Aberrant Neural Entrainment to Word‐Level Speech Patterns in Fragile X Syndrome: Evidence for a Statistical Learning Deficit

ABSTRACT

Fragile X syndrome (FXS), the most common inherited cause of intellectual disability and autism spectrum disorder, causes significant language and cognitive impairments. Statistical learning refers to the ability to extract patterns from sensory input through mere exposure and plays a central role in language acquisition. Surprisingly, statistical learning in FXS has not been explored. Given that children with FXS typically follow a delayed developmental trajectory for language, we hypothesized that they would show impaired statistical learning. To test this hypothesis, we used an EEG measure of neural entrainment to index statistical learning of hidden trisyllabic words within a continuous speech stream in children with FXS (n = 17) and in typically developing controls (n = 31). Children with FXS showed significantly reduced neural entrainment to words compared to controls, particularly in the superior temporal gyrus and transverse temporal gyrus (primary auditory cortex), providing evidence of statistical learning impairment. Notably, syllable‐level entrainment was preserved or even enhanced in FXS, indicating that word‐level deficits cannot be attributed to general auditory processing impairments. In addition, while typically developing controls showed an increase in word‐level entrainment over the course of learning, children with FXS failed to show a similar increase over time. Taken together, this pattern of results demonstrates that children with FXS can process rapid, lower‐order acoustic structure but struggle to integrate these syllables into longer, chunk‐like word representations. Overall, these findings suggest that statistical learning is impaired in FXS and also suggest neural entrainment to statistical structure as a potential therapeutic target.

1. Introduction

Statistical learning is the ability to extract patterns from sensory input through mere exposure, without instruction, reinforcement, or feedback (Aslin 2017; Saffran, Aslin, and Newport 1996). In the seminal demonstration of statistical learning, 8‐month‐old infants learned the underlying statistical structure of a continuous stream of repeating trisyllabic nonsense words (e.g., “tupiro”) after only 2 min of passive listening (Saffran, Aslin, and Newport 1996). Subsequent work has confirmed that statistical learning is a powerful learning mechanism that is present in infants, children and adults (Moreau et al. 2022; Raviv and Arnon 2018; Saffran et al. 1997; Saffran, Newport, and Aslin 1996) and that it supports widespread cognitive abilities, including visual perception (Fiser and Aslin 2001), motor learning (Theeuwes et al. 2024), social learning (Nencheva et al. 2025), and especially language acquisition (Saffran and Kirkham 2018).

Because statistical learning is foundational in language acquisition and other aspects of cognition (e.g., Kidd 2012; Romberg and Saffran 2010; Sherman et al. 2020), understanding statistical learning in neurodevelopmental disorders (NDDs) may provide new mechanistic insights for the altered patterns of learning and language outcomes associated with these disabilities (Saffran and Kirkham 2018). However, to date there are mixed findings regarding statistical learning abilities in NDDs as a whole. For example, children with Developmental Language Disorder (formerly known as Specific Language Impairment) seem to consistently show poorer performance on statistical learning tasks compared to typically developing children (Evans et al. 2009; Haebig et al. 2017; Obeid et al. 2016). In contrast, for autism spectrum disorder (ASD), there is behavioral evidence both for (Hu et al. 2024; Jones et al. 2018) and against statistical learning impairments (Brown et al. 2010; Haebig et al. 2017; Mayo and Eigsti 2012; Obeid et al. 2016). However, as has been previously pointed out (Saffran and Kirkham 2018), behavioral studies have been limited to relatively high functioning children who are able to successfully perform the tasks (e.g., making forced‐choice judgments about which syllable sequences are more familiar). Capturing statistical learning with neural measures would in principle allow for inclusion of children with more severe deficits who cannot perform behavioral tasks, and may also provide a more direct index of statistical learning than behavioral measures, which may be influenced by compensatory strategies (Livingston and Happé 2017).

Indeed, several studies to date that have used neural measures of statistical learning have consistently indicated reduced learning abilities in children with ASD (Jeste et al. 2015; Scott‐Van Zeeland et al. 2010; Wagley et al. 2020), in contrast to the mixed behavioral findings. Further, one of these studies found that within the ASD group, children with the most severe communication deficits showed the greatest impairments in learning at the neural level, as reflected by reduced signal increases in brain regions associated with statistical learning (Scott‐Van Zeeland et al. 2010), suggesting more pronounced statistical learning deficits in lower functioning children. Nonetheless, like behavioral studies, past neuroimaging studies (Jeste et al. 2015; Scott‐Van Zeeland et al. 2010; Wagley et al. 2020) have also included relatively high functioning children, with IQs ranging from borderline to normal. Thus, a next step in better understanding language impairments in NDDs would be to use neural measures to characterize statistical learning capacities in more profoundly affected children.

One useful measure of statistical learning that has not been previously used in the investigation of NDDs is neural entrainment (Obleser and Kayser 2019), which arises through the synchronization of brain oscillations to the frequency of repeating patterns. By adapting classical statistical learning paradigms and presenting individual syllables forming repeated words at a fixed rate, neural activity becomes synchronized to the individual syllables, and also—critically—to the component words (Batterink and Paller 2017; Buiatti et al. 2009; see Figure 1). This word‐rate entrainment is thought to reflect learners' moment‐to‐moment sensitivity to the statistical structures in the stream, with multiple neural and cognitive mechanisms potentially contributing to the response. Proposed mechanisms include general auditory tracking of acoustic regularities in the speech stream (Pinto et al. 2022; Sjuls et al. 2024), implicit sensitivity to transitional probabilities between syllables (Pinto et al. 2022), sequence prediction dynamics (Xu et al. 2024), Hebbian learning (Endress 2024), and perceptual integration of individual elements into higher‐order units (e.g., Batterink and Paller 2017; Buiatti et al. 2009). Despite these differing accounts, neural entrainment to words appears to consistently capture learners' real‐time sensitivity to statistical regularities in input. This interpretation is supported by findings that word entrainment (1) increases gradually over the course of exposure to the structured speech stream, reflecting the time course of learning (Batterink and Paller 2017, 2019; Choi et al. 2020; Elmer et al. 2021; Kabdebon et al. 2015; Moreau et al. 2022; Ordin et al. 2020; van der Wulp et al. 2025; Zhang et al. 2021); and (2) positively predicts learners' subsequent performance on behavioral measures of learning (Batterink 2020; Batterink and Paller 2017, 2019; Buiatti et al. 2009; Choi et al. 2020; Pinto et al. 2022; van der Wulp et al. 2025; see also Moser et al. 2021 for results from a nonlinguistic paradigm). Notably, not all studies have observed a relationship between word‐level entrainment and learning performance, possibly due to methodological issues such as low reliability and high interindividual variability in statistical learning measures, as well as a general reliance on explicit over implicit behavioral measures (Ordin et al. 2020; Smalle et al. 2022; Zhang et al. 2021; see Sjuls et al. 2024 for review; see Moreau et al. 2022 for a null association of entrainment with an implicit learning measure in children).

