Lack of neural evidence for implicit language learning in 9-month-old infants at high risk for autism

Abstract

Word segmentation is a fundamental aspect of language learning, since identification of word boundaries in continuous speech must occur before the acquisition of word meanings can take place. We previously used functional magnetic resonance imaging (fMRI) to show that youth with autism spectrum disorder (ASD) are less sensitive to statistical and speech cues that guide implicit word segmentation. However, little is known about the neural mechanisms underlying this process during infancy and how this may be associated with ASD risk. Here, we examined early neural signatures of language-related learning in 9-month-old infants at high (HR) and low familial risk (LR) for ASD. During natural sleep, infants underwent fMRI while passively listening to three speech streams containing strong statistical and prosodic cues, strong statistical cues only, or minimal statistical cues to word boundaries. Compared to HR infants, LR infants showed greater activity in the left amygdala for the speech stream containing statistical and prosodic cues. While listening to this same speech stream, LR infants also showed more learning-related signal increases in left temporal regions as well as increasing functional connectivity between bilateral primary auditory cortex and right anterior insula. Importantly, learning-related signal increases at 9 months positively correlated with expressive language outcome at 36 months in both groups. In the HR group, greater signal increases were additionally associated with less severe ASD symptomatology at 36 months. These findings suggest that early differences in the neural networks underlying language learning may predict subsequent language development and altered trajectories associated with ASD risk.

1 |. INTRODUCTION

Autism spectrum disorder (ASD) is a neurodevelopmental disorder characterized by impaired social communication and interaction, as well as restrictive behaviors and sensory processing atypicalities. While children with ASD often show heterogeneous language profiles, it is estimated that over half of the children with ASD have persisting language impairments throughout their lifespan (using MCFLIRT; Sperdin & Schaer, 2016). Indeed, delays in language development are one of the earliest parent-reported concerns related to ASD (Tager-Flusberg et al., 2005) and are considered to be an early behavioral marker of ASD risk (Zwaigenbaum et al., 2005). Affected children often exhibit deficits in both expressive and receptive language skills (Hudry et al., 2010; Kamio et al., 2007) that are predictive of future prognosis. Early language profiles have been shown to be related to ASD symptom severity during childhood (Gotham et al., 2012; Longard et al., 2017) and to predict future verbal as well as cognitive development in young children with ASD (Franchini et al., 2018). Prior studies have demonstrated that difficulties in the language domain are reflected in altered neural processing of language and other auditory stimuli in toddlers with ASD (e.g., Eyler et al., 2012; Lombardo et al., 2015; Redcay & Courchesne, 2008; Redcay et al., 2008; Swanson, Wolff, et al., 2017), but the timeline underlying the emergence of this atypical developmental trajectory is still unclear. In this study, we aim to examine the neural networks underlying implicit language learning in 9-month-old infants well before the onset of overt behavioral symptoms and language delay.

Implicit learning is key to language acquisition; a crucial step in early language acquisition is identifying word boundaries in continuous speech (i.e., word segmentation). Prior research has robustly demonstrated that statistical regularities and prosodic cues (i.e., stress) are crucial cues that aid in word segmentation. Milestone behavioral studies indicate that infants as young as 6–8 months of age are able to segment a stream of speech based solely on statistical regularities in the input (Saffran et al., 1996), demonstrating that infants can implicitly identify word-like units in continuous speech based only on transitional probabilities. Behavioral studies have also shown that by 9 months of age, infants learning English start to rely also on prosodic cues to identify potential word boundaries (Elizabeth K. Johnson & Jusczyk, 2001; Jusczyk & Houston, 1999; Romberg & Saffran, 2010), since 90% of words in conversational English have stress on their initial syllable (Cutler & Carter, 1987). Importantly, a recent study using functional near-infrared spectroscopy (fNIRS) showed that newborns use both statistical cues and the prosodic contours of their native language to segment speech (Flo et al., 2019), indicating that typically developing infants are also sensitive to prosodic information of their native language early on. However, infants at familial risk for ASD (Ference & Curtin, 2013) and toddlers diagnosed with ASD (Paul et al., 2008) do not display the characteristic preference for the prosodic contours of their native language compared to typically developing infants and toddlers, who start to weigh speech cues more heavily than statistical cues by 8–9 months of age (Elizabeth K. Johnson & Jusczyk, 2001). This is consistent with prior work in youth and adults showing that prosodic production is often the most affected aspect of speech production in ASD (Depape et al., 2012; Peppé et al., 2011) and that individuals with ASD commonly display deficits in extracting prosodic information from speech (Lindstrom et al., 2018). Taken together, these studies suggest that sensitivity to prosody may be affected in infants at risk for ASD before impairments in language can be overtly detected.

Infant siblings of children with ASD are at higher risk for developing the disorder with a recurrence rate of around 20% (Ozonoff et al., 2011), and these infants often exhibit difficulties with language even if they do not go on to develop autism. Indeed, infants at familial risk for developing ASD have higher rates of language delay in both the receptive and expressive domains (Garrido et al., 2017; Hudry et al., 2014; Paul et al., 2011). Although overt behavioral symptoms only begin to emerge towards the end of the first year of life and infants are typically not reliably diagnosed until age 3 (Jones et al., 2014), prior findings from infant sibling studies indicate that early differences in brain function that precede behavioral symptoms associated with ASD are already present before the first birthday (as described below; see Wolff et al., 2017 for review).

Passive functional magnetic resonance imaging (fMRI) has been an effective tool for examining the neural circuitry underlying language processing in the first year of life. This method has been used to successfully map neural responses to speech in typically developing infants (Dehaene-Lambertz et al., 2002, 2006, 2010), with reliable activation in fronto-temporal language areas observed in 2-day-old neonates (Perani et al., 2011). A growing body of work has utilized fMRI to examine early differences in the functional neural substrates for auditory sensitivity in infant siblings at risk for developing ASD. High-risk infants have been found to be less sensitive to vocal sounds than their low-risk counterparts; when presented with naturalistic auditory stimuli (e.g., coughing, yawning, sneezing, laughing) at 4–6 months of age, low-risk infants showed greater activation in the right posterior superior temporal sulcus, a region important for processing prosodic cues and the emotional content of speech, compared to the high-risk group (Lloyd-Fox et al., 2013). This finding was further supported by another study, which also demonstrated that infants at low risk for developing ASD showed voice specialization in the temporal cortex by 4–7 months of age, whereas high-risk infants showed atypical cortical processing of socially relevant auditory information with less voice sensitivity in the same regions (Blasi et al., 2015). Furthermore, these early brain measures related to later behavioral measures of language and social development (Blasi et al., 2015), suggesting that neural networks subserving language acquisition already show differences based on risk for ASD that can predict developmental trajectories across time. Studies in older infants and toddlers with ASD have demonstrated a similar pattern of reduced activation in bilateral superior temporal gyrus (Lombardo et al., 2015) in addition to other temporal and frontal regions (Redcay & Courchesne, 2008) that play a role in language and speech processing. Despite efforts to characterize neural responses to speech in infancy (Blasi et al., 2015; Perani et al., 2011) and language processing in toddlers with ASD (Dinstein et al., 2011; Eyler et al., 2012; Lombardo et al., 2015; Redcay & Courchesne, 2008; Redcay et al., 2008), little research has focused on the neural correlates of prosodic processing and early language learning in young infants at risk for ASD.

