Auditory perception is associated with implicit language learning and receptive language ability in autism spectrum disorder

Abstract

Background.

Autism spectrum disorder (ASD) is associated with language impairment as well as atypical auditory sensory processing. The current study investigated associations among auditory perception, implicit language learning and receptive language ability in youth with ASD.

Methods.

We measured auditory event related potentials (ERP) during an artificial language statistical learning task in 76 youth with ASD and 27 neurotypical (NT) controls. Participants with ASD had a broad range of cognitive and language abilities.

Results.

NT youth showed evidence of implicit learning via attenuated P1 amplitude in the left hemisphere. In contrast, among youth with ASD, implicit learning elicited bilateral attenuation that was increasingly evident with greater receptive language skill.

Conclusions.

Efficient early auditory perception reflects language learning and is a marker of language ability among youth with ASD. Atypical lateralization of word learning is evident in ASD across a broad range of receptive language abilities.

1. Introduction

Autism spectrum disorder (ASD) is a neurodevelopmental disorder associated with social communication deficits and restricted and repetitive behaviors. Individuals with ASD often show impairment in receptive/expressive language level and language pragmatics, as well as sensory-seeking or sensory-avoiding behaviors. Auditory processing, which provides a link between language comprehension and basic sensory perception, is well studied in ASD, with atypical responses apparent at both behavioral and neurophysiological levels (Kuhl, Coffey-Corina, Padden, & Dawson, 2005; Marco, Hinkley, Hill, & Nagarajan, 2011; Novick, Vaughan, Kurtzberg, & Simson, 1980). Atypical neurophysiological response to auditory stimuli among individuals with ASD has been recorded across multiple methodologies, including scalp electrophysiology (EEG and ERP) and functional neuroimaging (fMRI, MEG). However, the pattern of atypicality depends on the type and complexity of stimulus (O’Connor, 2012; Samson, Mottron, Jemel, Belin, & Ciocca, 2006). While some children with ASD appear to show enhanced discrimination of basic auditory input, such as pitch and musical tones (Heaton, Hudry, Ludlow, & Hill, 2008; Jones et al., 2009; Miller, 1989), there are conflicting results when complex acoustic stimuli are used. Some studies report delayed or impaired detection of novel stimuli (Cardy, Flagg, Roberts, & Roberts, 2005; Jansson-Verkasalo et al., 2003), others show faster latency to a mismatch negativity (MMN) ERP component reflecting auditory discrimination (Gomot, Giard, Adrien, Barthelemy, & Bruneau, 2002), and still others show no group differences (Čeponienė et al., 2003). These inconsistencies may reflect heterogeneity in the ASD phenotype, particularly as language ability has been correlated with amplitudes of early auditory perceptual ERP components in individuals with ASD (Bruneau, Bonnet-Brilhault, Gomot, Adrien, & Barthélémy, 2003; Kemner, Verbaten, Cuperus, Camfferman, & van Engeland, 1995; Lincoln, Courchesne, Harms, & Allen, 1995).

On the other hand, atypical neural activation is nearly always documented in ASD when speech-like stimuli are used (for review, see Groen, Zwiers, van der Gaag, & Buitelaar, 2008). Using ERP, basic auditory processing, including stimulus evaluation, is reflected in components that occur within the first 200 milliseconds after stimulus onset (e.g. N1, P1). Relative to neurotypical controls, youth with ASD show attenuated perceptual ERP amplitudes in response to auditory speech-like stimuli (i.e., vowels) but no difference in response to non-speech stimuli (i.e., tones) (Čeponienė et al., 2003; Whitehouse & Bishop, 2008). Interestingly, among high functioning youth with ASD, P1 amplitude to speech sounds may be moderated by active attention to stimuli (Whitehouse & Bishop, 2008). These studies have not reported group differences in latency of early perceptual ERPs, suggesting ASD is associated with reduced early perceptual attention to linguistic stimuli that is not due strictly to processing speed or orienting. Altogether, the existing literature is somewhat mixed in that some studies support a bottom-up process implicating basic auditory cortical dysfunction (Boddaert et al., 2003; Boddaert et al., 2004; Frith & Happé, 1994; Siegal & Blades, 2003) and others support a top-down hypothesis implicating attentional control over early auditory perception(Čeponienė et al., 2003; Haesen, Boets, & Wagemans, 2011; Whitehouse & Bishop, 2008). In both scenarios, atypical perceptual attention has downstream effects on higher-order processes, such as language encoding and comprehension, which could explain some functional language differences in ASD.