FIGURE 1.

Neural entrainment statistical learning paradigm. Syllables are presented at 3.33 Hz, forming hidden trisyllabic words presented at 1.11 Hz. Neural entrainment at 1.11 Hz, as measured by intertrial phase coherence (ITC), serves as an online neural index of statistical learning. Entrainment at 3.33 Hz reflects sensory tracking of the auditory signal.

These results suggest that despite its limitations (see Pinto et al. 2022; Sjuls et al. 2024), neural entrainment can be leveraged as a sensitive, real‐time, and relatively direct measure of statistical learning. Importantly, neural entrainment to words appears to occur relatively automatically, emerging without focused attention to the speech stream (Batterink and Paller 2019) and persisting even in minimally conscious patients (Benjamin et al. 2025; Xu et al. 2023). Neural entrainment may thus provide an index of statistical learning that is relatively less influenced by other cognitive factors (e.g., attention, executive control, decision‐making capabilities) than standard behavioral measures. Finally, because neural entrainment requires only passive listening on the part of the participant, it offers an especially useful tool for investigating statistical learning in children with NDDs, including those who are most severely affected and unable to perform behavioral tasks.

Neural entrainment at a given frequency of interest can be quantified both through spectral amplitude and intertrial phase coherence (ITC). Spectral amplitude peaks capture the strength of the neural response relative to neighboring frequencies, while ITC reflects the consistency of the phase of neural oscillations across trials. While both metrics often show similar significant increases in response to rhythmic auditory stimulation (Benjamin et al. 2021), relative to power estimates, ITC is more robust to background low frequency fluctuations (Forget et al. 2010), does not require normalization against background aperiodic (1/f) neural activity (Kabdebon et al. 2015), and is a more sensitive measure of synchronized neural activity (Ding et al. 2018; Ding and Simon 2013). Thus, ITC has emerged as a common method for capturing neural entrainment during statistical learning (Batterink and Paller 2017; Herrera‐Chaves et al. 2025; Sherman et al. 2023; Sjuls et al. 2024; Xu et al. 2023), an approach that is followed in the current study.

Fragile X syndrome (FXS) is the most common inherited cause of intellectual disability and the leading single‐gene cause of ASD, affecting approximately 1 in 4000 males and 1 in 8000 females (Hunter et al. 2014). Clinically, FXS includes intellectual disability, social interaction difficulties, generalized and social anxiety, attention deficits, sensory hypersensitivity, and language delays that range from mild to severe (Hagerman et al. 2017; Kaufmann et al. 2004). Approximately 50%–60% of males and 20% of females with FXS also meet criteria for ASD (Baker et al. 2019; Kaufmann et al. 2017). Language impairments in FXS can include pronounced delays and deficits in receptive language, vocabulary, morphology and syntax, conversational discourse, and processing of rapidly changing auditory information (Finestack et al. 2009; Hoffmann 2022). Sensory hyper‐responsivity and cortical hyper‐excitability are also characteristic, and consistently reported across autonomic and EEG measures (Baranek et al. 2008; Elmaghraby et al. 2025; Ethridge et al. 2019; Miller et al. 1999; Wang et al. 2017). Notably, in contrast to other NDDs such as idiopathic autism, the underlying biology of FXS is very well‐characterized. The disorder results from hypermethylation of the fragile X message ribonucleuprotein 1 (FMR1) gene, leading to loss of fragile X messenger ribonucleoprotein (FMRP), an RNA‐binding protein that regulates activity‐dependent translation at synapses. This protein normally acts as a translational repressor for hundreds of mRNAs critical for synaptic plasticity, and its absence results in excessive protein synthesis, disrupted synaptic pruning, and altered excitatory‐inhibitory balance. Further, FMRP regulates the translation of many genes that are independently implicated in ASD (Darnell et al. 2011; Iossifov et al. 2012), and synaptic plasticity mechanisms disrupted in FXS, particularly mGluR5 signaling and GABAergic dysfunction, are also implicated in other NDDs (Bear et al. 2004; Berry‐Kravis 2022). These shared pathways make FXS an informative single‐gene model for understanding statistical learning deficits that may extend to the broader NDD population. Nonetheless, we note that findings from FXS cannot be directly generalized to all NDDs without further investigation, as FXS‐specific mechanisms (e.g., loss of FMRP) may produce deficits distinct from those seen in other conditions.

Despite extensive documentation of language impairments in FXS, no studies have directly examined statistical learning in this population. This gap is particularly striking, given that statistical learning underlies many linguistic abilities affected in FXS, including word segmentation and syntax acquisition (Romberg and Saffran 2010). In the current study, we address this gap by using neural entrainment as a behaviourally‐independent measure of statistical learning. We test the hypothesis that children with FXS will show reduced neural entrainment (measured via inter‐trial phase coherence) to repeating patterns in auditory streams compared to age‐matched typically developing controls, reflecting a deficit in the online computations underlying statistical learning. By documenting statistical learning deficits in FXS, this work provides the first direct evidence of impaired implicit learning mechanisms in this population and establishes a foundation for future mechanistic studies and targeted interventions.