By adapting the word segmentation paradigm used in infant behavioral studies (Saffran, Aslin, et al., 1996), we previously conducted a series of fMRI studies assessing developmental changes in the neural networks subserving implicit language learning in neurotypical children and adults (McNealy, Mazziotta, & Dapretto, 2006, 2010, 2011). In these studies, participants listened to three speech streams of concatenated syllables containing either strong statistical regularities alone, strong statistical regularities and prosodic cues, or virtually no cues. Since translational probabilities (i.e., statistical regularities) and the identification of word boundaries are computed online while listening and processing each speech stream, brain regions where activity increased as a function of exposure to the speech streams (i.e., learning-related signal increases) were examined. Neural activity differed significantly across the three conditions with strong signal increases in left temporal regions observed only while listening to speech streams containing strong statistical regularities, particularly when these were paired with prosodic cues (McNealy et al., 2006, 2010). Importantly, a post-scan behavioral word discrimination task was administered to investigate whether participants were able to discriminate between words presented in the speech streams. Both children and adults displayed a positive relationship between activity in the left temporal cortex while listening to the speech streams and accuracy at discriminating words in the post-scan task, indicating that the observed signal increases over time did indeed reflect implicit word segmentation and learning of word boundaries during the primary fMRI task (McNealy et al., 2006, 2010). In other words, an increase in signal was correlated with behavioral performance, providing evidence that the signal increases over time reflect the implicit computation of statistical regularities that occurs over time.

Using this same fMRI paradigm, we have also shown diminished neural responses associated with implicit language learning in youth with ASD, suggesting that children with ASD may be less sensitive to cues that guide word segmentation (Scott-Van Zeeland et al., 2010). In particular, youth with ASD (ages 9–16 years old) did not show learning-related signal increases in response to the speech streams with prosodic cues and/or statistical regularities, whereas the age matched typically developing youth demonstrated learning-related signal increases in left temporo-parietal cortex and basal ganglia. Taken together, these studies suggest that the processing of statistical regularities and prosodic cues, which are particularly salient for the detection of word boundaries, is altered in ASD. Although these studies were all conducted in awake children and adults, a recent fNIRS study which used a similar paradigm found evidence of implicit word segmentation in neurotypical neonates who were tested while sleeping or quietly resting (Flo et al., 2019), paving the road to investigate how these processes are affected during the first year of life in infants at high risk for ASD.

Here, we used the same fMRI word segmentation paradigm utilized in prior fMRI studies (McNealy et al., 2006, 2010, 2011; Scott-Van Zeeland et al., 2010) to examine the early signatures of implicit language learning and prosodic processing in 9-month-old infants at high (HR) and low familial risk (LR) for developing ASD. Although a recent fNIRS study using a similar paradigm has shown that neurotypical neonates can utilize both statistical and prosodic cues to segment a speech stream (Flo et al., 2019), the 9-month time point marks a particularly important stage in development because at this age infants begin to place increasing emphasis on speech cues (i.e., prosody; Elizabeth K. Johnson & Jusczyk, 2001). As such, we focused on investigating the learning-related signal increases during the speech stream containing both strong statistical regularities and prosodic cues. Based on prior evidence of altered neural activity associated with implicit language learning in youth with ASD (Scott-Van Zeeland et al., 2010) and emerging evidence of reduced temporal cortex activation in response to vocal stimuli in high-risk infants (Blasi et al., 2015; Lloyd-Fox et al., 2013) as well as toddlers with ASD (Lombardo et al., 2015; Redcay &Courchesne, 2008), we expected to observe differences in the sensitivity of language-related temporal regions to speech cues that are critical for language acquisition. Given that word segmentation skills in young infants have been shown to predict expressive vocabularies at 2 years of age (Newman et al., 2006) and altered activation to speech stimuli in toddlers with ASD has previously been associated with both language skills and ASD symptomatology (Redcay & Courchesne, 2008), we further expected that measures of brain activity at 9 months would predict subsequent behavioral measures of language development and ASD symptom severity. Although general cognitive development may likely play a role in mediating the association between developmental trajectories leading to ASD and overall language development, here we focused on examining how brain-based measures relate to ASD symptomatology and language outcome in line with other findings from the infant sibling literature (e.g., Emerson et al., 2017; Franchini et al., 2018; Lombardo et al., 2015; Swanson, Shen, et al., 2017). Lastly, given ample evidence of altered functional connectivity in many neural networks in ASD (Hernandez et al., 2015; Hull et al., 2016; Picci et al., 2016), including reports of disrupted synchronization of language areas in toddlers with ASD (Dinstein et al., 2011) as well as altered developmental trajectories of connectivity between language-related regions in at-risk infants during the first year of life (Liu et al., 2020), we additionally conducted an exploratory analysis examining how functional connectivity within language networks may vary as a function of implicit learning during exposure to the speech stream containing statistical regularities and prosodic cues.

2 |. MATERIALS AND METHODS

2.1 |. Participants

Infants were assigned to risk cohorts based on family history. High-risk (HR) infants had at least one older sibling with a confirmed ASD diagnosis; initial parent reports of sibling diagnoses were confirmed by a review of documented evidence. Low-risk (LR) infants had no family history of ASD (i.e., no first- or second-degree relatives with ASD) or any other neurodevelopmental disorders. Informed consent was obtained from parents/legal guardians of infant participants under protocols approved by the Institutional Review Board (IRB). Exclusionary criteria for both groups included: 1) genetic syndromes or neurological conditions (e.g., fragile X syndrome, tuberous sclerosis), 2) chronic medical conditions or significant perinatal insult impacting development, 3) severe visual, hearing, or motor impairment, 4) non-English speaking families, and 5) contraindication for MRI.