Statistical learning is an example of a higher order language process that is reliant on earlier auditory perception (Emberson, Liu, & Zevin, 2013). Statistical learning facilitates language learning in the natural environment through probabilistic parsing of speech components to learn and comprehend spoken words (Saffran, 2003). Statistical learning deficits have been documented among individuals with ASD (Scott-Van Zeeland et al., 2010), but may be due to the linguistic stimuli rather than statistical learning broadly (for review, see Eigsti & Mayo, 2011), as decreased attention to verbal stimuli has been posited to explain deficits in linguistic statistical learning tasks (Arunachalam & Luyster, 2016). A major limitation to research on statistical learning in ASD has been the reliance on behavioral response to indicate learning, which involves comprehension of directions and working memory skills, which may be deficient in children with ASD and severe verbal impairment. This group makes up 30% of the ASD population, but is highly underrepresented in the extant literature (Tager-Flusberg & Kasari, 2013).

The P1 auditory ERP component presents an opportunity to evaluate neural response at the intersection between auditory perception and speech processing among youth with ASD, given its early timing (approximately 40 – 150 ms), predominance in children (Sharma, Kraus, McGee, & Nicol, 1997), observability during passive tasks (Sharma et al., 1997), and specificity to speech stimuli (Dehaene-Lambertz et al., 2005). Among neurotypical individuals, P1 amplitude and latency decrease with age, reflecting increased efficiency of speech processing (Courchesne, 1990; Sharma et al., 1997). A decrease in left-hemisphere fronto-parietal perceptual ERP amplitudes and latencies has also been shown in response to known speech versus non-speech or unfamiliar speech in typically developing adults (Dehaene-Lambertz et al., 2005). Likewise, decreased P1 amplitude has been found in response to known versus nonsense words in typically developing children (Mills et al., 2006). Given the negative correlation between P1 amplitude and age, attenuated P1 amplitude in these studies may indicate more efficient processing of familiar linguistic stimuli, as well as greater perceptual attention to unfamiliar speech stimuli.

The origin of the P1 has been hypothesized as the superior temporal gyrus (Eggermont & Ponton, 2002), which includes primary auditory cortex as well as Wernicke’s area, thus also marking a neuroanatomical intersection of lower- and higher-order speech perception. Prior studies of neurotypical individuals have generally focused on later ERP components as markers of statistical learning, such as the N400 (Sanders, Newport, & Neville, 2002) and MMN (Bonte, Mitterer, Zellagui, Poelmans, & Blomert, 2005). However, Sanders and colleagues did report an effect of statistical language learning on the N100 component in adults, which may be a developmental correlate of the P1 in children and adolescents (Sharma et al., 1997). Thus, there is reason to believe statistical learning can be detected at the early auditory perceptual phase. However, it remains unclear how early speech perception relates to implicit learning and receptive language abilities in youth with ASD.

The current study aimed to evaluate the associations among early auditory perceptual response, statistical learning for linguistic stimuli, and receptive language abilities in youth with ASD. Using ERP and a well-known artificial language task, we compared electrophysiological evidence of implicit language learning at the perceptual level among youth with ASD to neurotypical (NT) youth. Because the focus of our investigation was on the association between perceptual processing and language in the clinical group, we conducted the remaining analyses only with the ASD group. Thus, within our sample of ASD youth (who had a broad range of language and cognitive abilities), we next examined associations among early auditory perception, statistical learning, and receptive language. Our approach addressed two major limitations in prior ASD research: first, we eliminated confounding factors such as lexical store and semantic knowledge by using an artificial language paradigm to measure perceptual discrimination of familiar auditory stimuli; second, we were able to include individuals whose low cognitive and language abilities would have precluded an explicit behavioral response, by measuring implicit learning via neurophysiological parameters.

