Resting state aperiodic and periodic EEG activity in preschool-aged autistic children: differences from neurotypical peers and links to language skills

Abstract

Background

The neural mechanisms underpinning language development in autism spectrum disorder remain unclear. While prior studies have identified associations between resting-state EEG absolute power and language skills in autistic children, none have examined the distinct roles of aperiodic and periodic activity decomposed from the absolute power spectra on language development in autistic children.

Methods

At the group level, we examined resting-state power spectra differences between 64 neurotypical and 64 age-matched autistic children from 2 to 6 years old, as well as within the autistic group based on language impairment status. At the individual level, we examined whether aperiodic and periodic EEG features were associated with concurrent language skills measured by natural language samples and a standardized language assessment in participants with autism spectrum disorder.

Results

Autistic children exhibited higher aperiodic offset, but not aperiodic slope, compared to their neurotypical peers. While we did not find significant differences in peak alpha frequency and alpha peak amplitude between neurotypical and autistic children, after separating the autistic group by language impairment status, we found that autistic children with language impairment had significantly lower alpha peak amplitude compared to autistic children without language impairment and their age-matched neurotypical peers. Regarding the brain-language association, autistic children with lower aperiodic offset demonstrated better concurrent expressive and receptive language skills, but not nonverbal developmental quotient. Autistic children with higher alpha peak amplitude demonstrated better concurrent language and nonverbal developmental skills.

Limitations

Findings were based on cross-sectional data from children with relatively higher socioeconomic backgrounds. Future studies are needed to explore the longitudinal relation between resting-state EEG aperiodic and periodic features and language development in autistic children, while accounting for potential confounding effects of demographic variability and data quality.

Conclusions

The characteristic features of resting-state power differences vary when comparing autistic children to neurotypical peers versus comparing within the autistic subgroup based on language phenotypes. These findings underscore the importance of considering the heterogeneity of the autism spectrum when investigating the neural mechanisms underlying language development in autistic children.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13229-025-00700-1.

Background

Language skills are highly heterogeneous in children with autism spectrum disorder (ASD), a neurodevelopmental disorder defined by persistent challenges in social communication and interaction as well as restricted and repetitive behaviors and interests [1]. A significant portion of autistic children have language impairment (e.g. [2]), among whom around 30% remaining minimally verbal into adulthood despite receiving years of intervention [3]. Even among this minimally verbal subgroup, their expressive-receptive language profiles exhibit considerable variability [4]. Given this, it remains one of the most pressing questions to investigate what drives the heterogeneous language developmental patterns and outcomes in autistic children.

Electroencephalography (EEG) is a non-invasive neuroimaging technique that directly measures the electrophysiological dynamics of the brain [5]. Resting state EEG, which captures spontaneous neural oscillations when the brain is not responding to any specific task or sensory stimulation, can provide meaningful information about the neural mechanisms related to language development [6]. Resting-state EEG can be decomposed into different frequency bands (i.e., delta, theta, alpha, beta, and gamma), with each band associated with characteristic power distributions that reflect different aspects of neural activity, including cognitive processes, arousal states, and network connectivity [7]. Examining resting-state power differences between neurotypical and autistic children, as well as within the autistic group comparing those with and without language impairment, provides important insight into whether observed alterations in EEG features are primarily associated with an autism diagnosis or with language impairment.

Resting-state power differences between neurotypical and autistic children

Findings on power differences across frequency bands between neurotypical and autistic children have been mixed. Most studies on power differences in the alpha band (typically defined as 6–12 Hz) report that autistic children exhibited decreased absolute or relative alpha power compared to neurotypical peers [8–11]; see [12] for a comprehensive review), suggesting diminished cortical inhibition at rest [13]. However, other studies (e.g. [14, 15]) found no significant differences. In addition to these power differences, alterations in peak alpha frequency (PAF; the frequency at which absolute alpha power is maximal), a marker of neural network maturation that typically increases with age [16, 17], have also been observed in autism. Specifically, Dickinson et al. [18] found significantly lower PAF in autistic children, suggesting delayed maturation of neural networks relative to neurotypical peers.

Resting-state power differences between neurotypical and autistic children have also been examined in high-frequency bands (beta 12–30 Hz and gamma 30–55 Hz). While several studies reported increased absolute beta and gamma power in autistic children compared to neurotypical peers [19, 20], other have found decreased beta [14], and some reported no significant differences in gamma power during early childhood [21].

Given the mixed findings in previous studies, a few systematic reviews and meta-analyses have been conducted to synthesize resting-state power differences between neurotypical and autistic children. In a review of 14 EEG studies, Wang et al. [22] identified a distinct “U shaped” EEG power spectrum profile in individuals with ASD, with increased power in low (i.e., delta, theta) and higher frequency bands (i.e., beta, gamma) and decreased power in the alpha band compared to neurotypical controls. However, a recent comprehensive meta-analysis synthesizing 41 studies revealed slightly different findings, showing that autistic individuals exhibited reduced relative (but not absolute) alpha power, increased absolute and relative gamma power, but similar delta, theta, and beta power compared to neurotypical individuals [23]. Inconsistent findings in prior studies of resting-state EEG power differences may be due to the heterogeneity of ASD driven by complex interactions between genetic and environmental factors (e.g. [24]). Despite these mixed findings, prior research suggests an overall trend of reduced alpha power and elevated gamma power in autistic children.

Resting-state brain oscillations and language development

The relationship between resting-state EEG and language development has been investigated in neurotypical and autistic children, using both cross-sectional and longitudinal designs, with a particular focus on gamma power. While several studies have reported a positive association between gamma power and language abilities in neurotypical toddlers and children [25–28], this association was reported to be the opposite in autistic children: Wilkinson et al. [29] found that reduced frontal gamma power was associated with better expressive, but not receptive, language skills in 2-year-old toddlers at elevated likelihood for ASD. A similar finding was also reported for preschool-aged autistic children [21]. It has been hypothesized that the increased gamma power observed in autistic children reflects alterations in the balance between neural excitation and inhibition, and excessive excitation relative to inhibition may adversely affect language processing [30]. The relationship between resting-state EEG power and language development in autism has been primarily examined through correlation analysis in prior studies. There remains a need to extend the investigation to examine the potentially distinct resting-state EEG profiles of different ASD language subtypes, particularly those with and without language impairment.