Infants included in the study were from households where English was the primary language. The primary caregiver completed a language exposure form, which included information on the number of languages spoken at home and the number of waking hours the infant interacted with each household member for each language reported. Inclusionary criteria for this paradigm required overall English exposure to be greater than 75% or English exposure during interaction with the mother to be at 100%. There was only one infant whose overall English exposure was less than 75% (at 69%) but had 100% English exposure during interaction with the mother; accordingly, this subject was included in the analyses.

A total of 64 infants (36 HR, 28 LR) underwent MRI during natural sleep at 9 months of age; 9 HR and 9 LR infants did not successfully complete this fMRI scan (i.e., woke up before or during this scan). Data from three LR infants did not pass quality control during data preprocessing (as described below) and were also excluded, yielding a final sample of 43 infants (27 HR, 16 LR). Risk-based cohorts were matched on age, gender, race, family income, maternal education, and amount of English exposure (Table 1). In contrast with the HR infants, whose inclusionary criteria required them to be younger siblings of children with ASD, some infants in the LR group were first born children (N = 7). In order to ensure that birth order did not influence the results, we extracted and compared parameter estimates of activity between first-born LR infants and LR infants who had older siblings in all regions where there was any significant effect in the LR group; in this admittedly small sample, no significant differences emerged based on birth order (all ps > 0.3).

Table 1.

Subject Demographics

	LR N=16	HR N=27	P
Sex (Female/Male)	9/7	8/19	.08
Race (White/Non-white/Missing)	10/6/0	18/7/2	.56
Family Income			.98
Not Answered	1	1
<25 K	0	1
25–50 K	1	4
50–75 K	2	2
75–100 K	3	4
100–125 K	3	5
>125 K	6	10
Maternal Education			.23
Not Answered	1	1
Some college	1	3
College degree	3	13
Some graduate school	1	1
Graduate school degree	10	9
	Mean (SD)	Mean (SD)	P
Age at Scan (Months)	9.13 (.40)	9.19 (.28)	.58
Overall English Exposure (%)	93.29 (9.25)	94.38 (8.20)	.69
English Exposure with Mother (%)	97.81 (4.82)	97.69 (5.87)	.95
Mean Relative Motion (mm)	.07 (.07)	.11 (.09)	.11
Maximum Relative Motion (mm)	1.81 (5.13)	2.38 (3.93)	.68
Mean Absolute Motion (mm)	.45 (1.25)	.69 (1.09)	.53
Maximum Absolute Motion (mm)	1.45 (3.29)	2.49 (3.85)	.37
Number of Volumes Censored	1.38 (2.63)	3.15 (5.84)	.26
VABS Communication Score (36 months)	107.33 (14.13)a	96.25 (14.23)b	.04
CDI Expressive Language Score (36 months)	537 (175.22)a	464.59 (221.11)c	.35
ADOS-2 CSS (36 months)	3.00 (2.86)d	3.14 (1.93)e

^aN=12

^bN=17

^cN=20

^dN=14

^eN=22

2.2 |. Experimental stimuli and design

We used the speech stream exposure fMRI paradigm as first described in McNealy et al. (McNealy et al., 2006), which was previously adapted from the word segmentation paradigm used in infant behavioral studies (Saffran, Aslin, et al., 1996). Infants were exposed to three counterbalanced streams of nonsense speech presented in one continuous block where statistical regularities (i.e., transitional probabilities) and prosodic cues were manipulated across three conditions (Figure 1). Artificial speech streams were created using three different sets of 12 syllables following the same procedure used in prior infant and adult behavioral studies (Elizabeth K. Johnson & Jusczyk, 2001; Saffran, Aslin, et al., 1996; Saffran et al., 1996), whereby each stream was created by repeatedly concatenating 12 syllables. Each syllable was recorded separately using SoundEdit (Macromedia; Adobe Systems, San Jose, CA), ensuring that the average syllable duration (0.267 s), amplitude (18.2 dB), and pitch (221 Hz) were not significantly different across the experimental conditions and matched those used previously in the behavioral literature. For two speech streams, the 12 syllables were used to make four tri-syllabic artificial words, which were randomly repeated three times to form a block of 12 words, subject to the constraint that no word was repeated twice in a row. Five such different blocks were created, and this five-block sequence was then concatenated three times to form a continuous speech stream lasting 144 seconds, during which each word occurred 45 times. For example, the four words “pabiku,” “tibudo,” “golatu,” and “daropi” were combined to form a continuous stream of nonsense speech with no breaks or pauses (e.g., pabikutibudogolatudaropitibudo...). Within the speech stream, transitional probabilities for syllables within a word and across word boundaries were 1 and 0.33, respectively. Thus, as the words were repeated, transitional probabilities could be computed and used to segment the speech stream.

FIGURE 1.

Speech Stream Exposure Paradigm. In this paradigm, three sets of 12 syllables were used to create three sets of four tri-syllabic words (a). The Stressed and Unstressed Languages were formed by concatenating these words to form two artificial languages, whereas the Random Syllables stream was created by pseudorandomly concatenating individual syllables (b). Infants were exposed to three counterbalanced speech streams (c) containing statistical regularities and prosodic cues (Stressed Language), solely statistical regularities (Unstressed Language), and minimal cues to guide word segmentation (Random Syllables).

In the Stressed Language condition (S), the speech stream contained strong statistical regularities as well as prosodic cues introduced by adding stress (i.e., slightly increased amplitude, longer duration, and higher pitch) to the initial syllable of each word, one-third of the time it occurred. Stress was added by slightly increasing the duration (0.273 s), amplitude (16.9 dB), and pitch (234 Hz) of these stressed syllables; these small increases were offset by minor reductions in these parameters for the remaining syllables within the Stressed Language condition to ensure that the mean duration, amplitude, and pitch would not be reliably different across the three experimental conditions. The initial syllable was stressed since 90% of words in conversational English have stress on their initial syllable (Cutler & Carter, 1987). In the Unstressed Language condition (U), the stream contained only strong statistical regularities as cues to word boundaries. A third speech stream—the Random Syllables condition (R)—was also created to control for activity associated with merely listening to a series of concatenated syllables. In this speech stream, the 12 syllables were arranged pseudorandomly such that no three-syllable string was repeated more than twice in the stream; the frequency of any two-syllable strings was also minimized. The statistical likelihood of any one syllable following another was very low (with an average transitional probability between syllables in the stream of 0.1; range 0.02–0.22), thus providing minimal cues to word boundaries (i.e., statistical computations of transitional probabilities would be considerably more challenging than for the other two speech streams). The three speech streams (144 s each) were interleaved with periods of silence (30 s each). The order of presentation of the three experimental conditions was counterbalanced across participants. Importantly, each subject heard a different set of syllables associated with each of the speech streams, ensuring that any difference between conditions could not be attributed to different degrees of familiarity with a given set of syllables.

