Physics Department, Laboratory of Integrative Neuroscience, FCEyN UBA and IFIBA, Conicet, Pabellón 1, Ciudad Universitaria, 1428 Buenos Aires, Argentina

Cognitive Neuroimaging Unit, Institut National de la Santé et de la Recherche Médicale (INSERM), 91191 Gif sur Yvette, France

Institute of Cognitive Neurology (INECO), Favaloro University, Buenos Aires, Argentina

Programa Argentino para Niños, Adolescentes y Adultos con Condiciones del Espectro Autista (PANAACEA), Buenos Aires, Argentina

UDP-INECO Foundation Core on Neuroscience (UIFCoN), Diego Portales University, Santiago, Chile

Universidad Torcuato Di Tella, Almirante Juan Saenz Valiente 1010, Buenos Aires C1428BIJ, Argentina

Abstract

Background

The dimensional approach to autism spectrum disorder (ASD) considers ASD as the extreme of a dimension traversing through the entire population. We explored the potential utility of electroencephalography (EEG) functional connectivity as a biomarker. We hypothesized that individual differences in autistic traits of typical subjects would involve a long-range connectivity diminution within the delta band.

Methods

Resting-state EEG functional connectivity was measured for 74 neurotypical subjects. All participants also provided a questionnaire (Social Responsiveness Scale, SRS) that was completed by an informant who knows the participant in social settings. We conducted multivariate regression between the SRS score and functional connectivity in all EEG frequency bands. We explored modulations of network graph metrics characterizing the optimality of a network using the SRS score.

Results

Our results show a decay in functional connectivity mainly within the delta and theta bands (the lower part of the EEG spectrum) associated with an increasing number of autistic traits. When inspecting the impact of autistic traits on the global organization of the functional network, we found that the optimal properties of the network are inversely related to the number of autistic traits, suggesting that the autistic dimension, throughout the entire population, modulates the efficiency of functional brain networks.

Conclusions

EEG functional connectivity at low frequencies and its associated network properties may be associated with some autistic traits in the general population.

Background

In the seminal paper in which Kanner first characterized autism, he noticed that some of the symptoms observed in children were shared at sub-threshold levels by their parents

There is increasing evidence that ASD could be a condition of altered brain connectivity

To assist in the diagnosis of ASD, it is essential to find robust brain biomarkers that characterize ASD as the upper extreme of a dimension across the entire population. To date only one piece of research has addressed this issue: it showed that fMRI connectivity in a single link connecting the anterior cingulate cortex and the mid-insula diminishes in strength as the number of autistic traits increases

In a previous study we showed a diminution of long-range connectivity within the (low-frequency) delta band in an ASD population compared to control groups, leading to a ‘big-world’ organization of brain connectivity in ASD

Methods

Participants and assessment

In this study, there were 74 subjects of similar educational and cultural backgrounds (37 male, 37 female; mean age = 27.33, SD = 5.10; educational level = 19.2 years; SD = 2.94 years). None of the volunteers had a history of neurological or psychiatric conditions as determined by a semi-structured interview (Schedules of Clinical Assessment in Neuropsychiatry)

We requested participants to select someone who knew them well, preferably a close relative or partner, to complete the adult version of the SRS questionnaire (SRS-A, ^{-9}). Also, we observed a significant positive correlation between ADOS and SRS scores within the ASD group (correlation = 0.67,

EEG results

EEG measurements were taken in a Faraday cage with a Biosemi Active Two 128-channel 24-bit resolution system, with active electrodes (the first amplifying stage on the electrode improves the signal-to-noise ratio), digitalized at 512 Hz and low-passed DC-1/5th of the sample rate (-3 dB) by a fifth-order digital sync anti-aliasing filter. There were no additional hardware filters during acquisition. Temporal signals between 5 and 10 minutes were recorded during an eyes-closed rest while subjects sat on a reclining chair in a sound-attenuated room with a dim light. During the experiment, participants and EEG recordings were monitored to ensure that they maintained vigilance and did not fall asleep. After the acquisition, signals were re-referenced to the average of all electrodes. Segments containing movement artefacts were manually deleted (mean length of the remaining time series = 7.90 min, SD = 2.21), followed by an ICA-based rejection of residual artefact-laden ICA-components. After this pre-processing, we filtered the EEG signals on specific frequency bands: delta (0.5 Hz to 4 Hz), theta (4 Hz to 8 Hz), alpha (8 Hz to 12 Hz), sigma (12 Hz to 15 Hz), beta (15 Hz to 25 Hz) and gamma (25 Hz to 35 Hz).