2. Methods

2.1. Participants

Participants included a total of 17 children with FXS (full mutation confirmed by PCR/Southern blot) and 31 typically developing children (TDC; average IQ 85–115, no neuropsychiatric family history). Two additional typically developing controls were recruited but later excluded from the final dataset due to poor data quality (n = 1) and a technical problem with the EEG file that precluded data analysis (n = 1). Table 1 includes demographic and IQ scores for each group. In addition, Table S1 includes language scores on the CELF‐5, an assessment of receptive and expressive language abilities, for a subset of FXS and TDC participants who successfully completed the protocol. We pooled analyses across male and female individuals in order to maximize statistical power. Sample size reflects available recruitment during the study period. A sensitivity power analysis conducted using G*Power for a two‐sample comparison (α = 0.05, n = 17/31) indicates that we were powered to detect a standardized group difference of approximately d = 0.76 with 80% power (alpha = 0.05, one‐tailed). We used a one‐tailed test for this analysis as we had specific directional hypotheses. This means that the study would not be able to reliably detect effects smaller than Cohen's d = 0.76.

TABLE 1.

Demographic and IQ data.

Characteristic	FXS, N = 17 a	TDC, N = 31 a	p b
Sex	4 F (24%), 13 M (76%)	17 F (55%), 14 M (45%)	0.037
Age (years)	9.4 ± 3.0 (5.2–14.7)	9.3 ± 3.0 (5.1–15.5)	> 0.9
Full scale deviation IQ	50.8 ± 18.5	104.8 ± 7.5	< 0.001
Nonverbal deviation IQ	51.4 ± 17.403	105.3 ± 8.5	< 0.001
Verbal deviation IQ	50.1 ± 20.5	104.9 ± 9.5	< 0.001

^a n (%); mean ± SD.

^b Pearson's chi‐squared test; Welch two sample t‐test.

For the FXS group, behavioral and medication regimens were stable for ≥ 30 days (≥ 60 days for non‐stimulant psychotropics). Exclusion criteria were unstable seizure disorder, significant medical or neurological illness, unresolved acute illness, peripheral hearing loss, use of GABAergic or glutamatergic modulators, or inability to comply with study procedures.

All procedures were approved by the Cincinnati Children's Hospital Institutional Review Board (IRB #2015–8425). Written informed consent was obtained from parents or legal guardians, and children provided age‐appropriate assent when possible. For younger children or those with intellectual disabilities, willingness to participate was carefully assessed. Children were monitored for signs of distress or resistance (e.g., crying, attempting to leave, or refusing to wear the EEG net) and the session was paused or discontinued if visible distress was observed.

2.2. Stimuli

Syllables contributing to the continuous statistical learning speech stream were the same as those used by Batterink and Paller (2019). Syllables were recorded by a male native English speaker using neutral intonation and no co‐articulation between syllables. Each syllable was saved within a separate file, with the beginning of each sound file coinciding with the precise onset of the syllable. Syllables were concatenated to create 4 nonsense “words” (“tafuko,” “rugeme,” “repuni,” “fetisu”). To create the continuous speech stream, syllables were presented at a rate of 300 ms per syllable (3.33 Hz) in a predefined pseudorandom order, with the constraint that the same word did not repeat consecutively. Thus, trisyllabic words were presented at a rate of 1.11 Hz. The speech stream contained a total of 2400 syllables (800 words), with each word presented 200 times (12 min total duration).

2.3. Procedure

After electrode setup for EEG recording, participants were seated in a sound‐attenuated booth at a comfortable viewing distance from a computer monitor. The continuous speech stream was presented with Sony MDR‐V150 headphones at 65 dB while participants' EEG was recorded. To facilitate cooperation, participants viewed a silent video of their choice, presented without words or captions, while the speech stream was presented.

2.4. EEG Recording and Analysis

EEG data were recorded throughout the speech stream exposure period at a 1000 Hz sampling rate with an EGI NetAmp 400 with a 128‐channel HydroCel electrode net (Magstim/EGI, Eugene, OR).

2.4.1. Preprocessing

EEG preprocessing and analysis followed the basic approach used in our prior study that used the same recording system (Choi et al. 2020). First, a 60‐Hz notch filter and a band‐pass Butterworth filter from 0.5 to 20 Hz were applied to the raw data. Next, we performed artifact correction using the Artifact Blocking (AB) (Mourad et al. 2007) algorithm to attenuate artifacts (e.g., eye blinks, eye movements, and body movements). As recommended by Fujioka and colleagues, we applied the algorithm after removing the outer ring of channels from further analysis, and we set the threshold value within the artifact‐block algorithm at ±50 μV. Visual inspection of the individual data confirmed that the artifact‐block algorithm successfully corrected artifacts and reduced noise in the data. No data epochs are removed as part of this procedure; artifacts within the epoched datasets are simply attenuated. The artifact‐corrected data were then re‐referenced to a common average. Nonoverlapping epochs of 9.0 s, time‐locked to the onset of every 10th word and corresponding to a duration of 10 words, were then extracted, yielding 78 epochs per participant.

2.4.1.1. Source Localization

To conduct parallel analyses at the source level, the cleaned continuous scalp EEG data was used for source‐modeling with the fsaverage template brain in MNE 1.9 (Gramfort et al. 2010). Individual MRIs and electrode digitization were not available; therefore, source localization was employed primarily as an optimal spatial filter rather than for precise anatomical mapping. This approach was particularly important for separating activity in temporal and frontal regions, given the orientation of Heschl's gyrus and its proximity to frontal sources. The forward solution was computed with a precomputed BEM and ico‐5 source space (minimum source–sensor distance: 5 mm), and the inverse operator was estimated with λ² = 1/9 and fixed‐orientation dipoles. Source time courses (STCs) were parcellated into 68 cortical nodes using the Desikan–Killiany atlas (Desikan et al. 2006) and obtained for each epoch. Following source localization, the same nonoverlapping epochs of 9.0 s were extracted as in the scalp level analysis.