2.3 |. Behavioral measures

Developmental assessments examining language development and ASD symptomatology were conducted at 36 months of age (Table 1). The Vineland Adaptive Behavior Scales, Second Edition (VABS II; Sparrow et al., 2005), a parental clinical interview measure of abilities in everyday settings, was administered to assess the child’s personal and social skills in four main domains including communication, daily living skills, socialization, and motor skills. Here, we focused on the subscale assessing communication skills, which provides an overall index of receptive, expressive, and written language skills (e.g., names at least 10 objects, uses phrases with a noun and a verb, states own first name or nickname when asked). In line with prior research demonstrating that brain activation in toddlers with ASD was stratified by language outcome as measured by the Vineland communication subscale (Lombardo et al., 2015), domain standard scores from this subscale were used. A total of 12 LR and 17 HR infants contributed VABS II data at this time point.

The MacArthur-Bates Communicative Development Inventory (CDI; Fenson et al., 2007) Words and Sentences checklist was also completed at 36 months. This instrument is a parent-report standardized questionnaire that measures expressive language skills, sentence complexity, and grammatical development. Here, we focused on the expressive language subscale measuring vocabulary production since only expressive language can be measured by the CDI at this time point (i.e., receptive language is only assessed until 18 months in the CDI). Although the CDI is more commonly used in infants ∼30 months and under, we opted to use the measure as completed at 36 months of age to have a consistent time point for all of the behavioral measures examined in this study. A total of 12 LR and 20 HR infants contributed CDI data at this time point.

The Autism Diagnostic Observation Schedule-2 (ADOS-2; Lord et al., 2012) was administered at 36 months. This is a play-based clinical assessment and gold-standard method for measuring ASD symptom severity. Participants were administered the appropriate ADOS-2 module according to their developmental level. The calibrated severity score (CSS) was used to compare symptom severity across ADOS modules, whereby higher CSS ratings represent greater symptom severity (Gotham et al., 2009). A total of 36 infants (14 LR, 22 HR) contributed ADOS-2 data at this time point. Infants were evaluated for ASD diagnosis based on ADOS and best clinical estimate at 36 months. Six HR infants met criteria for ASD and two LR infants had ADOS CSS scores in the clinical range (i.e., greater than or equal to 4). Importantly, our brain–behavior correlation in this group was still significant when these two infants were excluded from the analysis (ρ₁₀ = 0.717, p = 0.02). Since our group assignment was based on familial risk, and the modest sample size did not allow us to perform separate analyses comparing infants who did go on to receive diagnoses versus those who did not, we opted to keep these two LR infants within their original group.

2.4 |. fMRI data acquisition

At 9 months of age, infants underwent MRI during natural sleep. Parents were instructed to put their infant to sleep using their regular bedtime routine; once asleep, swaddled infants were transferred to the scanner bed. Soft and malleable silicone earplugs as well as MiniMuffs Neonatal Noise Attenuators (Natus Medical Inc., San Carlos, CA) were used as hearing protection; headphones used to convey auditory stimuli further dampened scanner noise. Infants were placed on a custom-made bed, which fit inside the head coil, and secured on the scanner bed with a Velcro strap. To minimize movement, a weighted blanket was used and foam pads were positioned around each infant’s head. Infants were presented with the auditory stimuli through a set of MR-compatible stereo headphones (Resonance Technology, Northridge, California; Optoacoustics LTD., Or Yehuda, Israel). Stimuli were presented using E-Prime 2.0 Software (Psychology Software Tools, Sharpsburg, Pennsylvania) on a Dell Latitude E6430 laptop computer. A study staff member remained in the scan room for the duration of the scan to monitor infants for overt movement, waking, or signs of distress.

All MRI data were collected on a Siemens 3T Tim Trio (12-channel head coil; 21 HR, 11 LR) or Prisma scanner (32-channel head coil; 6 HR, 5 LR) during natural sleep (scanner was included as a covariate of non-interest in all subsequent analyses). One functional scan lasting 8 minutes and 48 seconds was acquired with a T2*-weighted spinecho functional sequence (Siemens Trio: TR = 3000ms, TE = 28ms, matrix size 64×64, FOV = 192mm, 34 slices, 3mm in-plane resolution, with 4mm-thick axial slices; Siemens Prisma: identical parameters except with 33 slices). For registration purposes, a matched bandwidth T2-weighted high-resolution echo planar scan was acquired co-planar to the functional scan to ensure identical distortion characteristics (Siemens Trio: TR = 5000ms, TE = 34ms, matrix size 128×128, FOV = 192mm, 34 slices, 1.5mm in-plane resolution, with 4mm-thick axial slices; Siemens Prisma: identical parameters except with TE = 45ms, 33 slices).

2.5 |. fMRI data analysis

Functional imaging data were preprocessed and analyzed using FMRIB’s Software Library (FSL; Smith et al., 2004). Preprocessing included skull stripping, rigid-body motion correction (using MCFLIRT; Jenkinson et al., 2002), spatial smoothing (Gaussian kernel of 5mm full width at half maximum), high-pass filtering (180s), and affine registration with 12 degrees of freedom to the subject’s corresponding high-resolution matched bandwidth coplanar image using FMRIB’s Linear Image Registration Tool (FLIRT; Jenkinson et al., 2002; Jenkinson & Smith, 2001), followed by affine registration with 12 degrees of freedom to a standard 1-year infant brain template (Shi et al., 2011) using FLIRT.

FSL’s fMRI Expert Analysis Tool (FEAT) was used for statistical analyses. We performed two separate first-level analyses: first, mean BOLD activity maps of the three conditions were modeled and, second, significant increases were modeled using a linear function to examine learning-related signal increases for each condition during exposure to the speech streams (further described below). Each condition was coded as a separate explanatory variable and convolved with a canonical (double-gamma) hemodynamic response function. Volume censoring was performed as an added motion correction step in addition to the standard rigid-body motion correction procedures; volumes identified as outliers due to excessive motion using root mean squared (RMS) intensity difference (as compared to the center volume, and confirmed by visual inspection) were censored. The two groups did not differ on the total number of volumes censored (p = 0.26). There were also no between-group differences in mean or maximum relative motion, or in mean or maximum absolute motion (all ps > 0.26).