Resting-state aperiodic and periodic neural activity

To date, the studies including autistic children have focused on absolute power to examine the association between resting-state EEG power and language skills. Absolute power comprises both aperiodic and periodic activity (Fig. 1a [31]). Specifically, aperiodic activity includes two interrelated but distinct components (Fig. 1b): the aperiodic offset, reflecting non-oscillatory, continuous background activity or broadband firing in the brain, and the aperiodic slope, which follows a 1/f power law distribution. The aperiodic slope has been suggested to reflect the balance of excitation (E) and inhibition (I) within cortical circuits, such that a lower, flattened slope is associated with increased excitation over inhibition, whereas a higher, steeper slope is associated with increased inhibition over excitation [32]. In contrast, periodic power refers to the oscillatory activity occurring within specific frequency bands, representing the portion of the power spectrum that rises above the aperiodic slope (Fig. 1c). Importantly, parameterization of the absolute power spectrum also improves the accuracy of peak frequency and amplitude measurements of oscillatory/periodic components [31, 33].

Fig. 1.

Resting-state EEG power between neurotypical (N = 64) and autistic (N = 64) children note. A. absolute power spectrum: N = 64 neurotypical (blue) and N = 64 autistic children (orange). Shading error describes 95% confidence intervals. B. SpecParam estimated aperiodic power spectrum. C. aperiodic-adjusted periodic power spectrum, calculated by subtracting the SpecParam estimated aperiodic power from the absolute power spectrum. D. SpecParam estimated aperiodic offset between TD and ASD groups. E. SpecParam estimated aperiodic slope between TD and ASD groups. F. alpha peak amplitude between 6 and 12 Hz. G. frequency of identified maximal peak in the alpha band. p values represent independent t-test comparisons between TD and ASD groups

A growing body of work suggests an E/I imbalance in individuals with ASD (e.g. [34, 35]), which may be related to the reduced levels of the neurotransmitter gamma-aminobutyric acid (GABA) in certain cortical regions of the brain [36, 37] and reduced thalamocortical connections resulting in alterations in neural regulation [38, 39]. An E/I imbalance has also been linked to greater autistic traits in young children with elevated likelihood of neurodevelopmental disorders [40], and decreased inhibition has been associated with lower listening comprehension skills in late childhood and early adolescence [41].

However, how the physiologically distinct aperiodic and periodic EEG features relate to language abilities has not been examined in preschool-aged children with ASD. It remains unclear whether the associations between gamma power and expressive language in autistic children identified in prior studies are driven by aperiodic or periodic components within this high-frequency range. In addition, previous research on resting-state EEG power in autistic children most often compared EEG patterns from neurotypical peers at the group level, without distinguishing different language profiles across individuals with autism. However, given the wide range of language abilities often observed in autistic children, examining both the continuous association between resting-state power and language skills, as well as the resting-state power differences between autistic children with and without language impairment, can provide deeper insights into the relationship between neural activity and language development in autism.

The current study

The first aim of the current study was to characterize and compare resting-state aperiodic components and periodic power between a large cohort of preschool-aged neurotypical and autistic children. We were specifically interested in examining the differences in 1) aperiodic offset, aperiodic slope, and aperiodic total power; and 2) peak alpha frequency and alpha peak amplitude between age-matched neurotypical and autistic children. We hypothesized that autistic children would have higher aperiodic offset and aperiodic total power, particularly aperiodic gamma power, compared to age-matched neurotypical peers due to increased non-oscillatory neural spiking activity. Given the growing hypothesis that autism may be related to E/I imbalance [34, 38], we hypothesized that autistic children would have a lower, flattened aperiodic slope, indicating increased excitation over inhibition, compared to neurotypical peers. In addition, in line with findings in Dickinson et al. [18], we hypothesized that autistic children would exhibit lower PAF and lower alpha peak amplitude than neurotypical peers.

The second aim of the study was to investigate the concurrent relationship between resting-state aperiodic and periodic EEG components and language skills in preschool-aged autistic children. To broaden our understanding of neural oscillations associated with language ability in preschool-aged children, we analyzed multiple EEG features, including aperiodic offset, aperiodic slope, aperiodic gamma power (given the significant association between gamma power and language in previous autism literature), peak alpha frequency, and alpha peak amplitude. We hypothesized that autistic children with lower aperiodic offset, lower aperiodic gamma power, higher peak alpha frequency, and higher alpha peak amplitude would have better language skills.

Methods

Participants

EEG data from a total of 154 children (81 neurotypical and 73 autistic) between 2 and 6 years old were analyzed. Children were recruited as part of two studies: Predicting and Optimizing Language Outcomes (POLO) project at Boston University and Brain Indicators of Development Growth (BRIDGE) project at Boston Children’s Hospital. EEG data were collected in similar ways on the same type of equipment in the two studies following similar protocols.

In POLO, children enrolled in the neurotypical (TD) group did not have a suspected or confirmed diagnosis of any neurodevelopmental disorder and did not have first-degree family members diagnosed with ASD. They also did not meet the autism spectrum cut-off on the Autism Diagnostic Observation Schedule, second edition (ADOS-2 [42]). Children enrolled in the autistic group had a clinical diagnosis of ASD that was confirmed by meeting the cut-off on the ADOS-2 administered by research-reliable clinicians. Autistic children with a co-occurring genetic disorder were not excluded from POLO. None of the POLO participants had a formal diagnosis of Fragile X syndrome, however two autistic children had specific genetic disorders (i.e., RBFOX1 gene mutation and 3q29 Deletion) and were included in our analysis. In BRIDGE, children enrolled in the neurotypical group did not have any diagnosis of a neurodevelopmental disorder. Participants in the autistic group had a clinical diagnosis of ASD provided by either a developmental behavioral pediatrician and/or clinical psychologist.

Overall inclusion criteria for POLO also included the following: 1) English was the primary language spoken at home or at school; 2) There was no uncorrected vision or hearing impairment; and 3) walking onset occurred before 24 months. Ethical approval of POLO was obtained from the University Institutional Review Board (IRB) at Boston University (IRB-Protocol 5372E). For the BRIDGE study, all children were exposed to English at least 75% of the time across their waking hours. Children were excluded for any history of prematurity ( < 34 weeks gestational age), known birth trauma, unstable seizure disorder, current use of anticonvulsant medications, and uncorrected hearing or vision impairment. IRB approval of BRIDGE was obtained from Boston Children’s Hospital (IRB-P00034676). All study procedures were conducted following the guidelines of the Declaration of Helsinki. Legal guardians of all participants provided written informed consent before the onset of the study in both sites.