2.4.2. Computation of Intertrial Phase Coherence and Statistical Analysis

For each participant, we quantified neural entrainment by measuring inter‐trial coherence (ITC) across the nonoverlapping epochs. ITC is a measure of event‐related phase‐locking or phase synchronization across trials, which ranges from 0 to 1, with 0 indicating purely non‐phase‐locked (i.e., random) activity at a given frequency band, and 1 indicating strictly phase‐locked activity. Higher ITC values indicate more consistency in the phase of the signal across individual trials (in our case, epochs time‐locked to word onsets). To the extent that statistical learning occurs, we expected to observe higher phase‐locking at the triplet frequency, reflecting greater neural entrainment to the hidden word structure. The Fast Fourier transform was applied to each epoch to decompose the signal into its frequency components, and ITC was then computed across all epochs across frequencies, including each frequency of interest (word and syllable). For scalp level analyses, this procedure was carried out for all electrode channels. For source level analyses, this procedure was conducted for all 68 nodes.

To assess statistical significance of ITC values, for each channel (scalp level analysis) or node (source level analysis), we converted raw ITC values to z‐scores of ITC (zITC), thereby normalizing each individual's ITC values against their own null distribution of non‐entrained activity. To accomplish this, we generated 100 surrogate datasets for each participant. To create each surrogate dataset, each epoch onset was individually shuffled by a different random value between −900 ms and 900 ms from the actual epoch onsets (following Batterink and Zhang 2022; Herrera‐Chaves et al., in press; Moreau et al. 2022). This procedure eliminates the alignment between the EEG signal and the auditory speech stream, while preserving the general features of the data and each epoch's general timing within the experimental task. ITC was then computed for each of the 100 surrogate datasets, resulting in a null distribution of ITC values (at each channel/node). Next, zITC for each electrode was computed using the standard z‐score formula (i.e., subtracting the mean of the surrogate ITCs from the observed ITC, and dividing this value by the standard deviation of the surrogate ITCs).

Given evidence that FXS involves sensory hypersensitivity and exaggerated neural responses to sensory stimulation (Baranek et al. 2008; Miyakoshi et al. 2025; Pedapati et al. 2025), we did not compute the Word Learning Index (the ratio of word‐ to syllable‐rate entrainment), as in some prior studies (Batterink and Paller 2017, 2019; Choi et al. 2020; Moreau et al. 2022; Ringer et al. 2025; Smalle et al. 2022). Instead, we analyzed syllable‐ and word‐rate entrainment separately to allow independent characterization of low‐level sensory tracking and sensitivity to word‐like structures in children with FXS, as described further below.

2.4.2.1. Scalp‐Level Statistical Analysis

Mean zITC values were averaged across electrodes to produce an average entrainment value for each participant, at each frequency of interest (Word, Syllable). We extracted zITC values for all channels because we observed relatively widespread ITC effects across the scalp at both frequencies of interest (see Figure 2). To test whether neural entrainment at each frequency of interest differed between groups, we conducted two separate univariate ANOVA with Group (TDC, FXS) as a between‐subject factor and mean zITC_word (or zITC_syllable) across channels as the dependent measure. We additionally tested whether each group showed significant neural entrainment at each frequency of interest by conducting a one‐sample t‐test against 0, two‐tailed.

FIGURE 2.

EEG results from scalp‐level analyses. (A) Z‐scored ITC values (zITC) as a function of group (TDC in black; FXS in red) and frequency. Children with FXS show a significant reduction in neural entrainment at the word frequency (1.11 Hz); no group differences are observed at the syllable frequency (3.33 Hz). EEG data is averaged across scalp electrodes. Shaded error bars indicate standard error of the mean. (B) Topographical plots showing distribution of zITC across the scalp, as a function of Group and Frequency. (C) zITC values at our two frequencies of interest, summarized from data shown in Panel A. Error bars represent standard error of the mean. * denotes significance at p < 0.05.

2.4.2.2. Source‐Level Statistical Analysis

We examined group differences in source‐level entrainment using linear mixed‐effects models that accounted for the nesting of nodes within cortical regions. Region, Hemisphere (and Node when applied) were sum coded, and Group was treatment coded with the TDC group serving as the reference group.

2.4.2.2.1. Word‐Level Entrainment

For word‐level responses (${\text{ZITC}}_{\text{word}}$), the primary model (WordModel1) included Group (TDC, FXS), Cortical Region (prefrontal, frontal, temporal, central, parietal, lingual, occipital), Hemisphere (left, right), and their interactions, with random intercepts for participant and node:

$${\text{ZITC}}_{\text{word}}~\text{Region}\times \text{Hemisphere}\times \text{Group}+\left(1|\text{SubjectID}\right)+\left(1|\text{Node}\right)$$

If a significant Group × Region interaction was found, this motivated a refined node‐level model (WordModel2):

$${\text{ZITC}}_{\text{word}}~\text{Node}\times \text{Group}+\left(1|\text{SubjectID}\right)$$

Follow‐up comparisons were conducted using estimated marginal means (emmeans), testing group differences at the node level without multiple‐comparison adjustment. For characterization of neural topographies, estimated marginal means were also used to test which nodes showed significant entrainment within each group (ZITC_word against 0).