Data analysis

MATLAB (MathWorks Inc, Natick, MA) was used for the analyses. We first explored the relation between SRS score and gender. We calculated the mean SRS score for men and women, and assessed their difference statistically using a t-test for two independent samples. To study the relation between SRS score and age, we conducted both a t-test comparing groups (low SRS score and high SRS score groups, obtained after a median split on the SRS score) and a regression between SRS score and age (Figure

Autistic traits and functional EEG connectivity

**Autistic traits and functional EEG connectivity.** (**a**) SRS scores. The SRS score is higher for men than women. Error bars show the SEM. (**b**) Scatter plot showing that age does not modulate SRS within our population. (**c**) Cumulative β-value distributions, showing that the entire distributions are shifted to negative values. The dotted red line is a cumulative value of 0.95. (**d**) Mean and SEM of β-value distributions, obtained from independent multivariate regression between SRS score and SL connectivity (including gender and age as regressors of no interest). See Table **e**) Scalp plot of β-values. A link is traced between two electrodes if that connection is significant (

To test whether synchronisation likelihood (SL) connectivity in spontaneous activity between electrodes covaries with SRS score, we conducted a functional connectivity analysis. The synchronisation between all pair-wise combinations of EEG channels was computed for all subjects with the SL method _{
f,p
}. A matrix entry _{
f,p
}
_{
f,p }matrices. To investigate connectivity changes associated with SRS score, we conducted an across-subjects multivariate linear regression, using least squares (as implemented in MATLAB function _{
f,p }and the SRS score for each subject, including gender and age of the subject as regressors of no interest. This lead to six matrices of beta (β) values, _{
f
}
_{
Delta
}
_{
f
}

To assess the _{
f }matrices statistically we performed a bootstrap analysis _{
f }matrix, and called it the observed mean _{
f }per frequency band

To further characterize the β-value distributions, we calculated the standard error of the mean (SEM) for each distribution _{
f }through a jackknife procedure _{
N }is the standard deviation of the β-values over the

To estimate the discriminative power of SL at characterizing autistic traits, we calculated a receiver operating characteristic (ROC) curve _{ROC}) quantifies how separable the two groups are: _{ROC} = 0.50 means that the two groups completely overlap (along the variable considered), while _{ROC} = 1 indicates that the two groups are perfectly separable by their respective SRS scores.

To address the issue of length of connections and their relation with SRS score, we defined four different regions grouping electrodes: frontal, occipital, lateral right and lateral left (Figure

SL connectivity between and within ROIs

**SL connectivity between and within ROIs.** (**a**) ROC curve value for all frequency bands, between and within frontal and occipital ROIs. Fronto-occipital SL connectivity produced ROC curves showing significant _{ROC} values. (**b**) ROC curve value for all frequency bands, between and within right and left ROIs. Left-right SL connectivity produced ROC curves showing significant _{ROC} values. (**c**) ROC curves after performing a median-split group separation, classifying subjects as high or low SRS score according to their gamma band SL value. Only the fronto-occipital SL shows significant classification (A_{ROC fronto-occipital} = 0.86; A_{ROC occipital} = 0.51; A_{ROC frontal} = 0.59). (**d**) Multivariate regressions obtained by averaging all gamma band SL values within (frontal and occipital) and between (fronto-occipital) ROIs, against SRS score (and gender and age as other regressors). Frontal and occipital SLs do not show significant regression with SRS score (occipital: P = 0.23, R-square = 0.01; frontal: P = 0.31, R-square = 0.001). The fronto-occipital SL, in contrast, does show good regression between SL and SRS score (P < 0.05, R-square = 0.09).

Graph theory metrics

We used graph theory metrics to summarize topological information. The connectivity matrix _{
f,p
} defines a weighted graph where each electrode corresponds to a node and the weight of each link is determined by the SL of the electrode pair. To calculate network measures, _{
f,p }matrices were converted to binary undirected matrices by applying a threshold _{
f,p }matrix to a binary undirected graph, we calculated the clustering coefficient

To quantify the impact of autistic traits on network properties, we performed two analyses. First we conducted a median-split analysis, grouping subjects as for the SL quantification analysis (conducting a t-test and a ROC analysis). Also we conducted a multivariate linear regression analysis between the small-world index and SRS score (including gender and age as regressors of no interest). To assess the regression analysis statistically, we conducted a bootstrap analysis repeating the methods used to assess the significance of the _{
f }matrices. We obtained the P-value corresponding to the observed small-world index, and set the

Results

First we measured the dependence of SRS score on the demographical covariates age and gender. As expected

Next we measured the relation between SRS score and connectivity. For each participant in this study, we calculated the synchronization likelihood across all pairs of channels. The element _{
ij
} and the values for total SRS, with gender and age as regressors of no interest. A positive β-value β_{ij} indicates that the SL between electrodes

The regression analysis showed an overall decrease of the mean connectivity (averaged across all electrode pairs) as the SRS score increased. This global decrease in connectivity was observed for all frequency bands, except in the alpha band and was significant for the delta, theta, beta and sigma bands (see Table

**Frequency band**

**Distribution mean and SD**
^{
a
}

**Significance (bootstrap P)**

^{a}All distributions are shifted towards negative values.