After checking for EEG data quality (see below EEG pre-processing and rejection criteria), 77 (95.06% of those who were able to complete EEG) neurotypical and 64 (87.67%) autistic children provided sufficient valid resting-state EEG data. For Aim 1, we further matched the ages of neurotypical and autistic children in our sample using 0.5-year age bands to control for the age effect on resting-state EEG features, given the rapid neurological development of the brain during early childhood. The final sample for Aim 1 included age-matched 64 neurotypical (M_age = 4.51 years, SD = 1.10) and 64 autistic children (M_age = 4.60 years, SD = 1.10), t_age (126) = −0.485, p = 0.628. We were not able to exactly match on biological sex in our final neurotypical and autistic sample for Aim 1 (TD Males n = 42, ASD Males n = 53). The autistic group was further categorized based on the status of language impairment (defined as 1 SD below the population norm on a standardized language measure of receptive language; see below): 37 (57.81%) autistic participants had language impairment (ASD-LI), whereas 26 (40.63%) demonstrated an age-expected language level (ASD-NL). One (1.56%) autistic participant didn’t complete the standardized language measure and was not assigned a language impairment status. For Aim 2, we included all 64 autistic children (M_age = 4.60 years, SD = 1.10) who provided sufficient valid resting-state EEG data to maximize the statistical power for the brain-behavior analyses (see Tables 1 and 2 for participant characteristics).

Table 1.

Demographic characteristics of child participants (N = 128)

	Combined N = 128	TD N = 64	ASD N = 64	Group Comparisons
	n (%a)	n (%a)	n (%a)	X2	p
Sex				4.08	0.043
Female	33 (25.8)	22 (34.4)	11 (17.2)
Male	95 (74.2)	42 (65.6)	53 (82.8)
Ethnicity				7.98	0.019
Hispanic	16 (12.5)	3 (4.7)	13 (20.3)
Non-Hispanic	111 (86.7)	60 (93.8)	51 (79.7)
Prefer not to answer or unknown	1 (0.8)	1 (1.6)	0
Race				17.55	0.002
White	76 (59.4)	36 (56.3)	40 (62.5)
Black or African American	8 (6.2)	0 (0)	8 (12.5)
Asian	18 (14.1)	9 (14.1)	9 (14.1)
American Indian/Alaska Native	2 (1.6)	0 (0)	2 (3.1)
More than one race	23 (18.0)	18 (28.1)	5 (7.8)
Prefer not to answer or unknown	1 (0.8)	1 (1.6)	0
Household Annual Income				8.06	0.045
>$100,000	93 (72.7)	53 (82.8)	40 (62.5)
$40,000–$100,000	25 (19.5)	9 (14.1)	16 (25)
$40,000 or less	3 (2.3)	0 (0)	3 (4.7)
Not Answered	7 (5.5)	2 (3.1)	5 (7.8)
Mother Education Level				14.04	0.003
Graduate/Professional degree	65 (50.8)	40 (62.5)	25 (39.1)
Bachelor’s Degree	44 (34.4)	21 (32.8)	23 (35.9)
Some college or special training after GED	14 (10.9)	1 (1.6)	13 (20.3)
High school/GED	5 (3.9)	2 (3.1)	3 (4.7)
Father Education Level				21.93	0.001
Graduate/Professional degree	62 (48.4)	39 (60.9)	23 (35.9)
Bachelor’s Degree	35 (27.3)	20 (31.3)	15 (23.4)
Some college or special training after GED	15 (11.7)	1 (1.6)	14 (21.9)
High school/GED	11 (8.6)	4 (6.3)	7 (10.9)
Junior High	2 (1.6)	0 (0)	2 (3.1)
Not Answered	3 (2.3)	0 (0)	3 (4.7)

Note. Chi-square tests were used to determine demographic characteristic differences between the TD and ASD group

Table 2.

Means and standard deviations of child characteristics variables by group

Child Characteristics	Participants						Group Differences
	TD (N = 64)			ASD (N = 64)
	M	SD	Range	M	SD	Range
Chronological Age (years)	4.51	1.1	2.0–6.6	4.60	1.1	2.3–6.8	t (126) = −0.485, p = 0.628
PLS-5 Receptive Language	115.44	14.94	79–147	77.84	23.23	50–130	U (127) = −7.97, p < 0.001
ELSA Expressive Language
NDW	175.57	58.57	46–267	78.27	85.26	0–353	U (95) = −5.35, p < 0.001
Utterance Rate	6.91	1.94	2.4–9.9	3.64	3.49	0–11.7	U (95) = −4.61, p < 0.001
Intelligibility (%)	91.97	6.83	71.8–100	70.44	31.05	2.3–100	U (84) = −3.38, p = 0.001
MSEL or DAS-II NVDQ	105.87	20.25	63.6–203.4	67.44	26.63	11.5–125.3	U (123) = −7.39, p < 0.001

Note. PLS-5 Receptive Language score = the PLS-5 Auditory Comprehension subscale standard score; NDW = number of different words; Utterance Rate = the number of utterances per minute; Intelligibility = the percentage of intelligible utterances. MSEL or DAS-II NVDQ = Mullen Scales of Early Learning or Differential Ability Scales nonverbal developmental quotient score. Independent t test was used to compare group difference in chronological age as it was normally distributed in both groups. Mann-Whitney U tests were used to compare group differences in other child characteristics, as they all violated normality, which was required for parametric tests

* p < 0.05. ** p < 0.01. *** p < 0.001

Statistical power analysis using G*Power 3.1.7 [43] indicated that the current sample size (N = 64 TD and N = 64 ASD between-group; N = 37 ASD-LI and N = 26 ASD-NL within ASD group) would allow us to detect medium (Cohen’s d = 0.499) to large effect sizes (Cohen’s d = 0.72) for two-tailed t-tests with 0.05 significance level and 0.8 power. Similarly, the sample size for the ASD group (N = 64) would allow us to detect medium effect sizes (f² = 0.158, R² = 0.166) for regression models (Fixed model, R² increase) and medium effect sizes (r = 0.246) for correlations with 0.05 significance level and 0.8 power. Note that the power calculations were based on individual tests, and the actual power to detect effects after correcting for multiple comparisons may be lower.