2. Methods and Materials

2.1. Participants

Participants with ASD were drawn from a sample of 113 youth with a previous diagnosis of ASD. Of these, a portion were excluded from the current study due to inability to tolerate the net (n = 4) or behaviors during testing that decreased EEG quality (n=33; e.g., excessive movement, talking, and/or insufficient artifact-free ERP trials). Participants with little or no ERP data tended to have lower cognitive skills (mean nonverbal IQ = 69) compared to participants who were ultimately included in the analyses. The final sample included 76 youth with ASD (ages 4-17 years) as well as 27 youth with neurotypical development (NT) and comparable age (t[101] = 1.418, p = .159) and gender (χ²[1] = 3.244, p = .072). All participants were from primarily English-speaking households. Youth with NT had no history of developmental disability and no immediate family member with an ASD diagnosis. Autism spectrum disorder diagnoses based on clinical criteria from the Diagnostic and Statistical Manual of Mental Disorders, 5^th Edition (DSM-5) (Association, 2013) were confirmed by a licensed psychologist following gold standard ASD evaluation that included the Autism Diagnostic Interview, Revised (ADI-R) (Lord, Rutter, & Le Couteur, 1994), the Autism Diagnostic Observation Schedule, 2^nd Edition (ADOS-2) (Lord et al., 2012), and review of the participant’s medical history. Participant demographics are presented in Table 1.

Table 1.

Participant Demographics

	NT	ASD	Difference Testing
N	27	76
Age in years	13 (2.3)	12.12 (2.9)	t(101) = 1.42, p = .159
Female n (%)	10 (37%)	15 (19.7%)	Χ2(1) = 3.24, p = .072
Nonverbal IQ	113.59 (15.99)	88.45 (28.55)	t(81.87) = 5.60, p < .001α
Receptive Language (PPVT-4 Standard Score)	119.81 (10.32)	96.11 (27.165)	t(100.48) = 6.42, p < .001α
ADOS-2 Module Distribution (Module 1/2/3/4)	n/a	7%/ 4%/ 88%/ 1%	n/a
ADOS-2 Total Calibrated Severity Score	n/a	7.50 (1.72)	n/a
Learned ERP Trials	32.15 (7.95)	29.99 (8.68)	t(101) = 1.14, p = .259
Unlearned ERP Trials	31.70 (8.50)	30.74 (9.03)	t(101) = 0.49, p = .629

Note: Values are means with standard deviations in parentheses. ERP trials represent the number of artifact-free trials available after processing for creation of individual subject averages.

^αEqual variances not assumed.

2.2. Procedures

Prior to testing, participants and caregivers provided written consent or verbal assent, as developmentally appropriate. Cognitive, behavioral and electrophysiological testing took place in the laboratory over the course of one or two days.

2.2.1. Cognitive and behavioral measurement.

Receptive language abilities of all participants were measured using the Peabody Picture Vocabulary Test, 4^th Edition (PPVT-4) (Dunn & Dunn, 2007), which provides an age-standardized score (M = 100, SD = 15). Nonverbal IQ (NVIQ) was derived from the Differential Abilities Scale, 2^nd Edition (DAS-II) (Elliott, 2007). When the participant’s cognitive abilities were below the expected testing range, a ratio IQ was calculated from age equivalencies on nonverbal subtests. ASD participants had a wide range of nonverbal intellectual ability (range = 19 to 154) and receptive language skill (range = 20 to 139). Compared to the NT group, ASD participants had significantly lower NVIQ (t[81] = 5.596, p < .001, equal variances not assumed) and receptive language skills (t[100] = 6.416, p <.001, equal variances not assumed).