2.4.2.2.2. Syllable‐Level Entrainment

For syllable‐level responses (${\text{ZITC}}_{\text{syllable}}$), we used an analogous region‐level model (as well as an analogous node‐level model, as above):

$${\text{ZITC}}_{\text{syllable}}~\text{Region}\times \text{Hemisphere}\times \text{Group}+\left(1|\text{SubjectID}\right)+\left(1|\text{Node}\right)$$

2.4.2.2.3. Word Precision Index

Within bilateral transverse temporal gyrus nodes (identified as key regions in our main analyses), we further quantified the precision of phase locking at the word frequency using a novel measure we term the Word Precision Index (WPI). This index reflects the selectivity of neural entrainment to words relative to nearby frequencies and was computed by subtracting the mean zITC values at the neighboring four frequency bins (word freq‐2, word freq −1, word freq +1, word freq +2) from zITC at the word frequency itself. Higher values indicate more precise (accurate) phase locking to the component words in the stream, while lower values indicate relatively imprecise phase locking. Computing the WPI allowed us to test the hypothesis that children with FXS may show not only reduced entrainment to words overall, but also “noisier” (less precise) word entrainment, characterized by erroneous phase‐locking to nearby frequencies. This hypothesis is motivated by prior evidence of imprecise neural entrainment to auditory stimuli in FXS (Ethridge et al. 2017; Pedapati et al. 2025). We tested for group differences in the word precision index using a two‐sample t‐test.

2.4.3. Time Course of Neural Entrainment Over Learning

In our prior studies in healthy infants, children and adults, we have observed an increase in neural entrainment at the word frequency over the exposure period, reflecting the progression of statistical learning over time (Batterink 2020; Batterink and Paller 2017, 2019; Choi et al. 2020; Moreau et al. 2022). These increases appear most consistently within the first half (~6 min) of exposure to the artificial language (Batterink and Paller 2019; Choi et al. 2020; Moreau et al. 2022), after which entrainment values fluctuate around an asymptotic value rather than continuing to increase (Batterink and Paller 2019). Therefore, to quantify the progression of learning over time in the current study, we focused on examining neural entrainment values at our frequencies of interest within the first half of exposure.

Since ITC is undefined at the single trial level, we used a jackknifing approach (Richter et al. 2015; Rimmele et al. 2023; Waschke et al. 2019) to estimate ITC values at the single epoch level at our two frequencies of interest (word and syllable). An advantage of this approach is that it provides single‐trial estimates of entrainment without the need to impose arbitrarily‐sized sliding windows. In this procedure, ITC is initially computed across all epochs (as in our main analysis). Next, ITC is recomputed on all epochs but the first one, which represents one leave‐one‐out jackknife replication. This procedure is repeated for each epoch (systematically leaving out one epoch each time), producing a total of N jackknife replications, where N is the total number of epochs. These jackknifed ITC values represent the values that the overall ITC would have had without that specific trial. We normalized these jackknifed ITC values at each frequency of interest by subtracting the average of the jackknifed ITC estimates for all other frequency bins under 5 Hz, excluding the word and syllable frequencies and their harmonics (i.e., frequencies spanning from 0.66 to 5.0 Hz, excluding frequency bins 1.11, 2.22, 3.33 and 4.44 Hz). For ease of interpretability, we then converted each of these normalized jackknifed ITC values to ITC pseudovalues (using the formula N*ITC_all − (N − 1)*Jackknifed_ITC). The ITC pseudovalue for a given epoch represents the contribution of that epoch to the overall ITC. That is, the higher the pseudovalue, the more a given trial contributes to increasing the overall ITC, as its phase at a given frequency resembles the mean phase at that frequency across trials. Finally, because ITC estimates at the trial level are inherently very noisy, we smoothed the pseudovalues using a moving average with a span of 5 datapoints (i.e., each nth pseudovalue was averaged with neighbors n ‒ 2, n ‒ 1, n + 1 and n + 2) to reduce the influence of outliers in our statistical analysis, which excludes the first two and final two datapoints from further analysis. We hypothesized that these trial‐level ITC estimates (pseudovalues) should increase over time in the TDC group, tracking the process of statistical word learning, and that this increase may be less robust in the FXS group, reflecting impairments in the statistical learning process.

We statistically tested whether neural entrainment at our frequencies of interest changed over time within each group, as well as whether the two groups showed significantly different trajectories. We initially modeled the smoothed ITC pseudovalues at each frequency of interest using a linear mixed model with Epoch Number, Group and their interaction as fixed factors, and participant as a random intercept (Time Course Model: [ITC_Pseudovalue_word ~ EpochNumber*Group + (1 | SubjectID)]). Group was again treatment coded with the TDC group serving as the reference group. In addition, we used emtrends to test our hypothesis that ITC to words would increase over time in the TDC group, and to characterize the trajectory of learning within each group alone. Emtrends is a function in R that allows for estimating linear slopes of a continuous predictor (e.g., epoch number) within levels of a factor (e.g., group), based on a fitted model. Given the results from our main analysis, we conducted this entire procedure at the scalp level (averaging pseudovalues across all electrodes) and at the source level. Source level analyses focused on bilateral transverse temporal gyrus, as our overall source localization analysis at the node level revealed this subregion to be a key hub for neural entrainment, showing the greatest entrainment response to both words and syllables relative to all other nodes/subregions across the brain (see Section 3; Figure 4; Tables S2 and S3).

FIGURE 4.

Neural entrainment to word frequency is reduced in FXS. Lateral views of cortical source‐localized z‐scores of ITC (zITC) for word frequency (top row) and syllable frequency (bottom row) processing for left and right hemispheres. Left column shows FXS group averages, middle column shows TDC group averages, and right column shows group numerical differences (FXS − TDC). Color scales represent zITC values, where positive values (red) indicate greater phase‐locking to the respective frequency compared to surrogate data, and negative values (blue) indicate reduced phase‐locking. Source localization was performed using the Desikan–Killiany atlas with 68 cortical nodes. Note that different scales are used for word and syllable frequencies and for difference effects between groups at each frequency.