In the first analysis modeling mean BOLD activity maps of the three conditions (i.e., mean activation model), the simple effects of each experimental condition (Stressed Language, Unstressed Language, Random Syllables) were modeled with respect to baseline (i.e., rest). In the second analysis modeling signal increases (i.e., learning-related signal increase model), we investigated changes in neural activity as a function of exposure to the speech stream within each activation block where each activation condition (Stressed Language, Unstressed Language, Random Syllables) was modeled with a linear function with respect to baseline. In other words, each condition was examined using a linear signal increase model to investigate changes in neural activity as a function of exposure to the speech stream within the activation block (i.e., learning-related signal increases). By doing so, we aimed to identify the neural substrate of implicit learning (i.e., word segmentation) as it occurred over each activation block. Indeed, we previously demonstrated that both children and adults displayed a positive relationship between brain activation while listening to the speech streams and accuracy at discriminating words in a post-scan behavioral word discrimination task, indicating that the observed signal increases over time do reflect implicit word segmentation and learning of word boundaries during the primary fMRI task (McNealy et al., 2006, 2010). For all analyses, fixed-effects models were run separately for each participant, and then combined in a higher-level mixed-effects model to investigate within- and between-group differences.

Based on prior findings showing the strongest learning-related signal increases during the speech stream containing both statistical regularities and prosodic cues (Stressed Language condition) in neurotypical children and adults (McNealy et al., 2010, 2011), as well as a lack thereof in youth with ASD (Scott-Van Zeeland et al., 2010), we chose to focus our analyses primarily on this Stressed Language condition. In the mean activation model, we ran a between-group contrast to explore how LR and HR infants differed in their activation during the Stressed Language condition versus baseline (i.e., S). In the learning-related signal increase model, we examined signal increases that occurred during the Stressed Language condition (i.e., S↑) within each group. To specifically examine signal increases in response to prosodic cues, we contrasted the Stressed Language condition to the Unstressed Language condition (i.e., S↑ > U↑). Based on prior work demonstrating that toddlers with ASD are less sensitive to vocalizations that involve prosodic processing, planned between-group contrasts focused on investigating how the groups differed in learning-related signal increases in response to prosodic cues (i.e., S↑ > U↑).

Higher-level group analyses were carried out using FMRIB’s Local Analysis of Mixed Effects State with mixed effects modeling (FLAME 1+2) since equal variances were not assumed between the HR and LR groups. Scanner was included as a covariate of non-interest in these group-level analyses. Initial within-group analyses were prethreshold masked with an anatomical mask that included brain regions that showed significant activation during this paradigm in older children and adults (McNealy et al., 2006, 2010, 2011; Scott-Van Zeeland et al., 2010) and/or are regions known to be involved in processing language (Figure S1 and Table S1; regions anatomically defined from a standard infant brain atlas (Shi et al., 2011)). Between-group analyses for each condition (Stressed Language, Unstressed Language, Random Syllables) were prethreshold masked using a combined mask across both risk groups for each condition (HR+LR, liberally thresholded at z > 3.1, uncorrected; Figure S2). All imaging results are presented at z > 3.1, p < 0.001, cluster corrected for multiple comparisons at p < 0.05.

2.6 |. Brain-behavior data analysis

Behavioral correlations were conducted in R (R Core Team, 2018) to examine if brain-based differences in implicit language learning would predict later language development and ASD symptomatology at 36 months of age. Behavioral data were included for all infants in the sample who had aged to 36 months by the time of data analysis (Table 1). To specifically hone in on how brain activation to prosodic cues could be related to later impairment and subsequent outcome, we focused these correlations on brain regions in which activity increased as a function of exposure to the speech streams that were identical except for the addition of prosodic cues. That is, we examined the strength of association between parameter estimates extracted from regions which showed significant differences in activation between the Stressed and Unstressed Language conditions under the signal increase model (i.e., S↑ > U↑) and behavioral measures collected at 36 months of age, including the VABS communication subscale, CDI expressive language score, and ADOS-2 CSS. Behavioral correlations were conducted within each group separately; Spearman’s correlations were conducted since the behavioral data were not normally distributed.

2.7 |. Psychophysiological Interaction (PPI) analysis

An exploratory, post hoc psychophysiological interaction (PPI) analysis (Friston et al., 1997; O’Reilly et al., 2012) was conducted to examine functional connectivity between a specific seed region of interest (ROI) and the rest of the brain as a function of implicit learning. Since the most differences were found during exposure to the speech stream containing both statistical regularities and prosodic cues (Stressed Language), we focused the PPI analysis on this condition. We chose to examine functional connectivity of Heschl’s gyrus (HG) during stimulus presentation in light of prior work demonstrating altered resting-state functional connectivity between primary auditory cortex and other language and sensory processing regions in ASD (Abrams et al., 2013; Linke et al., 2017). Anatomical HG seeds derived from a 1-year infant brain atlas (Shi et al., 2011) were defined as spheres (4-mm radius) around the center of gravity within each region. These were binarized and summed together to create a bilateral HG seed and PPI analysis was conducted for this ROI using FEAT. Three regressors were implemented in the first-level analyses: the first was the psychological variable (i.e., condition type) convolved with a double-gamma hemodynamic response function, the second was the physiological variable (i.e., the time course of the seed ROI), and the third was the interaction between the physiological and psychological variables (PPI). Given that the Stressed Language condition was our primary contrast of interest, we focused our PPI analysis on this contrast to examine increasing functional connectivity as the infants were presented with the speech stream containing both statistical and prosodic cues. That is, we modeled learning-related signal increases during the Stressed Language condition to calculate significant increases in functional connectivity during this condition. Between-group contrasts were prethreshold masked using a combined mask across both groups (HR and LR, liberally thresholded at z > 3.1, uncorrected) for the Stressed Language condition and corrected for multiple comparisons at a threshold of z > 3.1, p < 0.001, cluster corrected for multiple comparisons at p < 0.05.

3 |. RESULTS

Passive listening to all three conditions (compared to silent baseline) activated bilateral temporal cortices in both risk groups (Figure S3). Between-group comparisons revealed that LR infants showed significantly more activation in the left amygdala during the Stressed Language condition compared to HR infants (Figure 2b and Table 2). There were no regions where HR infants showed greater activation than LR infants during this condition.