2.2.2. EEG word segmentation paradigm.

The current study employed a passive, auditory statistical learning task to evaluate segmentation of tri-syllabic, artificial, unstressed (i.e. lacking prosodic cues) nonsense word combinations (McNealy, Mazziotta, & Dapretto, 2010, 2011; Saffran, Aslin, & Newport, 1996; Scott-Van Zeeland et al., 2010). We specifically selected a task that did not require an explicit behavioral response (e.g. pressing a button in response to “learned” words) so that we could include individuals with very low language skills and without prior linguistic knowledge. However, prior research indicates that neurotypical youth and youth with high functioning ASD are equally successful with this task (Mayo & Eigsti, 2012). Participants viewed a static line drawing of a robot at visual angle 10.6° horizontal and 11.3° vertical, which remained on the screen throughout all experimental phases. In the resting phase (part 1), participants were instructed to “study the picture” for one minute. In the exposure phase (part 2), participants were told, “it wants to talk to you,” and were subsequently exposed to two minutes of a continuous stream of English language syllables (e.g. tutibupabikudogala) identical to those in McNealy et al. (2011). The syllable stream was organized pseudo-randomly such that tri-syllabic combinations were statistically related via transitional probabilities of 1.0 within “learned words” and 0.33 within “unlearned words.” In this way, four learned combinations were randomly embedded 45 times each, and 12 unlearned combinations were randomly embedded 15 times each. None of the unlearned words began with the same syllable as one of the learned words; thus, the learned words could theoretically be distinguished at word onset. No prosodic cues were provided to aide segmentation. In the test phase (part 3), the participant was told, “it has more to say,” followed by presentation of four continuous cycles of 12 trisyllabic learned and 12 trisyllabic unlearned words presented pseudo-randomly for a duration of four minutes. The words were separated by a random jitter of between 500-750 milliseconds of silence. Syllable length ranged from 245 to 287 milliseconds each. Participants were not given explicit instructions other than to remain still and attentive to the robot image on screen, nor were they asked to provide a response at any stage of the experiment. Analyses were restricted to the final, test phase of the experiment.

2.2.3. EEG recording.

Continuous EEG was acquired from a high-density 128-channel GSN Hydrocel net using Net Station 4.5.6 software integrated with a 400-series high-impedance amplifier (Electric Geodesics Inc, EGI, Eugene, OR). Electrode impedances were below 50 kOhms to maximize signal-to-noise ratio, within the standard range for high-impedance amplifiers. During acquisition, EEG signals were referenced to the vertex electrode, analog filtered (0.1 Hz high-pass, 100 Hz elliptical low-pass), amplified, and digitized with a sampling rate of 1000 Hz. Researchers observed and marked periods of off-task behavior or behaviors interfering with the EEG signal (e.g., external noise such as talking).

2.2.4. EEG pre-processing.

Continuous EEG was filtered offline between 0.1-40 Hz and the test phase was segmented into epochs that contained a 100 ms baseline period pre-stimulus (i.e. start of the trisyllabic word) onset and 1000 ms post-stimulus onset. Segmentation accounted for computer timing offsets (measured monthly) and digital finite impulse response filters. Epochs marked during acquisition as contaminated by movement or improper attention were removed. Automatic artifact detection rejected channels containing voltage shifts greater than 100 μV for each trial. If a channel was rejected for more than 50% of epochs, it was excluded from analysis. Automated rejection decisions were reviewed and verified manually by trained experimenters before bad channel rejection. Epochs were re-referenced to the average reference, baseline corrected, and lastly, averaged across epochs within individual. Participants were included if they provided greater than 14 artifact-free epochs per condition in the test phase of the experiment. ASD and NT groups did not differ on number of artifact-free trials.

Based upon previous literature (Friederici, 2005; Junge, Kooijman, Hagoort, & Cutler, 2012; O’Connor, 2012; Sanders et al., 2002; Sharma et al., 1997) and visual inspection of the grand-averaged waveforms across ten a priori identified scalp electrode clusters (see Supplementary Materials), peak amplitude and latency were extracted for the P1 component between 40 and 120 milliseconds within left (EGI GSN Hydrocel electrode sensor #20, 23, 24, 27 and 28) and right (EGI GSN Hydrocel electrode sensors #3, 117, 118, 123 and 124) frontal regions of interest (ROI). Examination of the topographic plots indicated our a priori selected ROIs captured the P1 and were consistent across groups (Figure 1A). A late, slow wave component was also apparent between 200 and 1000 milliseconds in these ROIs. However, because the focus of our study was on early perceptual response, and this late wave was confounded by the onset of the second and third syllables in the trisyllabic words, we elected to evaluate only the P1 in the current study.