Measures

Receptive and expressive language skills

The Preschool Language Scales, Fifth Edition (PLS-5 [44]) is a standardized assessment of developmental language skills in children from birth to 7 years, 11 months. The PLS-5 has excellent test–retest reliability with average coefficients from 0.86 to 0.95, excellent internal consistency with an average coefficient of 0.91, and good validity [44]. The POLO study only administered the PLS-5 Auditory Comprehension subscale to measure participants’ receptive language and used natural language sample (see below) to measure their expressive language. The PLS-5 Auditory Comprehension subscale standard score was used to determine whether autistic participants had language impairment. Children with a PLS-5 standard score more than 1 SD below the normed average score (i.e., auditory comprehension standard score lower than 85) were considered to have language impairment (ASD-LI), Otherwise, they were considered to have language skills in the normal range (ASD-NL).

Eliciting Language Samples for Analysis (ELSA [45]) is a protocol designed to assess expressive language and communication in autism through a natural language sample. ELSA includes eight play-based activities to elicit expressive communication and takes 15 to 25 minutes to administer. ELSA has demonstrated good reliability and validity and can be used for autistic individuals with minimal or low-verbal abilities [45]. Each ELSA administration was transcribed per verbatim by a trained transcriber and checked by a second transcriber. Any disagreements were resolved through consensus. All ELSA transcriptions were then processed through the Systematic Analysis of Language Transcripts (SALT) software [46] to derive the following variables to reflect the participants’ expressive language skills: number of different words (NDW), frequency of utterances per minute (utterance rate), and the percentage of intelligible utterances (intelligibility).

Nonverbal developmental quotient

Nonverbal developmental quotient (NVDQ) of the participants was measured by either the Mullen Scales of Early Learning (MSEL [47]) or the Differential Ability Scales, Second Edition [48] depending on the ages of the participants. Participants younger than 68 months were assessed using the MSEL, whereas those older were administered the DAS-II. The MSEL [47] is a standardized assessment to evaluate early motor, visual, and language skills for children from birth to 68 months old. In this study, NVDQ was measured by the MSEL Visual Reception and Fine Motor subscales. The MSEL NVDQ score was calculated by averaging the age equivalent scores of the Visual Reception and Fine Motor subscales and dividing by chronological age and then multiplied by 100. Higher NVDQ scores indicate higher levels of nonverbal developmental skills.

The Differential Ability Scales, Second Edition [48] is a standardized assessment for assessing cognitive abilities of children from ages 2:6 to 17:11 and has demonstrated good reliability and validity [48]. The Early Years Battery Upper Level (for children aged 3:6 to 6:11) was used in this study. NVDQ was assessed through two subscales, Matrices and Picture Similarities. The DAS-II Early Years NVDQ score was calculated by averaging the age equivalent scores of the Matrices and Picture Similarities subscales and dividing by chronological age, multiplied by 100. Higher NVDQ scores indicate higher levels of nonverbal developmental skills.

EEG data acquisition

Resting-state EEG was collected in an electrically shielded laboratory setting. The participants sat in a comfortable chair with their parents and a research assistant by their side in a dimly lit, sound-attenuated room with a low electrical signal background. For the POLO study, children watched a silent screensaver of abstract moving objects for about 5 minutes. For the BRIDGE study, children watched a preferred video of their choice with the sound turned off for about 5 minutes. Resting-state EEG was recorded in the eyes-open condition using a 128-channel Hydrocel Geodesic Sensor Net (Electrical Geodesics Inc., Eugene, OR) suitable for each participant’s head circumference, which was connected to a NetAmps 300 amplifier (Electrical Geodesic Inc.). Electrooculographic electrodes of the net were removed for the participants’ comfort. Data were referenced to the single vertex electrode, Cz, and impedances were controlled below 50kΩ. The sampling frequency was 1000 Hz for both POLO and BRIDGE. Participants’ behaviors were video recorded during EEG collection. Data segments in which participants were talking, touching the net, or when the caregiver was speaking were manually identified and were excluded from analyses.

EEG pre-processing

The resting-state EEG raw NetStation (Electrical Geodesics Inc.) files were exported to MATLAB (version R2022a) for preprocessing using the Batch Electroencephalography Automated Processing Platform (BEAPP [49]) with the integrated Harvard Automated Preprocessing Pipeline for EEG (HAPPE [50]). A 1-Hz high-pass and a 100-Hz low-pass filter were applied to all files. All data were resampled to 250 Hz. Artifact removal by the HAPPE pipeline includes removing 60 Hz line noise, rejecting bad channels, and removing artifacts using combined wavelet-enhanced independent component analysis (ICA) and Multiple Artifact Rejection Algorithm (MARA). As in prior studies with similar length of data (e.g. [51]), a total of 36 channels were selected for MARA and further analysis, including the 10–20 electrodes (i.e., 22, 9, 33, 122, 24, 124, 11, 45, 36, 108, 104, 58, 96, 52, 92, 62, 70, 83) and additional channels selected to cover the whole brain region: 28, 19, 4, 117, 13, 112, 41, 47, 37, 55, 87, 103, 98, 65, 67, 77, 90, 75 (see Supplementary Figure S1 and Supplementary Material for a sensitivity analysis restricted to electrodes within the central region of interest). Channels removed from bad channel rejection were then interpolated after artifact removal. Data were referenced to the average reference, detrended to the signal mean, and segmented into 2-second non-overlapping segments. Any segments with retained artifact were rejected using HAPPE’s amplitude and joint probability criteria.

EEG rejection criteria

Using HAPPE data quality metrics [50], EEGs were rejected if they had fewer than 20 good segments (2s per segment with a total of at least 40s), percent of good channels < 80%, percent of independent components rejected as artifact > 80%, mean retained artifact probability > 0.3, and percent of EEG signal variance retained after artifact removal < 25% (see Table 3 for EEG data quality comparison between the neurotypical and autistic groups). Of the 154 participants with EEG recordings, four autistic children were excluded due to fewer than 20 segments, three from the autism and one from the neurotypical group were excluded for lower than 25% of EEG signal variance retained after artifact removal, and five (two autistic and three neurotypical) were excluded due to usability concern documented during data collection, resulting in a final 141 EEGs (77 from neurotypical and 64 from autistic children). After matching the ages of neurotypical and autistic children, Table 3 showed the EEG data quality between 64 neurotypical and 64 autistic children in the final sample.

