman, 1992; Kirby et al., 2015; Landsberger et al., 2018) and for children who are hard of hearing (Kirby et al., 2015). For children with CIs, our results align with those reported recently by Jeddi et al. (2019). SMRT performance of 75 children with unilateral CIs was weakly but significantly correlated (r = .243, p = .036) with age (range: 8–12 years; Jeddi et al., 2019). Our results contrast, however, with results from several other studies. In Horn et al. (2017), the SRD performance of 10 pediatric CI recipients (age range: 7–15 years) was similar to that of 15 adult CI recipients. In Gifford et al. (2018), the SMD performance of 36 pediatric CI recipients (age range: 5.6–17.9 years) was not significantly correlated with the child's age (r = .31, p = .06); however, they reported that 58% of the pediatric recipients scored at chance level (23%–43% correct). Using the same criteria, only 15% of the pediatric participants scored at chance level in this study. Thus, floor effects may be partially responsible for the lack of correlation between age and SMD performance in the Gifford et al. (2018) study. In Landsberger et al. (2018), the SMRT performance of 20 pediatric CI recipients (age range: 5.8–13 years) was not correlated with the child's age. Across these studies, differences in the type of spectral resolution task, the age range of the pediatric CI population, device configuration (unilateral vs. bilateral), and the number of participants likely contribute to the mixed results.

Overall, SMD performance by these pediatric CI recipients was poorer than that of their age-matched TH peers. This poorer performance is likely associated with the limited spectral resolution of current CI systems. Both the CI and TH participants, however, exhibited similar increases in SMD performance with age, and data from both groups indicate that maximum performance is reached around 10 years of age. Interestingly, in the studies that reported better spectral resolution performance by older TH children (Allen & Wightman, 1992; Kirby et al., 2015; Landsberger et al., 2018), the age at which performance reaches a peak or is adultlike is similar to that found here, namely, 9–10 years old. These results suggest that, for children with CIs less than 10 years old, performance on the SMD test depends on a child's ability to perceive differences in the stimuli and their ability to perform this type of task.

The SMD scores of these pediatric CI recipients (M = 67%, range: 22%–100%) are similar to those of adult postlingually deafened CI recipients (M = 61%, range: 20%–100%) and higher than those of pediatric CI recipients (M = 44.9%, range: 23%–87%) reported in Gifford et al. (2018). Notably, on average, the 104 pediatric CI recipients in this study were slightly younger (mean age at test = 9 years) and received their first CI(s) earlier (mean age at first CI = 2.1 years) than the 36 pediatric CI participants in the Gifford et al. (2018) study (mean age at test = 10.9 years, mean age at first CI = 4.6 years). Furthermore, the participants in this study listened to the spectral modulation stimuli with bilateral devices, either two CIs or bimodal devices, whereas participants in the Gifford et al. (2018) study were tested with only a unilateral CI. Early access to sound, via CIs, coupled with bilateral listening may have contributed to the current pediatric group performing similarly to adult CI recipients on the SMD test. Although a direct comparison of speech perception skills for the populations in these studies is not strictly valid due to the different speech tests that were administered, the average word recognition scores “in noise” for these pediatric CI participants (66%) were similar to the average word recognition scores “in quiet” for the adult and pediatric CI groups in the Gifford et al. (2018) study (52% and 61%, respectively).

Objective 2. Identify Demographic, Cognitive, and Audiological Factors That Are Associated With SMD Performance in CI Recipients

In addition to age at test, SMD performance in CI recipients was significantly associated with their nonverbal intelligence. Specifically, children's SMD performance is better with both age and cognitive ability. However, once the influence of these two variables was considered, variables associated with auditory development (age at first CI, unaided PTA, and duration of acoustic experience) and memory (forward and backward digit span) did not make significant independent contributions to SMD performance. Kirby et al. (2019) found that both nonverbal IQ and visual working memory were significantly positively correlated with pediatric HA recipients' performance on the SMRT. For their participants, measures of verbal working memory and executive function, however, did not correlate significantly with their SMRT performance. The authors conclude that the positive correlations between the two cognitive measures (nonverbal IQ and visual working memory) and spectral resolution tasks may reflect the overall cognitive demands of the task, rather than a direct cognitive influence on spectral processing. The results from this study of pediatric CI recipients support a similar interpretation.