Figure 1.

Topographic ERP plots indicating the location of the P1 wave in NT and ASD youth at 60-110 ms, by stimulus condition. The plots demonstrate comparable ROIs, as well as a pattern of asymmetry to learned words among the NT controls. NT=neurotypical. ASD=autism spectrum disorder.

2.3. Analyses

All analyses were done in IBM SPSS version 19. Effects of group, condition and ROI on ERP components were evaluated with repeated measures ANCOVAs. Age was included a covariate given its prior association with P1 amplitude. Sex was not included in the models due to lack of main effect (p = .493) or theoretical basis for inclusion (Cunningham, Nicol, Zecker, & Kraus, 2000; Russo, Zecker, Trommer, Chen, & Kraus, 2009). To present the results in a concise and clear manner, we describe the results of all main effects regardless of statistical significance, but only interaction effects that are statistically significant. We report Greenhouse-Geisser corrected values and conducted LSD-corrected post-hoc analyses where indicated. Continuous associations between receptive language skill and ERP values were evaluated using linear regressions. Significant contributions of independent variables in the regression models were evaluated with R² change statistics.

3. Results

3.1. Group differences in implicit learning

A summary of P1 values by group and ROI are presented in Table 2. Our first analyses tested for effects of ASD diagnosis on P1 amplitude to learned versus unlearned words. We conducted a 2x2x2 group (NT, ASD) by condition (learned, unlearned words) by frontal ROI (right, left) repeated measures ANCOVA with age as a covariate.

Table 2.

P1 Maximum Amplitude Values by Group (μV)

	ASD		NT		Total
	M	SD	M	SD	M	SD
Left Frontal ROI
Learned	3.25	2.02	2.79	1.55	3.13	1.91
Unlearned	3.03	2.31	3.41	1.62	3.13	2.15
Right Frontal ROI
Learned	3.14	1.82	3.66	1.65	3.28	1.78
Unlearned	3.27	2.11	2.85	1.38	3.16	1.95

3.1.1. P1 Amplitude.

Results indicated no main effects of condition, ROI, or group (p’s > .590; η_p²’s < .01) on P1 maximum amplitude. There was a significant three-way interaction between condition, ROI and group (F[1,100] = 8.491, p = .004, η_p² = .078). Post hoc pairwise comparisons revealed that among the NT group only, there was evidence of implicit learning as reflected by unique patterns of ROI laterality for each condition: P1 amplitude was smaller in the left relative to right frontal ROI to learned words only (F[1,100] = 5.172, p = .025; η_p² = .049). This effect was not evident among the ASD group (F[1,100] = 2.015, p = .159; η_p² = .020; See Figure 2). Within hemisphere, although the direction of condition effects was opposite for each group (i.e. in the right hemisphere, the ASD group had greater P1 amplitude to learned words while the NT group had greater P1 amplitude to unlearned words, and vice versa in the left hemisphere), condition differences within group and hemisphere were not statistically significant (p’s > .089). Likewise, within hemisphere, there were no overall group differences in P1 amplitude collapsed across condition (p’s >.600). There was also a main effect of age (F[1,100] = 8.289, p = .005; η_p² = .077), reflecting attenuation of P1 amplitude with greater age( r = −.275, p = .005), consistent with previous reports (Sharma et al., 1997; Courchesne, 1990).

Figure 2.

A) Left frontal and right frontal waveforms by diagnostic group. B) Implicit learning among the NT group was evidenced by smaller P1 amplitude to learned words in the left versus right ROI (*p = .025). No differences were apparent in the ASD group.