Table 3.

EEG data quality comparison between neurotypical (N = 64) and autistic children (N = 64)

	TD (N = 64)		ASD (N = 64)		Za	p
	M	SD	M	SD
Number of Kept Segments	111.47	30.25	84.86	34.48	4.20	< 0.001
Percent of Good Channels	93.53	3.77	92.84	4.05	0.77	0.442
Percent of Rejected Independent Components	31.03	13.72	36.80	10.76	−2.65	0.008
Mean Retained Artifact Probability	0.09	0.04	0.11	0.04	−2.35	0.019
Percent of EEG Signal Variance Retained	71.89	17.68	64.89	17.33	2.19	0.029

Note. ^a Mann-Whitney U tests were used to compare group differences in EEG data quality metrics between the TD and ASD group. Standardized test statistic Z-values were reported

EEG power spectra analysis and parameterization

Using the BEAPP Power Spectral Density (PSD) module, the PSD at each electrode, for each 2-second segment, was calculated using a multitaper spectral analysis [52] with three orthogonal tapers. For each electrode, the PSD was first averaged across segments and then averaged across the electrodes across the whole brain region (see Supplementary Figure S1). The original frequency resolution of the PSD was 0.5 Hz, which was then linearly interpolated to 0.1 Hz intervals for consistent frequency alignment across participants.

The power spectra were then decomposed for aperiodic and periodic components using a modified version of SpecParam v1.0.0 [31] (https://github.com/fooof-tools/fooof; in Python v3.11.4). SpecParam was modified to better fit the data in the study (see [51] Supplemental Figs. 7–8 for more details on how the original algorithm was modified). The SpecParam model was used in the fixed mode (no spectral knee), evaluating spectra between 2 and 55 Hz, with peak width_limits set to [0.5, 18.0], max_n_peaks = 7, and peak_threshold = 2. Mean estimated error for the full sample was 0.013 (SD = 0.008; range 0.003–0.047). Mean R² for the full sample was 0.997 (SD = 0.0029; range 0.980–0.999). Neurotypical and autistic groups did not differ in either mean estimated error [U (128) = 0.253, p = 0.801] or mean R² [U (128) = −0.257, p = 0.797]. There was also no significant difference in R² [U (63) = −1.90, p = 0.07] between ASD-NL and ASD-LI groups. The ASD-NL group had slightly higher mean estimated error (M = 0.016, SD = 0.01) than the ASD-LI group (M = 0.011, SD = 0.01), U (63) = 2.485, p = 0.013.

Two aperiodic features, aperiodic offset and slope, were extracted for analysis after parameterization. As previously described in Wilkinson et al. [51], we define the aperiodic offset as the aperiodic power at 2.5 Hz, given the elevated levels of error observed in SpecParam estimates at frequencies below 2.5 Hz (see Supplementary Figure S2). Main findings of the study remained the same using aperiodic features parameterized from the original and modified versions of SpecParam. Periodic power spectra were calculated by subtracting the SpecParam-estimated aperiodic power from the absolute power spectrum. We calculated periodic power using the integral under the curve for the following frequency ranges: delta (2.5-4 Hz), theta (4-6 Hz), alpha (6-12 Hz), low beta (12-20 Hz), high beta (20-30 Hz), and gamma (30-55 Hz). Next, to characterize peaks for each child, a savgol filter (scipy.signal.savgol_filter, window length = 101, polyorder = 8) was used to smooth the periodic power spectra, and then peak maxima were identified within the alpha frequency range (6-12 Hz). Code used for EEG processing and SpecParam analyses, including modifications made to SpecParam, can be found on the Open Science Framework (https://osf.io/u3gp4/files/osfstorage; 01- 03).

Data preparation and analytic plan

All statistical analyses and visualizations were conducted in Python (version 3.11.4) and R (version 4.4.1). The distributions of scores from the standardized measures were first tested for normality through the Shapiro-Wilk test, univariate histograms, and normal Q-Q plots. For the behavioral variables, the scores from the PLS-5 receptive language subscale, ELSA expressive language measures, and MSEL or DAS nonverbal developmental quotient in autistic children all violated normality. Therefore, nonparametric tests were used for analyses that involved these variables.

For the first aim of the study, we used independent sample t-tests to compare aperiodic and periodic components between age-matched neurotypical and autistic samples. For the second aim of the study, we used partial nonparametric correlations (Spearman) and censored regression models to examine the concurrent association between resting-state EEG features and the variables of language skills and NVDQ while controlling for chronological age as a covariate. Partial nonparametric correlations (Spearman) were used for variables that were not normally distributed but showed sufficient variability across the range without a significant amount of data clustered at floor (i.e., NVDQ scores). Censored regression models [53] were used for variables with a significant amount of data clustered at floor (e.g., NDW with 20% of the sample having a 0; PLS auditory comprehension standard score with 27% of the sample at floor) and conducted through the R package AER [54]. Here are two example codes of conducting a censored regression model in R:

tobit_model < −tobit (NDW ~ whole_aperiodic_offset + Age, left = 0, data = dat), where the left-censoring point is 0 because it is the lowest score a child can get for NDW (i.e., having 0 word; nonverbal).

tobit_model < −tobit (ACSS ~ whole_aperiodic_offset + Age, left = 50, data = dat), where the left-censoring point is 50 because it is the lowest standard score a child can get in the PLS-5 auditory comprehension subscale.

Given that we tested the same EEG feature with multiple behavioral variables, we used the Benjamini-Hochberg method [55] to account for multiple comparisons with a false discovery rate (FDR) of 5%, such that no more than 5% of the significant results to be false positives. FDR correction was applied separately within each cluster of tests.

Results

Aperiodic offset and slope differences

Autistic children exhibited significantly higher aperiodic offset (M = 0.31, SD = 0.17) than their neurotypical peers (M = 0.25, SD = 0.14), t (126) = 2.289, p = 0.024; Cohen’s d = 0.405, 95% CI: [0.054, 0.754] (Fig. 1 D). However, there was no significant difference in aperiodic slope between the two groups, t (126) = 0.195, p = 0.846 (Fig. 1 E). Within the autistic group, there were also no significant differences in aperiodic offset, t (61) = 1.66, p = 0.101, and slope, t (61) = 1.50, p = 0.139, between autistic children with and without language impairment. Aperiodic offset and slope were not significantly correlated with each other, r (128) = 0.154, p = 0.083.