FIGURE 2.

Between-group differences in mean activation during the Stressed Language condition. Compared with the HR group, LR infants showed significantly more activation in the left amygdala during the S condition.

Table 2.

Within- and Between-group Activation and Signal Increases

Significant Contrast	Region	L/R	Max Z	Peak (mm)			Size (voxels)
				x	y	z
Mean Activation Model (vs. baseline)
LR > HR: S	Amygdala	L	3.59	−22	1	−12	73
Learning-Related Signal Increase Model
LR: S↑	STG	L	4.21	−42	−25	12	102
LR: S↑ > U↑	STG	L	4.79	−36	−27	13	202
	STG	R	5.97	59	−23	3	141
	HG	L	4.66	−36	−23	12	78
LR > HR: S↑ > U↑	STG	L	4.27	−35	−27	11	92
	HG	L	4.26	−36	−23	12	35
Exploratory PPI Analysis
LR > HR: PPI S↑	Insula	R	3.99	31	13	−1	183

LR: low risk; HR: high risk; STG: superior temporal gyrus; HG: Heschl’s gyrus

Since transitional probabilities and the identification of word boundaries are computed online, regions where activity increased as a function of exposure to the speech streams were examined to detect learning-related signal increases over time. As noted previously, we focused on the Stressed Language condition, which showed the strongest signal increases in neurotypical children and adults (McNealy et al., 2009, 2010). LR infants displayed significant signal increases in left STG for this speech stream (S↑; Figure 3a and Table 2). In contrast, no significant signal increases were observed in the HR group during this condition. Comparing between conditions to examine signal increases specific to the processing of prosodic cues during the Stressed compared to the Unstressed Language condition, the LR group showed greater signal increases in bilateral STG and left HG (S↑ > U↑; Figure 3b and Table 2). Between-group comparisons revealed that LR infants showed greater signal increases than HR infants in left STG and HG for this contrast as well (S↑ > U↑; Figure 4a and Table 2). There were no regions where HR infants showed greater signal increases compared to the LR group for this contrast.

FIGURE 3.

Learning-related signal increases in the LR group were associated with later language development. LR infants showed more signal increases in the left STG during the Stressed Language condition (a) and in bilateral STG as well as left HG during the Stressed compared to Unstressed Language conditions (b). Greater signal increases in the left STG/HG were associated with better VABS communication scores at 36 months (ρ₁₀ = 0.771, p = 0.003; c).

FIGURE 4.

Correlations between signal increases and later language development as well as ASD symptomatology. During the Stressed compared to the Unstressed Language condition, LR infants showed significantly more learning-related signal increases in left STG/HG compared to HR infants (a). In the HR group, greater signal increases in left temporal cortex were associated with better CDI expressive language scores (ρ₁₅ = 0.605, p = 0.01); b) and less severe ADOS-2 CSS (ρ₂₀ = −0.416, p = 0.027; c) at 36 months. HR infants with an ASD diagnosis at 36 months are denoted with the filled circles.

Next, we examined how early differences in language-related neural activity at 9 months of age might relate to later language development and ASD symptomatology. In the LR group, we focused on the regions that showed differences when contrasting the two language conditions to investigate the effect of signal increases to prosodic cues on later language development (LR: S↑ > U↑; Figure 3b). Greater signal increases in left STG/HG at 9 months of age were associated with better VABS communication scores at 36 months (ρ₁₀ = 0.771, p = 0.003; Figure 3c). There was no significant relationship with the CDI expressive language score. Between-group differences for this same contrast were also found in this left temporal region (LR > HR: S↑ > U↑; Figure 4a). In the HR group, infants who showed greater signal increases in this left STG/HG cluster at 9 months of age had better CDI expressive language scores (ρ₁₅ = 0.605, p = 0.01; Figure 4b) and lower levels of social communicative impairments as indexed by the ADOS-2 CSS at 36 months (ρ₂₀ = −0.416, p = 0.027; Figure 4c). There was no significant relationship with the VABS communication scores.

Lastly, in our exploratory PPI analysis, we found increasing functional connectivity between bilateral HG and right anterior insula in the LR group compared with the HR group over the course of the Stressed Language condition (Figure 5 and Table 2).

FIGURE 5.

Between-group differences in PPI analysis of functional connectivity during the Stressed Language condition. Compared to the HR group, LR infants show increasing connectivity between bilateral HG and the right anterior insula.

4 |. DISCUSSION

To our knowledge, this is the first study to examine early neural sensitivity to speech prosody and the modulation of language-network connectivity as a function of implicit learning in infants at risk for ASD. Our main goal for this study was to characterize the neural mechanisms underlying implicit language learning and prosodic processing in infants at risk for ASD and further examine how early differences in brain activity may relate to later behavioral outcome. Compared to HR infants, LR infants showed significantly more signal increases in left temporal regions during exposure to the speech stream containing both strong statistical regularities and prosodic cues. Interestingly, signal increases at 9 months were related to better language outcome at 36 months of age in both groups. In addition, for HR infants, greater signal increases at 9 months were associated with less severe ASD symptomatology at 36 months. During exposure to the speech stream containing not only strong statistical regularities but also prosodic cues to word boundaries, LR infants showed greater increasing functional connectivity between bilateral primary auditory cortex and right anterior insula compared with HR infants. Overall, our findings suggest that early differences in the neural networks related to language learning may predict later trajectories in language development and ASD symptomatology well before delays in language acquisition and ASD symptoms can be overtly observed at the behavioral level.

A growing body of work has begun to identify brain regions that are important for language processing in infants and toddlers, including the STG (Blasi et al., 2011, 2015; Dehaene-Lambertz et al., 2002; Eyler et al., 2012; Perani et al., 2011). Consistent with prior work showing brain responses to language during natural sleep, we found that all infants showed activation in bilateral temporal cortices in response to all three speech streams. Compared to HR infants, LR infants showed greater activation in the left amygdala during the Stressed Language condition, a region implicated in salience detection that has been associated with social impairments in ASD (see Zalla & Sperduti, 2013 for review). While the prosodic cues in this study were generated by primarily manipulating lexical stress, the stress cues featured a higher pitch (in addition to slightly increased amplitude and longer duration, as described previously), which is one of the cues that conveys affective prosody (Liebenthal et al., 2016). There is increasing evidence that, in addition to orienting attention to socially relevant stimuli in the environment (Sander et al., 2003; Zalla & Sperduti, 2013), the amygdala plays a role in the processing of affective prosody (Adolphs, 2002) and early language acquisition (Dehaene-Lambertz et al., 2010). Babies begin to learn language in the context of parent-child interactions, which are commonly characterized by infant-directed speech loaded with salient cues such as exaggerated prosody (i.e., stress; P. Kuhl & Rivera-Gaxiola, 2008). Indeed, when listening to their mother’s voice, 2-month-old infants show activation in regions involved in emotional processing, including the amygdala and the left posterior STG (Dehaene-Lambertz et al., 2010), suggesting that the amygdala may play a role in gating early language acquisition. This is supported by volumetric studies in neurotypical infants reporting associations between amygdala volume at 6 months and language outcomes from 6 months to 4 years of age (Ortiz-Mantilla et al., 2010). Thus, our findings reporting differences in amygdala activation in infancy may reflect atypical processing in the HR group whereby these infants may not be recruiting the normative regions for processing available prosodic cues in the input. This is consistent with prior work in HR infants with early language delay showing altered brain–behavior associations between amygdala volume and later language skills (Swanson, Wolff, et al., 2017).