3.1.2. P1 Latency.

Results indicated no main or interaction effects of condition, ROI, or group on P1 latency (p’s > .280, η_p²’s < .013). There was a main effect of age (F[1,100] = 6.614, p = .012; η_p² = .062) reflecting a negative association between age and average P1 latency (r = −.262, p = .008).

3.2. Receptive language and implicit learning in ASD

Next, we tested for a moderating effect of language ability on implicit learning within our ASD group. We conducted 2x2 (condition x ROI) repeated measures ANCOVAs with both age and receptive language skill (PPVT-4 standard score) as covariates predicting P1 amplitude and latency.

3.2.1. P1 amplitude.

There were no main effects of condition, ROI or receptive language on P1 amplitude among youth with ASD (p’s ≥ .080; η_p²’s < .017). However, there were significant interactions between condition and receptive language skill (F[1,73] = 4.333, p = .041; η_p² = .056) and between ROI and receptive language skill (F[1,73] = 5.818, p = .018; η_p² = .074), suggesting both implicit learning and laterality effects associated with P1 amplitude. We further investigated these interactions using linear regression models, described below. Additionally, there was a main effect of age (F[1,73] = 10.037, p = .002; η_p² = .121) reflecting a negative association between P1 amplitude and age within the ASD group (r = −.346, p = .002).

3.2.2. P1 latency.

Results indicated no main or interaction effects of condition, ROI, or group on P1 latency (p’s > .310, η_p²’s < .013). There was a marginal effect of age (F[1,73] = 3.175, p = .079) reflecting a trend toward faster P1 latency with greater age (r = −.210, p = .071).

3.3. Receptive language effects in ASD

We examined associations between auditory perception and receptive language ability using linear regressions predicting differences in P1 amplitude across condition (learned – unlearned) and ROI (left – right) separately. In the first analyses evaluating condition effects, we averaged P1 amplitude across ROIs. In the second analyses examining ROI effects, we averaged P1 amplitude across conditions. The independent variables entered one at a time in the model were age, NVIQ, ADOS-2 Social Affect Calibrated Severity Score (SA-CSS), and receptive language (PPVT-4 standard score). We included NVIQ to test the alternative hypothesis that differences in P1 amplitude were more strongly related to nonverbal cognition than language, given that our ERP paradigm was a statistical learning task. We included the ADOS-2 SA-CSS to test for effects of social impairment on statistical language learning.

3.3.1. Condition.

Higher receptive language was associated with attenuated P1 to learned words (B = −.028, SE = .013, p = .043; ΔR² = .056) over and above age (B = .010, SE = .082, p = .903; ΔR² = .001), NVIQ (B = .011, SE = .013, p = .370; ΔR² = .012) and SA-CSS (B = −.041, SE = .155, p = .791; ΔR² = .000). See Figure 3.

Figure 3.

Within the ASD group, receptive language ability was related to decreased P1 amplitude to learned words, averaged across both left and right frontal ROIs.

3.3.2. ROI.

Receptive language was not associated with a difference in P1 amplitude across ROIs (B = −.013, SE = .011, p = .229; ΔR² = .019) over and above the non-significant contributions of age (B = −.020, SE = .067, p = .770; ΔR² = .002), NVIQ (B = −.003, SE = .010, p = .773; ΔR² = .051), and SA-CSS (B = .067, SE = .127, p = .598; ΔR² = .008).

Results of both regression analyses were comparable when we excluded four outlying individuals with PPVT standard scores < 30.

4. Discussion

4.1. Statistical Learning

The NT group showed an effect of implicit learning as reflected by unique ROI laterality patterns, with a smaller left frontal P1 amplitude response to learned words relative to the right ROI. This is consistent with previous literature suggesting more efficient processing of familiar acoustic speech sounds in NT samples. Evidence for implicit learning at this early stage of perception implies a close link between lower-order auditory processing and speech language. Mills et al. (2006) suggested this pattern reflects increased attention to the less familiar stimuli. Alternately, the consistent negative association between P1 and age across all analyses supports prior assertions that the P1 component reflects a neural process that becomes more efficient (i.e. attenuates) with development and/or language experience (Courchesne, 1990; Sharma et al., 1997).