Aperiodic and periodic power differences

Given that prior studies have primarily used canonical frequency bands for analysis, we first examined group differences in aperiodic and periodic power across these frequency bands to assess whether previously reported alterations in power in ASD changed after parameterization. A significant difference was observed between neurotypical and autistic children in aperiodic total power (defined as the integral of the aperiodic spectra from 2.5–55 Hz), t (126) = 2.128, p = 0.035; Cohen’s d = 0.376, 95% CI: [0.026, 0.725]. Given the significant association between absolute gamma power and language skills in young autistic children in prior studies (e.g. [21; 29]), we also examined aperiodic gamma differences between the two groups. Neurotypical and autistic children did not show significant differences in either aperiodic or periodic gamma power, t (126) = 1.82, p = 0.071, and t (126) = 1.135, p = 0.259. In addition, there were no significant differences in the periodic total power (defined as the integral under the periodic power spectrum from 2.5–55 Hz) or periodic power in any of the canonical frequency bands (i.e., delta, theta, alpha, beta, and gamma) between neurotypical and autistic children, p > 0.05 (see Fig. 1 C; see supplementary materials). Within the autism group, autistic children with language impairment showed significant decrease in periodic alpha power within 6–12 Hz, t (61) = −2.1, p = 0.04, compared to those without language impairment. No other differences in aperiodic total power, aperiodic gamma power, periodic total power, or periodic power in any other canonical frequency band were observed (see supplementary materials).

Peak alpha frequency and alpha peak amplitude

There were no significant differences in peak alpha frequency, t (126) = 1.188, p = 0.237, or alpha peak amplitude, t (126) = 1.062, p = 0.29, between the TD and ASD groups (Fig. 1 F and 1G). However, after separating the autistic group based on language impairment status, autistic children without language impairment (ASD-NL) had significantly higher alpha peak amplitude (M = 0.44, SD = 0.12) than autistic children with language impairment (ASD-LI; M = 0.38, SD = 0.08), t (61) = 2.39, p = 0.02; Cohen’s d = 0.612, 95% CI: [0.096, 1.123] (see Fig. 2). There was no significant difference in peak alpha frequency between autistic children with and without language impairment, t (61) = −0.996, p = 0.323.

Fig. 2.

Alpha peak amplitude between autistic children with and without language impairment. Note. Alpha peak amplitude is defined as the maxima within the alpha band (6–12 Hz). Shading error describes 95% confidence interval

Given that language impairment status was associated with alpha peak amplitude, we also compared alpha peak amplitude between ASD-LI and age-matched neurotypical children. Significantly higher alpha peak amplitude was found in the neurotypical group (M = 0.45, SD = 0.12) compared to the ASD-LI group (M = 0.38, SD = 0.08), t (72) = −2.89, p = 0.005; Cohen’s d = −0.671, 95% CI: [−1.138, −0.201].

Resting-state EEG features and language abilities in ASD

Aperiodic power

We assessed whether aperiodic and periodic EEG features were associated with language abilities within the ASD group. Regarding expressive language, a significant censored regression model emerged when examining the predictive value of aperiodic offset on NDW while controlling for age, W (2) = 21.1, p < 0.001. Aperiodic offset negatively predicted the number of different words autistic children demonstrated during naturalistic interactions, b = −167.86, p = 0.021. Similarly, a significant censored regression model emerged when examining the predictive value of aperiodic offset on the number of utterances per minute, controlling for age, W (2) = 15.61, p < 0.001. Aperiodic offset negatively predicted the number of utterances autistic children demonstrated per minute, b = −7.35, p = 0.018. In terms of receptive language, aperiodic offset negatively predicted the PLS auditory comprehension standard score, b = −54.75, p = 0.017, model W (2) = 5.773, p = 0.040, after controlling for age. After controlling for multiple comparisons using the Benjamini-Hochberg procedure [55], all three censored regression models remained significant (FDR-adjusted p < 0.05). Given the strong association between language skills and nonverbal cognitive abilities in autism (e.g. [56],), we also examined the association between aperiodic offset and NVDQ controlling for age, which did not withstand multiple comparisons, p = 0.105 (Fig. 3 A).

Fig. 3.

Scatterplots of resting-state EEG and behavioral variables in autistic children. Note. A. Scatterplots of aperiodic offset, language skills, and NVDQ after controlling for age in autistic children. B. Scatterplots of alpha peak amplitude, language skills, and NVDQ after controlling for age in autistic children

Given the association between absolute gamma power and expressive language skills in previous studies with autistic toddlers and children, we also tested how aperiodic and periodic gamma power were associated with language abilities in autistic children. Initial censored regression models showed significant predictive value of aperiodic but not periodic gamma power on NDW (b = −178.88, p = 0.018) and utterance rate (b = − 7.55, p = 0.014). However, aperiodic gamma power no longer predicted either of these measures of expressive language once controlling for age in the regression models. Similarly, aperiodic gamma power did not significantly predict receptive language in autistic children, controlling for age (b = −29.22, p = 0.20).

Periodic power

Autistic children with higher alpha peak amplitude demonstrated more diverse vocabularies - NDW, b = 306.84, p = 0.004, model W (2) = 23.48, p < 0.001, more utterances per minute, b = 14.36, p = 0.0014, model W (2) = 20.01, p < 0.001, higher speech intelligibility, b = 82.59, p = 0.019, model W (2) = 20.82, p < 0.001, higher receptive language level, b = 109.89, p = 0.001, model W (2) = 10.28, p = 0.005, and higher NVDQ, rho (57) = 0.310, p = 0.017, controlling for age (Fig. 3 B). All these associations remained significant after applying the Benjamini-Hochberg FDR correction to control for multiple comparisons (FDR-adjusted p < 0.05). In contrast, PAF was not significantly associated with any variables of language skills or NVDQ. Notably, aperiodic offset and alpha peak amplitude were not significantly associated with each other in our autistic sample, r (64) = 0.009, p = 0.944, indicating two distinct potential neural mechanisms underlying language development in autistic children.

Given that aperiodic offset and alpha peak amplitude both significantly predicted NDW, utterance rate, and receptive language in autistic children while controlling for age, we further examined the relative predictive values of the two EEG features on each of these language variables by considering them together in the same censored regression model and found that both remained significant in predicting each of the language variables (see Supplementary Material).