3. Results

3.1. Scalp‐Level Analysis

Consistent with our hypothesis, children with FXS showed significantly reduced neural entrainment at the word frequency compared to the TDC group (zITC_word Group effect: (F (1, 46) = 5.11, p = 0.028); Figure 2). Further, while TDC participants showed highly significant word‐level entrainment as a group (ZITC_word > 0: t(30) = 4.23, p < 0.001), word entrainment in the FXS group was not significant (ZITC_word > 0: t(16) = 0.90, p = 0.38). In contrast, syllable‐level entrainment did not differ significantly between the two groups (zITC_syllable Group effect: F (1, 46) = 0.033, p = 0.86), and both groups showed highly significant entrainment at the syllable frequency (both p values < 0.001). Figure S1 shows the data within the FXS group disaggregated by sex. Overall, these results suggest a highly specific and aberrant pattern of reduced neural entrainment to words in the FXS group.

3.2. Source‐Level Analysis

3.2.1. Word Entrainment

The TDC group showed highly robust neural entrainment to words, which was significant across the brain but particularly strong in the temporal cortex (WordModel1: mean parameter estimate: 0.27, SEM = 0.060, p < 0.001; temporal cortex estimate: 0.40, SEM = 0.056, p < 0.001; see Figures 3 and 4; Table S2). No significant main effects of hemisphere or interactions with hemisphere were observed, indicating that entrainment was similarly robust across both hemispheres. A complementary node‐level analysis showed significant entrainment (zITC > 0) across many nodes, with maximal values observed in the transverse temporal cortex (estimate = 1.47, p < 0.001), superior temporal gyri (estimate = 1.08, p < 0.001) and middle temporal gyri (estimate = 0.72, p < 0.001); see Table S3 for full list of significant nodes by group.

FIGURE 3.

EEG results from source‐level analyses; all data shown is drawn from bilateral transverse temporal gyrus (primary auditory cortex), where maximal neural entrainment was observed. (A) Z‐scored ITC values (zITC) as a function of group and frequency. Children with FXS show a significant reduction in neural entrainment at the word frequency (1.11 Hz). Shaded error bars indicate standard error of the mean. Inset—entrainment values at the word and neighboring frequencies are used to compute the word precision index. Children with FXS show significantly reduced precision in word tracking. (B) zITC values at our two frequencies of interest, summarized from data shown in Panel A. Error bars represent standard error of the mean. * denotes significance at p < 0.01.

Of relevance to our main hypothesis that children with fragile X would exhibit reduced neural evidence of statistical learning compared to healthy controls, our initial model revealed a robust Group × Temporal Region interaction (Temporal Region × Group: −0.25, SEM = 0.078, p = 0.002). A follow‐up contrast confirmed that the FXS group showed significantly lower word entrainment than the TDC group within temporal cortex (estimate = 0.33, SE = 0.12, p = 0.004); no other regions showed significant group differences (p > 0.1). A second model at the Node level demonstrated that FXS children showed significantly reduced entrainment within bilateral superior temporal gyrus (TDC − FXS estimate: 0.64, SE = 0.26, z ratio = 2.49, p = 0.013) and within bilateral transverse temporal gyrus (TDC − FXS estimate: 0.90, SE = 0.26, z ratio = 3.49, p = 0.005). In addition, FXS children also showed significantly reduced precision in neural entrainment to the word frequency within the transverse temporal gyrus node, as reflected by the word precision index, relative to TDC children (t(46) = 2.31, p = 0.025; Figure 3A, inset). This result reflects the FXS group's “blurred” entrainment profile of above‐zero entrainment in frequency bins adjacent to the word frequency, indicative of imprecise or inaccurate phase‐locking. In summary, relative to typically developing children, children with fragile X show significantly reduced entrainment to words in both superior temporal and transverse temporal gyrus.

3.2.2. Syllable Entrainment

Within the TDC group, highly robust neural entrainment effects at the syllable frequency were observed (estimate = 2.98, SE = 0.29, p < 0.001). These entrainment effects were maximal over temporal and central cortex (Temporal Region: estimate = 1.71, SEM = 0.32, p < 0.001; Central Region: estimate = 1.53, SEM = 0.51, p = 0.004). Again, no significant main effects of hemisphere or interactions with hemisphere were observed, indicating that entrainment was similarly robust across hemispheres. A complementary node‐level analysis showed significant entrainment (zITC > 0) across all 68 nodes, with maximal values again observed in the transverse temporal cortex (estimate = 8.33, p < 0.001), superior temporal gyri (estimate = 6.84, p < 0.001) and supramarginal gyri (estimate = 6.53, p < 0.001); see Table S4 for all nodes by group.

No significant main effect of group was observed (p = 0.15), although the fragile X group was associated with numerically higher zITC estimates. In addition, in contrast to the general pattern of reduction observed for word entrainment, fragile X children showed significantly stronger syllable entrainment in the left hemisphere and in parietal cortex compared to typically developing children (Group × Hemisphere: estimate = 0.31, SEM = 0.010, p = 0.002; Group × Parietal Region: estimate = 0.053, SEM = 0.025, p = 0.035). This result suggests that the diminished word‐level entrainment in fragile X cannot be attributed to a general reduction in neural responsiveness to auditory stimuli.

3.3. Time Course of Neural Entrainment

3.3.1. Scalp Level (All Scalp Electrodes)

Our time course model revealed that word entrainment in the TDC group increased over the first half of exposure as hypothesized, as reflected by a significant epoch number effect (t(1726) = 2.90, p = 0.004). There was no interaction between epoch number and group (t(1726) = −0.94, p = 0.35), suggesting that the two groups showed a similar increase in neural entrainment over the first half of exposure. However, separate hypothesis‐driven contrasts using emtrends against 0, designed to characterize the trajectory of learning within each group on its own, indicated that while the increase over time in the TDC group was significant (as also shown by the main model), the increase in word entrainment over time within the FXS group was not significant (FXS: slope = 0.0004, SE = 0.0005, t(1726) = 0.98, p = 0.33; TDC: slope = 0.001, SE = 0.0003, t(1726) = 2.9, p = 0.004).