Accordingly, within our ASD group we also found an association between greater receptive language ability and smaller P1 amplitude to learned versus unlearned words. The association was robust to the contributions of age, nonverbal cognition and ASD symptom severity in the regression model, underscoring the specificity of auditory processing to receptive language skill. This pattern suggests that implicit learning did occur in youth with ASD who had average or high receptive language skills. Our results are consistent with a previous behavioral study using the same paradigm (Mayo & Eigsti, 2012) as well as a functional MRI study which found a negative association between communication deficits and neurobiological evidence of word learning(Scott-Van Zeeland et al., 2010).

Basic auditory sensory processing deficits are implicated in the etiology of language difficulties associated with ASD (Marco et al., 2011), but the direction of causality remains unclear. The bottom-up theory (Emberson et al., 2013; Groen et al., 2008) posits auditory perception impairments drive deficits in later encoding and semantic mapping. This would also be consistent with heightened sensitivity to pitch discrimination (Heaton et al., 2008) and behavioral sensitivity to noise described in earlier research (Jones et al., 2009). However, this does not rule out an opposite or bidirectional effect wherein familiarity with language also facilitates basic auditory processing of familiar words (Conway, Bauernschmidt, Huang, & Pisoni, 2010). Although our study did not explicitly test whether differences in ASD are due to atypical cortical processing versus reduced attention to speech stimuli, the lack of group differences in P1 amplitude overall is not supportive of the latter hypothesis. The most likely scenario includes bidirectional effects between brain development and experience. In other words, language impairment in ASD may be attributed to inefficient auditory perception that interferes with encoding, as well as disruptions in later recognition that moderate the perceptual phase. Given the heterogeneity of the ASD phenotype, we would further suggest individual differences in the strengths of these effects. The design of our task precluded dissociation between perception of the 2^nd and 3^rd syllable onsets and later encoding of the words; however,future work should attempt to estimate unique contributions of each phase to language acquisition and skill.

4.2. Lateralization of Language Processing

One of the most robust findings in cognitive neuroscience is specialization of the left-hemisphere of the brain for language processes (Casey, Galvan, & Hare, 2005; Durston & Casey, 2006; Eyler, Pierce, & Courchesne, 2012). Unlike the youth with NT, learning was reflected bilaterally across left and right ROIs among youth with ASD in our study. These results are consistent with prior literature suggesting individuals with ASD do not show left-hemisphere dominance for language tasks (Dawson, Finley, Phillips, & Galpert, 1986; Eyler et al., 2012; Herringshaw, Ammons, DeRamus, & Kana, 2016; Kana & Wadsworth, 2012; Mason, Williams, Kana, Minshew, & Just, 2008). Atypical lateralization of receptive language is evident in ASD very early in life (Finch, Seery, Talbott, Nelson, & Tager-Flusberg, 2017). Among neurotypical individuals, broader activation across brain hemispheres in response to speech may indicate lack of familiarity with language and/or immature language circuitry (Friedrich & Friederici, 2008, 2010; McNealy, Mazziotta, & Dapretto, 2006; McNealy et al., 2011; Spironelli & Angrilli, 2009). In contrast, bilateral activation in ASD has been proposed as a compensatory mechanism facilitating effective language processing in ASD (for meta-analysis, see Herringshaw et al., 2016). Our results do not support that hypothesis, however, given lack of significant association between receptive language skill and laterality in the linear regression analyses. The current study adds to the extant literature by documenting lack of left hemisphere specialization 1) during implicit language learning, 2) at the very early perceptual phase of linguistic processing, and 3) among an ASD cohort who had a broad range of receptive language ability.