Discussion

This study extended the investigation of resting-state EEG differences across diagnostic groups (TD vs. ASD) and ASD language phenotypes (autistic children with and without language impairment) by characterizing and comparing two physiologically distinct components from the EEG power spectrum: aperiodic components and periodic power. Using high-density whole-brain resting-state EEG data from an age-matched large sample, we found that autistic children demonstrated higher aperiodic offset, but not aperiodic slope, compared to their age-matched neurotypical peers. While we did not find significant differences in peak alpha frequency and alpha peak amplitude between neurotypical and autistic children, after separating the autistic group by language impairment status, we found that autistic children with language impairment had significantly lower alpha peak amplitude compared to autistic children without language impairment and their age-matched neurotypical peers. Autistic children with lower aperiodic offset demonstrated better concurrent expressive and receptive language skills, but not NVDQ. Autistic children with higher alpha peak amplitude demonstrated better concurrent language and nonverbal developmental skills. Given that aperiodic offset and alpha peak amplitude were not significantly associated with each other in our sample, our findings suggest that these two neural features may independently contribute to language development in autistic children.

Higher aperiodic offset but comparable slope in autistic children

Recent research highlights the importance of considering both aperiodic and periodic components when analyzing and interpreting EEG data [31], particularly given that large portions of resting power spectrum consist of nonrhythmic aperiodic signals [57]. After disentangling aperiodic and periodic components in resting-state power, we observed a higher aperiodic offset in autistic children compared to their age-matched neurotypical peers. Although increased aperiodic offset during the first year of life may be related to increases in gray matter volume, neuronal density, and synaptic number [51], aperiodic offset typically stabilizes after 1 year of age and then later decreases with age [51, 58, 59]. It is argued that reductions in aperiodic activity are associated with broader cortical maturation processes, including synaptic pruning and the refinement of neural circuits [58]. Therefore, the increased aperiodic offset observed in autistic children may reflect delayed cortical maturation and pruning relative to typical development. In addition, past work suggests that aperiodic offset reflects broadband neuronal firing [60] although note that this study was based on intracranial recordings [61]). A higher aperiodic offset observed in our autistic group, despite being chronologically age-matched, may indicate increased background neuronal firing compared to neurotypical children. Further research is needed to clarify the relation between aperiodic offset derived from extracranial recordings and broadband neuronal firing in young children.

In contrast, we did not find any significant difference in aperiodic slope between neurotypical and autistic children. Although aperiodic offset and slope are often highly correlated, a recent study of infants with and without a family history of autism also reported an aperiodic offset-specific finding, such that infants with later ASD diagnoses had greater developmental increases in aperiodic offset, but not aperiodic slope, from 3 to 12 months [62]. Both our findings suggest that aperiodic offset and slope may not necessarily co-vary in young children with ASD, highlighting the importance of examining these EEG features separately. In addition, prior studies suggested that individuals with ASD exhibit an E/I imbalance, characterized by decreased inhibition and increased excitation [34, 35]. This imbalance leads to heightened neural activity and more random broadband firing. Our finding of comparable aperiodic slope suggests that differences in E/I balance may not be salient in autistic children during the preschool years, a period of rapid neurodevelopment. In addition, although it has been suggested that aperiodic slope is linked to E/I balance [63, 64], the biological mechanism underlying this association still warrants further investigation [65]. Our finding of a similar aperiodic slope between neurotypical and autistic children does not necessarily indicate comparable E/I balance in the two groups, as we do not have pharmacological modulation in vivo to directly validate the association between aperiodic slope and E/I balance in our participants. In addition to aperiodic slope, several other proxy markers for E/I balance have been proposed, such as the kappa coefficient of neural avalanches and the detrended fluctuation analysis exponent [66]. Further research is needed to identify the most robust method for estimating E/I balance in autistic children, ideally by integrating multiple proxy markers. Finally, it should be noted that, in contrast to some prior studies that used SpecParam-estimated aperiodic offset and slope, we calculated aperiodic offset as the aperiodic power at 2.5 Hz rather than as the extrapolated y-intercept at 0 Hz, given the elevated levels of error observed in SpecParam estimates at frequencies below 2.5 Hz (see Supplementary Figure S2). In addition, when aperiodic offset is defined at the y-intercept, it is mathematically coupled with the slope, as changes in spectral steepness necessarily shift the y-intercept. By defining offset at a non-zero reference frequency, we substantially reduced collinearity between aperiodic offset and slope (see Supplementary Figure S3), facilitating a clearer interpretation of their respective associations with language variables.

Lower alpha peak amplitude in autistic children with language impairment

We also examined aperiodic-adjusted periodic power differences between neurotypical and autistic children. Given the dominant role of alpha rhythm at rest and prior mixed findings regarding absolute and relative alpha power between neurotypical and autistic children, we focused on periodic power differences in the alpha band. While we found no significant differences in alpha peak amplitude between neurotypical and autistic children, we did observe lower alpha peak amplitude in autistic children with language impairment compared to autistic children without language impairment. Given that decreased alpha power at rest has been linked to reduced cortical inhibition [13], our finding may reflect diminished neural inhibition in autistic children with language impairment compared to their autistic peers without language impairment. Importantly, our finding suggests that differences in alpha peak amplitude are more closely related to language impairment than to autism diagnosis alone, although the directionality of this association remains unclear. This finding underscores the importance of accounting for heterogeneity within the autism spectrum, as different subgroups may demonstrate distinct patterns of neural activity.

More advanced language skills are associated with reduced aperiodic offset and higher alpha peak amplitude in autistic children

We also investigated the continuous association between resting-state power and language skills in autistic children. We found that lower aperiodic offset and higher alpha peak amplitude were associated with better expressive and receptive language skills in autistic children while controlling for age. Within the limited studies that have investigated the association between resting-state power and language development in autistic children (e.g. [21, 29]), language skills were solely measured by broad standardized language assessments, which reflected language levels relative to typical development. In addition to using a standardized language measure to assess receptive language skills, we used natural language samples to capture more fine-grained details and multiple dimensions of expressive language profiles in autistic children [67], including lexical diversity characterized by the number of different words, communication fluency through the number of utterances per minute, and speech intelligibility measured by the percentage of intelligible utterances. Using this approach, we found that aperiodic offset was significantly associated with lexical diversity and communication fluency in autistic children during naturalistic interactions and their receptive language skills measured by a standardized measure. Similar associations were found with alpha peak amplitude, which was also related to speech intelligibility and nonverbal developmental skills.