The same analysis computed for syllable‐level entrainment showed a reduction in syllable entrainment over time within the TDC group (epoch number estimate: ‒0.00074, SE = 0.00036, t(1726) = −2.09, p = 0.037). This decrease over time was not significantly different between the two groups (t(1726) = 0.32, p = 0.75).

3.3.2. Source Level (Transverse Temporal Gyrus)

Given that bilateral transverse temporal gyrus nodes were identified as key regions in our main analyses, we focused our source‐level time course analysis on these nodes. Resembling the scalp‐level analysis, within the TDC group, word entrainment increased over time (epoch number: t(1726) = 3.18, p = 0.0015; see Figure 5). There was a marginally significant interaction between epoch number and group (t(1726) = −1.74, p = 0.081), indicating that children with fragile X showed a marginally smaller increase in word entrainment over time relative to the TDC group. Separate contrasts with emtrends, designed to characterize the trajectory of entrainment within each group, showed that whereas word entrainment significantly increased over time within the TDC group, word entrainment within the FXS group did not increase significantly (FXS: slope = 0.00016, SE = 0.00085, t(1726) = 0.19, p = 0.85; TDC: slope = 0.002, SE = 0.00063, t(1726) = 3.18, p = 0.002). We conclude that the TDC group alone showed an increase in word entrainment over time, though we note that with only a trend‐level interaction, we cannot conclude that the two groups showed different trajectories over time.

FIGURE 5.

Time course of neural entrainment to words. ITC estimates at the word frequency for each epoch over the first half of exposure within each group. Only the TDC group shows an increase in neural entrainment at the word frequency over time, characterized by an increase over the first phase of learning, followed by a leveling off. In contrast, the FXS group shows no clear evidence of an increase over time. Shaded regions indicate standard error of the mean, computed from the empirical data at each epoch within each group. Solid lines show model‐predicted marginal means (lines of best fit) across epochs for each group.

The same analysis computed on syllable entrainment revealed no significant change over time within the TDC group (Epoch Number: t(1726) = 0.25, p = 0.80). Compared to the TDC group, the fragile X group showed initially smaller entrainment at baseline (Group: t(55) = −2.7, p = 0.009), coupled with a significantly greater increase over time (Epoch Number × Group: t(1726) = 2.49, p = 0.013). Separate contrasts with emtrends indicated that syllable entrainment significantly increased over time within the FXS group (slope = 0.002, SE = 0.0007, t(1726) = 3.29, p = 0.001), in contrast to the stable pattern observed within the TDC group (slope = 0.0001, SE = 0.0005, t(1726) = 0.25, p = 0.80).

4. Discussion

Supporting our hypothesis that children with FXS would show impairments in the computations underlying statistical learning, we have identified a selective impairment in word‐rate neural entrainment (~1.1 Hz) and reduced learning‐related slope over time in children with FXS. The reduction in statistical learning was robust at the scalp and source levels, with peak effects localized to the primary auditory cortex and adjacent speech‐sensitive cortical regions (represented by the transverse temporal gyri and superior temporal gyrus [STG]). The observed dissociation between impaired word‐level entrainment alongside potentially heightened syllable‐rate entrainment (~3.3 Hz) represents a particularly striking and meaningful distinction, implicating a deficit in longer‐timescale integration, rather than a generalized auditory or attentional limitation in FXS. These findings also highlight word‐level entrainment as a candidate biomarker for language‐related vulnerabilities and treatment monitoring in FXS.

Neural oscillation models of speech processing show that cortical rhythms parse speech across nested timescales, with low frequency rhythms jointly supporting segmentation from syllables to words (Ding et al. 2016; Gross et al. 2013; Poeppel and Assaneo 2020). This hierarchical organization reflects cross‐frequency coupling mechanisms where slower oscillations modulate faster ones, enabling the brain to package incoming speech information into units of appropriate temporal granularity (Giraud and Poeppel 2012). While our entrainment measure does not directly reflect cross‐frequency coupling mechanisms, our word versus syllable dissociation generally maps onto this temporal hierarchy, suggesting that in FXS, the coordinated multi‐frequency dynamics needed for constructing hierarchical linguistic representations, particularly the integration processes operating at word‐level timescales (~1 to 2 Hz), are selectively compromised, while syllable‐level processing remains intact (Zioga et al. 2023).

At a systems level, loss of FMRP in FXS results in disruption of activity‐dependent synaptic regulation (Antoine et al. 2019; Gibson et al. 2008), producing local hyperexcitability, degraded temporal fidelity (Wang et al. 2023), and excitation–inhibition imbalance. These cellular perturbations manifest as abnormalities in large‐scale oscillatory dynamics. In response to an auditory chirp stimulus, participants with FXS demonstrated reduced gamma‐band phase locking, markedly elevated background gamma power, and increases in frontotemporal information flow, suggestive of impaired top‐down regulation (Pedapati et al. 2025). Together with resting‐state evidence of disrupted alpha–theta coupling (Pedapati et al. 2022; Wang et al. 2017), reduced alpha power (Van Der Molen et al. 2014; Van der Molen and Van der Molen 2013), and altered alpha/gamma connectivity (Schmitt et al. 2022), these findings converge on a model of local hyperexcitability and impaired cross‐frequency coordination, which in turn reduce signal‐to‐noise ratio and weaken the scaffolding required for integrating information across timescales. This overall framework helps explain our EEG results: word‐level entrainment (a slower rhythm reflecting word‐level sensitivity or integration) is reduced, whereas syllable‐level entrainment (a faster, locally driven rhythm) remains intact or is even heightened under conditions of hyperexcitability.