To investigate neural correlates of language-related implicit learning, we further examined brain regions that showed increases in neural activity as a function of exposure to the speech streams. Prior fNIRS research on neonates using a similar paradigm with a learning phase followed by a testing phase revealed differential neural responses during the testing phase, demonstrating that these neonates successfully segmented the speech stream and learned the words presented during the exposure phase (Flo et al., 2019). That is, the infants were not only sensitive to statistical and prosodic cues, but they also capitalized on these cues to segment the speech stream and distinguish words from non-words during the testing phase. In the present study, post-scan testing was not conducted to assess the degree to which the infants successfully segmented the speech stream, thus preventing us from unequivocally linking the observed signal increases to implicit learning. However, our fMRI findings are in line with prior fNIRS evidence demonstrating that the neural signatures of implicit learning can be detected using this paradigm in very young infants. LR infants showed significant signal increases in a left temporal cortex during the Stressed Language condition; this is consistent with findings in neurotypical children showing similar signal increases in temporal cortices while listening to the condition with prosodic cues (McNealy et al., 2010, 2011). In contrast, we found evidence of disrupted language network activity in the HR group; HR infants did not show any significant signal increases over time for any speech stream, consistent with a prior study where older youth with ASD did not show any significant signal increases for any speech stream (Scott-Van Zeeland et al., 2010). This suggests that early on in development, HR infants are already less sensitive to implicit statistical and speech cues that guide word segmentation during language learning.

Prior research has demonstrated that infants start to weigh speech cues more heavily than statistical cues by 8–9 months of age (Johnson & Jusczyk, 2001). Thus, to isolate the signal increases bolstered by the presence of prosodic cues, we directly contrasted the two speech streams that differed only with respect to the presence of prosodic cues. The LR group showed significantly stronger signal increases in bilateral auditory cortex regions (i.e., left HG and bilateral STG) during the Stressed compared to the Unstressed Language condition, indicating that the presence of prosodic cues was associated with increased activity in these critical language processing regions. This finding is consistent with prior work demonstrating that infants show bilateral activation to language stimuli that becomes left-lateralized over the course of development (Emerson et al., 2016). Interestingly, these early differences in the neural responses to speech cues that guide word segmentation at 9 months of age were related to later language development. In the LR group, greater signal increases in the left STG/HG region for this contrast were associated with better communication skills as assessed by the VABS at 36 months of age, supporting the notion that signal increases in this region are important for prosodic processing and language acquisition in early infancy. HR infants did not show any signal increases for this contrast, indicating that these regions were not engaged in the task in the HR group as they were in the LR group; this could account for why the VABS communication scores were related to the learning-related signal changes observed in the LR group, but not in the HR group.

Compared to the HR group, LR infants showed more signal increases in left temporal cortex (left STG and HG) during exposure to the speech stream with both prosodic and statistical cues versus the speech stream with statistical regularities only. This finding is in line with prior neuroimaging studies in toddlers with ASD reporting reduced left hemisphere activity and abnormal right-lateralized responses to speech/language stimuli (Eyler et al., 2012; Redcay & Courchesne, 2008; Redcay et al., 2008). Although we did not observe any evidence of increased right hemisphere activity in our HR group and we did not explicitly examine differences in laterality, our findings suggest that LR infants recruit language-relevant regions in the left hemisphere to process prosodic cues, whereas HR infants may not be capitalizing on these important cues to word boundaries in the same way. Although neurotypical newborns can already segment speech based on statistical regularities and prosodic contours (Flo et al., 2019), prior behavioral work has demonstrated that by 8–9 months of age, typically developing infants begin to weigh prosodic cues more heavily than statistical regularities to identify word boundaries (Johnson & Jusczyk, 2001; Johnson & Seidl, 2009). Indeed, our results show that in 9-month-old LR infants, over and above the availability of statistical regularities in the input, the presence of prosodic cues elicited signal increases in primary auditory and language processing regions. Since the ability to correctly segment words from continuous speech has been shown to be predictive of future language development (Newman et al., 2006), we also tested whether this neural response to prosodic cues, which we take to index implicit learning, may be associated with later development in the HR infants.

Importantly, signal increases in the left STG/HG region where LR infants showed greater signal increases compared to the HR group were significantly related to later language development and ASD symptomatology. HR infants showing more signal increases in this area at 9 months of age had higher (i.e., better) expressive language scores at 36 months of age. Furthermore, within the HR group, greater signal increases in left STG/HG were also related to less severe ASD symptoms at 36 months of age. Taken together, these findings suggest that greater signal increases in this left temporal region are associated with better language outcome and less severe ASD symptomatology at 36 months of age for HR infants. Although we detected correlations between left temporal regions and later outcome, prior work in toddlers with ASD has related greater activation in right temporal regions with better receptive language skill and less severe autism severity scores (Redcay & Courchesne, 2008). Several key differences across the two studies may account for the seemingly divergent findings. We studied a sample of infants who were at elevated familial risk for ASD, whereas Redcay and Courchesne’s (2008) sample included toddlers with an ASD diagnosis. Furthermore, Redcay and Courchesne (2008) related overall activity in response to language to behavior; in contrast, here we examined the relationship between learning-related signal increases and later behavioral outcome. Our findings suggest that HR infants who showed more normative patterns of signal increases had better language outcome and less severe ASD symptomatology at 36 months. By contrast, the findings in Redcay and Courchesne (2008) may reflect compensatory mechanisms in this clinical group of toddlers. Thus, the brain measures detected here may reflect an earlier profile during development that can be meaningfully related with later behavioral outcome to shed insight on early mechanisms underlying altered developmental trajectories associated with risk for ASD. Furthermore, altered neural activation has also been shown to be stratified by language outcome; compared with ASD toddlers who do not exhibit language delays, ASD toddlers with speech delays show attenuated neural responses in the networks underlying language and speech processing (Lombardo et al., 2015).