Lower aperiodic offset indicates reduced broadband random neuronal firing in the brain [61] and may reflect alteration in cortical maturation, including greater synaptic pruning or the refinement of neural circuits [58]. Increased inhibition, especially in frontal brain circuits, is thought to reduce neural noise and potentially enables more efficient processing of complex inputs, such as language, during early childhood [68], thereby potentially supporting the development of language. In addition, maturational increases in alpha peak amplitude have been associated with increased thalamocortical connectivity [69]. Thus, observed associations between alpha peak amplitude and language development may be related to the maturation of thalamocortical circuits that is essential for sensory processing and higher-order cognitive processes such as learning and attention [70]. Alpha oscillations also play a critical role in the timing of cortical processes [71] and provide an infrastructure for neural communication within and between brain regions [13, 72], all of which are essential for language processing and learning. Therefore, it may be that more efficient processing of language input and neural communication in the brain contributes to better language development in autistic children during early childhood.

Taken together, our findings highlight that aperiodic components, particularly aperiodic offset, are the characteristic features of resting-state power differences between preschool-aged neurotypical and autistic children. In contrast, alpha peak amplitude shows distinct patterns between autistic children with and without language impairment, despite being similar between neurotypical and autistic children. In addition, we found that lower aperiodic offset (indicating reduced broadband random neuronal firing in the brain) and higher alpha peak amplitude (potentially reflecting greater thalamocortical connectivity) both independently contribute to more advanced language skills in autistic children. These findings provide unique insights into the potential neural mechanisms underlying different language developmental patterns in autistic children.

Limitations

This study has several limitations that offer directions for future research. The cross-sectional nature of our data and analysis does not allow us to disentangle the developmental sequence of language abilities and resting-state EEG components. Future longitudinal studies are needed to explore whether early variations in resting-state EEG periodic and aperiodic features are predictive of later language development or vice versa. Moreover, the majority of participants in our study were from relatively high socioeconomic backgrounds, reflected in annual household incomes and parental education levels. Therefore, our findings may not generalize to autistic children from more socioeconomically diverse backgrounds. Many of our participants were also nonverbal. Although using standard scores allows for comparing one’s receptive language level to the chronological age expectations, it may underestimate the variability of those whose language skills are far below what is expected at their chronological age. Future research should consider alternative ways to capture the full range of their communication skills. In addition, the ASD group had fewer females, a more diverse ethnic and racial background, relatively lower household income, parents with slightly lower educational attainment, and relatively poorer data quality. Due to the sample size in the current study, it was not feasible to statistically control for all these covariates in the analyses. Therefore, interpretation of the findings should be cautious, given that these demographic factors may confound the association between resting-state EEG features and language skills in autistic children, and variability in data quality may have influenced the observed differences in aperiodic components. Last, our study focused on preschool-aged neurotypical and autistic children between 2.5 and 6 years old. Given that the preschool age is a rapid developmental period for both language skills and brain maturation, our relatively broad age range may have introduced variability that could influence the interpretation of neural and behavioral associations. It is also important to note that children experiencing the resting state differently may lead to potential variations in the data [73]. While some children may be able to remain still and relax their top-down cognitive control to engage in a resting state, others may find it difficult. Watching a predetermined video of a screensaver of abstract moving objects rather than a preferred video may have affected some children’s engagement. Therefore, observed variations in resting-state EEG may be influenced by individual differences in internal state or arousal levels rather than being exclusively driven by the underlying biological factors.

Conclusions

Despite these limitations, this is the first study to characterize and compare aperiodic and periodic activity of resting-state EEG across both diagnostic groups (neurotypical vs. autistic) and different ASD language phenotypes (autistic with and without language impairment). Notably, resting-state EEG power exhibits distinct differences when comparing between autistic children to neurotypical peers versus comparing within autistic subgroup based on language impairment. These findings underscore the importance and necessity of considering the diverse phenotypic profiles of autistic participants in future studies.

Supplementary Material

Below is the link to the electronic supplementary material.

Abbreviations

ADOS-2

Autism Diagnostic Observation Schedule, Second Edition

ASD

Autism Spectrum Disorder

ASD-LI

Autism Spectrum Disorder with Language Impairment

ASD-NL

Autism Spectrum Disorder with Language Skills in the Normal Range

BEAPP

Batch Electroencephalography Automated Processing Platform

BRIDGE

Brain Indicators of Development Growth Study

DAS

Differential Ability Scales

EEG

Electroencephalography

E/I balance

Excitation/Inhibition balance

ELSA

Eliciting Language Samples for Analysis

HAPPE

Harvard Automated Preprocessing Pipeline for EEG

MSEL

Mullen Scales of Early Learning

NDW

Number of Different Words

NVDQ

Nonverbal Developmental Quotient

PAF

Peak Alpha Frequency

PLS

Preschool Language Scales

POLO

Predicting and Optimizing Language Outcomes Study

PSD

Power Spectral Density

TD

Typically Developing

Authors contributions

Yanru Chen: conceptualization, project administration, methodology, visualization, formal analysis, data curation, writing – original draft, writing – review and editing. Meagan Tsou: project administration, data curation, writing – review and editing. Charles A. Nelson: funding acquisition, supervision, resources, writing – review and editing. Helen Tager-Flusberg: funding acquisition, supervision, resources, writing – review and editing. Carol L. Wilkinson: conceptualization, funding acquisition, supervision, methodology, resources, writing – review and editing. All authors read and approved the final manuscript.

Funding

This work was supported by the National Institute on Deafness and Other Communication Disorders (NIDCD) (P50 DC018006, MPIs: Tager-Flusberg/Kasari; K23 DC07983, PI: Wilkinson) and the Harvard Catalyst K12/CMeRIT Award (PI: Wilkinson).

Data availability

The datasets generated and analyzed during the current study are not publicly available due to privacy or ethical restrictions but are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

All study procedures were conducted according to the guidelines of the Declaration of Helsinki, including informed consent and assent prior to study inclusion, and approved by the Institutional Review Board at Boston University (IRB protocol #5372E) and Boston Children’s Hospital (IRB-P00034676).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.