The relation between audiological factors and spectral resolution performance is fairly consistent across studies of pediatric patients. In several studies (Gifford et al., 2018; Landsberger et al., 2018), earlier receipt of a CI was not associated with better SMD performance. This is consistent with our results: Age at first CI was not related to SMD performance. Interestingly, none of the audiological variables usually associated with better speech perception correlated significantly with better SMD performance in this study. SMD performance appears to depend more on physical and cognitive maturity than on audiological characteristics (e.g., age at first CI, unaided PTA, and duration of acoustic experience). It seems, then, that the audiological factors that influence speech perception are distinct from those that influence spectral resolution ability. Speech perception ability is influenced by early access to early spoken language (e.g., unaided PTA and age at first CI), whereas spectral resolution ability seemingly is not.

Objective 3: Determine the Unique Contribution of SMD Scores to Segmental and Suprasegmental Speech Perception Performance After Controlling for Demographic, Cognitive, and Audiological Variables

Results from this study of pediatric CI users showed that better performance on the SMD task, a psychophysical test of spectral resolution, is associated with better speech perception abilities. This result is similar to that from many adult CI studies (Anderson et al., 2012; Gifford et al., 2018; Henry & Turner, 2003; Henry et al., 2005; Saoji et al., 2009; Spahr et al., 2011) and from some pediatric CI studies (Jeddi et al., 2019; Jung et al., 2012). Unlike other studies, however, our speech perception battery included tests of both segmental and suprasegmental speech perception. Recently, in Davidson et al. (2019), these two types of speech perception were found to account for independent variance in vocabulary and language skills of a large number of pediatric CI recipients. Hence, it seemed worth determining the role, if any, of SMD performance to each of these two types of speech perception. In this study, measures of segmental perception included perception of words in noise and vowels and consonants in quiet. Suprasegmental perception included tests of talker discrimination, syllabic stress discrimination, and emotion identification. For each speech perception battery, a single standard score was created using PCA to represent performance. In both regression analyses, SMD scores added significant variance after accounting for factors previously shown to affect speech perception (i.e., age at test, age at first CI, nonverbal IQ, duration of acoustic hearing, and unaided PTA). Interestingly, SMD performance accounted for nearly twice the amount of unique variance for the suprasegmental PCA score (23%) than for the segmental PCA score (13%). Overall, the regression models accounted for more total variance in suprasegmental speech perception (53%) than in segmental speech perception (33%). It is unclear why SMD performance is more highly related to suprasegmental speech perception than to segmental speech perception. Suprasegmental speech perception (e.g., talker discrimination and emotion identification) is thought to rely primarily on perception of F0 (or “voice pitch” for speech). Notably, Davidson et al. found that, for those with better residual hearing, acoustic experience with an HA (in conjunction with a CI) for a period of time benefited both segmental and suprasegmental perception; however, the magnitude was greater for suprasegmental perception. This was the case even for those children who had moved from bimodal devices to BCIs. Examining the acoustic variables as described here allows for a more nuanced description of listening experience at each ear compared to describing participants solely by their currently used hearing devices (device groups). These SMD stimuli (with spectral modulation frequencies of 0.5 and 1 cycles/octave), however, have spectral variations that are very coarse and would seemingly be unrelated to the fine spectral differences present in the suprasegmental stimuli (e.g., in the talker discrimination task, the stimuli from the eight adult female talkers have average F0s that range from 163 to 237 Hz; much smaller than the 1.0 cycle/octave differences in the SMD stimuli).

Conclusions

The pediatric CI recipients in this study, who listened with two devices (BCIs or bimodal), performed as well on an SMD task as adult CI recipients with postlingual hearing loss listening with one CI (Gifford et al., 2018). Although these pediatric CI recipients, on average, did not perform as well as their TH peers, their slope of improvement in SMD with age did not differ from that of the TH group. Thus, based on our regression asymptotes, SMD scores obtained from children younger than approximately 10 years of age may underestimate mature performance. The significant association of nonverbal cognitive ability with SMD performance in pediatric CI recipients indicates that difficulties inherent in the task, as well as auditory capacity, affect performance. Higher scores on the SMD task predicted better speech perception scores, even after accounting for age at test. We observed a stronger association between SMD scores and suprasegmental perception than between SMD scores and segmental speech perception. This result suggests that SMD performance may also predict spoken language proficiency, given that suprasegmental perception plays a major role in spoken language development (Davidson et al., 2019). SMD scores may provide a useful nonlinguistic tool for predicting speech perception benefit from a sensory aid, with cautious interpretation based on the child's age and nonverbal cognitive ability.

Funding Statement

This research was supported by National Institute on Deafness and Other Communication Disorders Grant R01 DC012778 (Principal Investigator: L. S. Davidson).