It is notable that we found the strongest group differences within STG and transverse temporal gyrus. The finding of robust entrainment within these regions in the healthy control group aligns well with prior work that has implicated both the transverse temporal gyrus (Herrera‐Chaves et al., in press; McNealy et al. 2006) and surrounding STG (Cunillera et al. 2009; Henin et al. 2021; Herrera‐Chaves et al., in press; Karuza et al. 2013; McNealy et al. 2006) as regions forming a central hub in auditory‐linguistic statistical learning. Transverse temporal gyrus is the first cortical region to receive and process raw auditory input, while the STG is a critical locus for the processing of speech sounds, integrating lower‐level, acoustic‐phonetic information with higher‐order contextual and linguistic information to support speech perception (Bhaya‐Grossman and Chang 2022; Yi et al. 2019). Through recurrent connections, the STG generates context‐dependent representations that capture longer temporal sequences and give rise to word representations or perceived whole‐word forms (Bhaya‐Grossman and Chang 2022; Zhang et al. 2026). In the case of statistical learning, the STG's ability to integrate bottom‐up, incoming speech (e.g., a word‐final syllable such as “ko”) with the prior context (e.g., the two syllables “ta” and “fu” as in the word “tafuko”) may give rise to a unique distributed pattern of neural activity, which may then form the basis for identification of repeated patterns in continuous speech, as reflected by entrainment (Endress 2024). Circuits within STG may thus provide a core, automatic mechanism for statistical learning, operating without intentionality or top‐down control—a hallmark of statistical learning. The finding that entrainment is impaired within STG in FXS converges broadly with other evidence suggesting that the STG is a vulnerable region in FXS. For example, individuals with FXS show age‐related decreases in the volume of the STG (Reiss et al. 1994), a pattern not seen in controls, and reduced gray matter volume of the left STG (Sandoval et al. 2018). FXS individuals also exhibit reduced gamma band synchronization to an auditory chirp stimulus in left and right temporal lobes, as revealed through source‐localized EEG (Pedapati et al. 2025).

Learning‐over‐time dynamics also provide mechanistic insight into why individuals with FXS acquire new information more slowly. Word‐rate entrainment in controls, as in previous studies, increased with exposure, representing the accumulating knowledge of word‐like regularities (Batterink 2020; Batterink and Paller 2017, 2019; Choi et al. 2020). In children with FXS, the word learning rate was essentially flat, suggesting reduced capacity to accrue longer timescale learning, and is consistent with broader findings of slower learning rates and reduced adaptation in FXS across multiple domains (Goel et al. 2018; Knox et al. 2012; Schmitt et al. 2023). Though results fell short of showing significant group‐level differences, the flat entrainment trajectory in children with FXS is suggestive of reduced statistical learning abilities in this group, and may contribute to their known deficits across many components of language (Finestack et al. 2009; Hoffmann 2022).

A strength of the study is that we used a behaviorally‐independent, neural measure of statistical learning, which allowed for inclusion of more severely affected participants in our FXS sample who would have had difficulty performing behavioral tasks. However, a limitation of the current study is that the FXS sample was underpowered to test possible correlations between our entrainment measure of statistical learning and cognitive outcomes such as IQ, especially given the wide age range of the sample. In healthy adults, both EEG (van der Wulp et al. 2025) and behavioral measures of statistical learning (Siegelman and Frost 2015) have been found to be largely unrelated to general cognitive abilities, such as IQ, verbal working memory, and rhythmic abilities, suggesting that in normal development, statistical learning operates as a basic mechanism that does not limit higher‐level cognition. In contrast, if functioning of the normal statistical learning capacity breaks down or is atypical, as is the case in NDDs such as FXS, statistical learning might relate more strongly to general cognitive outcomes, in line with prior findings that populations outside of typical development show weaker statistical learning (Evans et al. 2009; Gabay et al. 2015; Lammertink et al. 2017; Vandermosten et al. 2019; Zhang et al. 2021). Testing a possible direct link between altered statistical learning and cognitive outcomes will be an important direction for future work. In addition, while our observed group differences had large effect sizes (Cohen's d ~ 0.8), our samples were underpowered to detect possible true effects that were smaller in size. Future work incorporating large sample sizes may allow for detection of smaller, more nuanced differences between groups that were not uncovered in the current study.

To conclude, these results indicate aberrant neural entrainment to higher‐order regularities in FXS, consistent with a specific deficit in assembling multi‐syllabic units from continuous speech, rather than a global impairment of auditory responsiveness. An important distinction in our findings is the preserved syllable‐level entrainment at 3.3 Hz versus impaired word‐level entrainment at 1.1 Hz in FXS participants. The pattern suggests that children with FXS can synchronize to rapid, lower‐order acoustic structure but struggle to gain sensitivity to the longer, chunk‐like representations needed for word segmentation. Similarly, the time‐course analysis demonstrated a characteristic build‐up of word‐locked entrainment in control participants, but no reliable increase in FXS. In a clinical context, there is early evidence that entrainment growth can be enhanced (i.e., through transcranial magnetic stimulation) (Smalle et al. 2022) and thus may represent mechanistically grounded biomarker (Sahin et al. 2019) that could quantify both learning capacity and treatment efficacy in FXS interventions targeting language development.

Funding

This work was supported by the Eunice Kennedy Shriver National Institute of Child Health and Human Development (5R01HD108222‐03).

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: CELF‐5 scaled score descriptive statistics by group.

Table S2: Results of main linear mixed effects model applied at the source level.

Table S3: Significant node‐level word entrainment by group.

Table S4: Significant node‐level syllable entrainment by group.

Figure S1: ITC data within the FXS group, disaggregated by sex. Although we are underpowered to detect sex differences, with only four female FXS participants, the data shown present ITC as a function of sex for descriptive purposes. Error bars represent standard error of the mean.

Data Availability Statement

The data that support the findings of this study are openly available in Zenodo at https://doi.org/10.5281/zenodo.17435554. Data analysis scripts are available at https://doi.org/10.5281/zenodo.18746668.