Delta

-0.030 ± 0.013

0.01

Theta

-0.017 ± 0.014

0.01

Alpha

-0.005 ± 0.011

0.53

Sigma

-0.007 ± 0.0072

0.10

Beta

-0.008 ± 0.006

0.04

Gamma

-0.007 ± 0.005

0.01

Next, we investigated the hypothesis that long-range connections are more informative about individual autistic traits than short connections. We averaged SL values for all pairs of electrodes within and between frontal, occipital and lateral clusters of electrodes (Figure _{ROC} are listed in Additional file

A_{ROC} for all frequency bands, for the combination of ROIs frontal, occipital and fronto-occipital.

Click here for file

The SL between electrode pairs between the left and right clusters also classified subjects better than the SL obtained from electrode pairs within left or right clusters (Figure _{ROC} values are listed in Additional file

A_{ROC} for all frequency bands, for the combination of ROIs right, left and right-left.

Click here for file

SL connectivity between long and short distances. (a) ROC curve value for all frequency bands, between and within frontal and occipital ROIs. (b) Long-distance SL connectivity produced ROC curves showing significant

Click here for file

Finally we examined the hypothesis that changes in connectivity result in a different network topology for participants with high or low autistic traits. Specifically we hypothesized that the small worldness of the network, a property which weights the compactness and clustering of a network using optimal architectures for information storage and propagation _{ROC} = 0.75; Figure

Topology of networks obtained from delta band _{f,p }matrices

**Topology of networks obtained from delta band **_{f,p }**matrices.** (**a**) Small-world indices, after a median-split of subjects into high and low SRS groups. Low SRS subjects have a higher small-world index (low SRS group = 0.38, high SRS group = 0.22; T-value = 4.17; **b**) ROC curve for the discrimination between low and high SRS score based on the small-world index. A_{ROC} is highly significant (A_{ROC} = 0.74, **c**) Regression between small-world index and SRS score (and age and gender), showing the negative relation between small-world index and SRS score (β-value = -0.08;

Significance and size effects of the t-test between low and high SRS groups for all frequency bands.

Click here for file

Significance and size effects of the regression analysis between the small-world index and SRS score for all frequency bands.

Click here for file

Discussion

The main purpose of this study was to characterize and compare resting-state functional brain networks in typical subjects along the autistic dimension. In agreement with our three working hypothesis we observed that: (1) the SRS score in the general population is indexed by the strength of connections between EEG electrodes – long-range connections are more predictive than short-range connections of an individual’s SRS score, (2) the lower part of the EEG spectrum is the most informative for individual autistic traits and (3) the small worldness of the network (and hence its optimality for storage and transfer of information) increases as the SRS score diminishes.

Long-range intra-cortical and feedback cortico-cortical connections, which are thought to be altered in ASD, are revealed by the slow cortical potentials of the EEG

Changes in connectivity patterns have an impact on the global organization of a network, which in turn determines the efficiency of information transfer and storage

In order to become useful tools and assist ASD diagnosis, brain-imaging techniques must find robust biomarkers to characterize ASD and its relation with sub-threshold traits in the general population. Only one piece of research has addressed this issue; Di Martino et al.

One limitation of the present study is the lack of a measure of specificity regarding possible comorbidities in the autistic traits we measured. The SRS score might capture, along with autistic traits, traits characterizing other psychiatric conditions

Conclusions

The present study demonstrates that a resting-state EEG can identify robust and monotonic changes associated with SRS score, a graded measure of autistic traits in the general population. Our results show a decrease in functional connectivity, mainly for the delta and theta bands, is associated with an increased number of autistic traits. When inspecting the impact of autistic traits on the global organization of the functional network we found that the optimal properties of the network are inversely related to the number of autistic traits, suggesting that the autistic dimension, throughout the entire population, modulates the efficiency of functional brain networks.

Abbreviations

ADOS: Autism Diagnostic Observation Schedule; ASD: autism spectrum disorder; EEG: electroencephalography; fMRI: functional magnetic resonance imaging; ROC: receiver operating characteristic; SEM: standard error of the mean; SL: synchronisation likelihood; SRS: Social Responsiveness Scale; ICA: independent component analysis; ROI: region of interest.

Competing interests

All authors report no financial relationships with commercial interests.

Authors’ contributions

PB and MS conceived the experiment. PB, SC, LA, JA, LB and AT collected the data. PB and MS analysed the data. PB, SC, MS, AI and FM wrote the paper. All authors read and approved the final manuscript.

Acknowledgments

This work is funded by CONICET and UBACYT. MS is sponsored by the James McDonnell Foundation 21st Century Science Initiative in Understanding Human Cognition - Scholar Award. MS and PB are supported by the Human Frontiers Science Program and a post-doctoral grant (PB). AI is supported by CONICET, CONICYT/FONDECYT Regular (1130920), INECO Foundation grant, and grants FONDECYT (1130920) and PICT 2012–1309. FM is supported by a grant, PICT 2012–0412. LB is supported by a post-doctoral grant from CONICET, Argentina.