In addition to atypical patterns of brain activity, altered functional connectivity is characteristic of developmental disorders involving language impairment including ASD (Hernandez et al., 2015; Hull et al., 2016; Picci et al., 2016). Indeed, there is ample evidence of disrupted connectivity in toddlers with ASD (Emerson et al., 2017; Keehn et al., 2013; McKinnon et al., 2019; Pruett et al., 2015). Toddlers with ASD display disrupted synchronization of language areas (Dinstein et al., 2011) and at-risk infants already show altered development of functional connectivity between language-related regions by 1.5 months of age (Liu et al., 2020), suggesting that disrupted functional connectivity early in life may be predictive for language impairments associated with the disorder. Recent work in adults with ASD using psychophysiological interaction (PPI) analyses have examined differences in the functional coupling (i.e., connectivity) of brain regions during task-based fMRI. Hoffman et al. (2016) examined task-related connectivity in adults with and without ASD using a task where participants listened to human vocal sounds, environmental sounds, and animal sounds. Individuals with ASD showed reduced connectivity between left temporal regions and frontal cortex compared to controls when listening to the human voice sounds compared with the other sounds. In a separate study, adults with ASD showed reduced connectivity between the superior temporal sulcus and amygdala during emotional prosody processing compared to the control group (Rosenblau et al., 2017). These studies indicate that in addition to exhibiting differences in neural activity in response to auditory stimuli, individuals with ASD show atypical connectivity between different brain regions implicated in language processing. This body of work informed our final exploratory analysis, which revealed that compared to the HR group, LR infants showed increasing connectivity between bilateral primary auditory cortex (i.e., HG) and the right anterior insula during the speech stream containing both statistical regularities and prosodic cues. The right anterior insula is the hub of the salience network (Seeley et al., 2007), a key neural network involved in orienting attention to the most salient aspects of the environment (Uddin, 2015). Since language development is strongly driven and influenced by social interaction (Kuhl, 2010; Kuhl & Rivera-Gaxiola, 2008; Lytle & Kuhl, 2017), connectivity with the right anterior insula may be cuing the brain to attend to salient information in the auditory input that would guide word segmentation and language acquisition more broadly. Our results thus suggest that connectivity with the right anterior insula may play a role in supporting normative development of networks for language processing and implicit learning during the first year of life. Consistent with prior findings demonstrating disrupted connectivity of the salience network in youth with ASD (Uddin et al., 2013), altered connectivity with the hub of this network observed in infants at risk for ASD could result in reduced attention to developmentally relevant social stimuli, which could have cascading effects on language development.

The lack of differential responsivity to prosodic cues in HR infants observed in our study is consistent with a growing body of work demonstrating that HR infants who later develop ASD are less sensitive to social cues available in their environment. Indeed, HR infants as young as 4–6 months of age already show less activation to social stimuli in both the visual and auditory domains compared to LR infants (Lloyd-Fox et al., 2013). Altered attention to salient cues that aid language acquisition may foreshadow later language development. This is particularly critical because longitudinal studies investigating language outcome from childhood through adulthood in ASD demonstrate that after the age of 6, language development prognosis is relatively stable through adulthood (Pickles et al., 2014), highlighting the importance of intervening early on in development.

This study presents some limitations that warrant discussion. Of note, among the HR infants for whom data were available at 36 months of age, only six HR infants qualified for an ASD diagnosis. Accordingly, given the small sample, we could not conduct meaningful comparisons between infants who did receive a diagnosis and those who did not. Given that infants at high familial risk for developing ASD are expected to exhibit a range of heterogeneous outcomes including other neurodevelopmental, learning, or language disorders (Miller et al., 2016; Zwaigenbaum et al., 2007), our findings may be better interpreted as an indicator of atypical trajectories in language development that may reflect a more general susceptibility to language impairment common to other developmental disorders. Future studies should employ larger sample sizes that can be subgrouped by outcome (e.g., HR-typically developing, HR-ASD, HR-broader ASD phenotype/developmental delay) in order to better elucidate the underlying mechanisms that contribute to the heterogeneity observed in this infant sibling population. In addition, while this study focused on alterations in neural activity and stimulus-related functional connectivity of brain regions associated with language processing, HR infants also exhibit altered structural connectivity of the white matter tracts connecting these regions (Liu et al., 2018; Wolff et al., 2012; Wolff, Swanson, et al., 2017). Indeed, altered structural connectivity has been identified as a possible biomarker for ASD (Ameis & Catani, 2015). Accordingly, future research should aim to investigate the impact of environmental/experiential factors on early functional as well as structural brain differences in at-risk populations.

Extending prior work showing evidence of word segmentation in neurotypical newborns (Flo et al., 2019) as well as disrupted activity subserving this process in older youth with ASD (Scott-Van Zeeland et al., 2010), this is the first study to examine implicit learning and prosodic processing in a cohort of infants at high familial risk for ASD. In line with prior evidence in older children with ASD, HR infants already show diminished neural responses associated with language processing by 9 months of age. Compared to the HR group, LR infants showed greater neural activity as well as stronger functional connectivity in regions that play a key role in salience detection. Our findings indicate that early differences in the neural networks underlying language learning may predict altered trajectories in language development and ASD diagnosis long before delays in language acquisition can be overtly observed at the behavioral level. Future larger-scale studies in infants at elevated risk for ASD should aim to parse the heterogeneity of this sample and investigate how distinct neural signatures associated with implicit learning may predict different outcomes and thus inform early and targeted interventions.

RESEARCH HIGHLIGHTS.

• A passive listening stimulus-evoked fMRI paradigm was used to examine the neural correlates underlying word segmentation in 9-month-old infants at high/low familial risk for ASD

• High-risk infants showed limited brain activity associated with prosodic processing and implicit language learning compared to low-risk infants

• Distinct patterns of neural responses at 9 months were associated with later measures of language development and ASD symptomatology at 36 months of age

• These findings suggest that speech cues important for language learning may be less salient and therefore command less attention in high-risk infants

• Early atypicalities in the neural networks subserving language processing/learning may predict language acquisition and altered developmental trajectories in at-risk infants