Although atypical lateralization is often conceptualized as an endophenotype mediating the association between genetic expression and language ability (Annett, 1985) there is more likely a bidirectional relation between language behaviors and the organization of neural networks associated with language processing (Bishop, 2013; Conboy & Mills, 2006; Mills et al., 2006). If increased lateralization of language cortex depends on language use and exposure, then symptoms of ASD, including reduced social drive and/or attention to language, would also contribute to atypical asymmetry. Given documented bilateral attenuation of the P1 in response to non-speech auditory stimuli in NT adults (Dehaene-Lambertz et al., 2005), our results could reflect a pattern wherein youth with ASD did not recognize the artificial language as speech, but rather as nonverbal acoustic stimuli. However, in our study we did not find an association between P1 laterality and social communication deficits as measured by the ADOS-2, suggesting that decreased social attention did not contribute to ERP results. Future work should aim to further investigate the effect of social motivation on statistical learning in youth with ASD using experimental methods described by Kuhl and colleagues (2005).

Animal research suggests a unidirectional bias in cortical compensation wherein the right hemisphere is better equipped to compensate for an early lesion in the left hemisphere (Morgan & Corballis, 1978), which makes language a particularly strong candidate for lateralization differences in ASD. One hypothesis that derives from our work is that if individuals with ASD engage broader neural circuitry to process language, this could create a resource deficit that contributes to other language abnormalities, such as odd speech prosody and poor integration of nonverbal communication. This is partly supported by the finding that individuals with ASD have more difficulty comprehending spoken language when it is accompanied by nonverbal gesturing, presumably due to the higher cognitive load involved in multimodal integration (Silverman, Bennetto, Campana, & Tanenhaus, 2010).

Several alternate explanations for our results are possible. First, prior literature has pointed to increased incidence of left- or ambidextrous-handedness in ASD as an indication of broader atypical cortical organization. However, evidence to support the link between handedness and language deficits is both inconsistent and weak (Escalante-Mead, Minshew, & Sweeney, 2003; Finch et al., 2017; Soper et al., 1986). Second, statistical learning itself may also be a lateralized process. For example, visual statistical learning tasks have been shown to predominantly engage the right hemisphere (Rauch et al., 1997; Roser, Fiser, Aslin, & Gazzaniga, 2011). However, within the ASD group, we found a specific association between receptive language ability and P1 attenuation, over and above NVIQ. Thus, it is unlikely our results reflect domain-general statistical learning differences between NT and ASD youth. Third, prior studies suggest that although individuals with ASD can be successful with statistical language learning, they may require a longer exposure period than we used in the current study (Evans, Saffran, & Robe-Torres, 2009; Mayo & Eigsti, 2012). Finally, individuals who have language impairment or developmental dyslexia (but no ASD) have also generally shown a pattern of reduced left-hemisphere lateralization of language processes in functional neuroimaging studies (Badcock, Bishop, Hardiman, Barry, & Watkins, 2012; De Guibert et al., 2011; Whitehouse & Bishop, 2008). Therefore, language impairment, rather than ASD, may be the primary explanation for our results. This final alternative explanation cannot be ruled out in our study and presents a possible future direction via evaluation of youth with specific language impairment or NT youth with a greater range of language abilities.

Our study was unique in that we recruited a large sample of ASD participants with a broad range of intellectual and language abilities. Despite the high rates of severe language disabilities in ASD, few studies have examined neurophysiological correlates of language processing in individuals with ASD and very low language skills (for an exception, see Cantiani et al.Cantiani et al., 2016). Our EEG task was designed to accommodate this wide range of functioning in that it did not rely on any previous lexical or semantic knowledge, nor did it require a response that was dependent on motoric or working memory abilities. Similar strategies with neurophysiological methods will facilitate better understanding of perceptual and linguistic neural circuitry in individuals with ASD and severe language and cognitive impairment, including genetic subtypes of ASD associated with intellectual disability (Earl et al., 2017).

5. Conclusions

Language learning is dependent on basic auditory perception as well as higher order language processes, such as word segmentation, encoding and recognition. Individuals with ASD may have atypical organization of neural networks which could contribute to related language deficits, such as difficulty with speech pragmatics. Our results support a bidirectional brain-behavior association such that auditory perception influences language learning, while language use and experience in turn facilitate efficient perceptual processing.