EEG functional connectivity in infants at elevated familial risk for autism spectrum disorder

Background: Many studies have reported that autism spectrum disorder (ASD) is associated with atypical structural and functional connectivity. However, relatively little is known about the development of these differences in infancy and on how trajectories may vary between sexes. Methods: We used the International Infant EEG Platform (EEG-IP), a high-density electroencephalogram (EEG) dataset pooled from two independent infant sibling cohorts, to characterize such neurodevelopmental deviations during the first years of life. EEG was recorded at 6, 12, and 18 months of age at typical (N=97) or high familial risk for ASD (N=98), determined by the presence of an older sibling with a confirmed ASD diagnosis. We computed the functional connectivity between cortical EEG sources during video watching using the corrected imaginary part of phase-locking values. Results: Our findings showed low regional specificity for group differences in functional connectivity but revealed different sex-specific trajectories between females and males in the group of high-risk infants. Specifically, functional connectivity was negatively correlated with ADOS calibrated severity scores, particularly at 12 months for the social affect score for females and for the restrictive and repetitive behaviors for males. Limitations: This study has been limited mostly due to issues related to the relatively small effective sample size inherent in sibling studies, particularly for diagnostic group comparisons. Conclusions: These results are consistent with sex differences in ASD observed in previous research and provide further insights into the role of functional connectivity in these differences.

findings, a general pattern of long-range EEG and magnetoencephalograpy (MEG) 83 underconnectivity distinguishes autistic from neurotypical individuals (17). These participant 84 samples have often been heavily skewed toward males, resulting in a gap in understanding the 85 development of functional networks in autism in males relative to females. The few studies that 86 have explored the effect of biological sex on functional connectivity as it relates to autism have 87 reported that autistic females show increased connectivity compared to both autistic males (18) 88 and neurotypical females (19). Further, one study reported reduced functional connectivity 89 associated with ASD symptoms specifically in females (16). 90 Once expanded and replicable in independent samples, these findings would lend support to 91 the "female protective effect" (FPE) which states that compared to males, more severe 92 etiological factors are necessary for autistic expression in females (20-22). FPF has been 93 proposed to explain the relatively higher prevalence of autism in males, a discrepancy that 94 appears to remain even after accounting for known ascertainment biases (23). 95 An effective way to examine the FPE is to study very early development in autism, before risk 96 signs become compounded and amplified by atypical interactions within the brain and with the 97 external environment. Despite the rapid increase in studies on early trajectories of brain 98 development in infants who later develop autism (14), little is known about early functional 99 connectivity, particularly as it relates to autism and biological sex. Only a few ASD studies have 100 estimated connectivity in infancy (24-28). While the findings have not been conclusive, current 101 evidence suggests that cortical network maturation differs in autistic individuals, with initial 102 overconnectivity within the first year of life followed by underconnectivty beginning in 103 toddlerhood (17). To date, very few infant-sibling studies have included biological sex as a 104 variable of interest when studying functional connectivity in ASD, with no significant sex-related 105 results (24,25). This situation might be due in part to the study of biological sex in sibling studies 106 being complicated by the relatively low number of children who go on to develop autism and the 107 high sex imbalance (mostly male) in the diagnosed subsample. 108 In the current study, we focus on the properties of the functional connectome in early 109 development to better understand variations between infants with an elevated likelihood for reporting results for categorical outcomes, we also perform dimensional analyses using Autism 117 Diagnostic Observation Schedule (ADOS) severity scores to addressed previous critics 118 regarding the use of categorical diagnostic outcomes. Previous studies have found categorical 119 analyses to mask significant heterogeneity in the nature and severity of symptoms for children 120 who develop autism (31). This issue also extends to children who do not receive an ASD 121 diagnosis but experience problems in other developmental domains such as language and 122 attention (32). Moreover, dimensional analysis allow us integrate developmental trajectories in 123 children at risk for ASD who do or do not experience challenges and gain further insights into 124 resilience processes (14). 125

Methods 126
Sample 127 The sample used for this study was taken from the EEG Integrated Platform (EEG-IP; van  Oregon) Net Station software. Scalp EEG was recorded at 500 Hz using a vertex reference and 152 re-referenced offline using a robustly interpolated average. The London and the Seattle 153 datasets were notch-filtered at 50 Hz and 60 Hz, respectively, to remove power line 154 contamination. 155 Semi-automated pre-processing was done with the EEG-IP-Lossless pipeline (29) using an 156 Octave interpreter running on a Compute Canada cluster. Pre-processing involved 157 comprehensive data annotation to identify artifacts and non-stationarity in scalp channels and 158 independent components. This pipeline provided an initial automated classification of the 159 independent components as being either valid brain activity or capturing some artifacts, such as 160 electromyographic, electrocardiographic, electrooculographic, and power line contamination. 161 Quality control included an expert review of all data annotations and confirmation of artifacts 162 informed by initial classification, topographies, activation time-series, dipole fit, and power 163 spectrum. For an expanded description of pre-processing criteria and artifact thresholds, see 164 (29,30). Once the flagged artifacts were removed from the data, the reconstituted scalp EEG 165 was epoched into 1-second non-overlapping windows for source reconstruction and calculation 166 of functional connectivity. We used relatively short time windows, as they have shown to be 167 advantageous for estimating functional connectivity (34). 168

Source reconstruction 169
Most EEG connectivity studies in autism have been performed on scalp electrode signals (17), 170 which are known to have various limitations compared to source analyses, such as poorer 171 signal-to-noise ratio, the impossibility to relate observations to brain structures, and the 172 confounding effect of volume conduction, reference electrodes, and common sources (35-39). 173 Although tools for EEG source reconstruction are now widely available, they have been used 174 only in a few autism studies (40,41). For infants with ASD or at high risk of ASD, the lack of age-175 matched templates has resulted in the use of head templates built from an adult population, 176 such as the Montreal Neurological Institute (MNI) brain (41), which is likely to distort source 177 estimations in ways that are not well-established. For this study, we used a recently developed 178 set of infant structural head templates (42) to perform EEG cortical source reconstruction, 179 investigate functional connectivity in infants, and identify potential ASD risk and resilience 180 factors. To avoid confounding a potential effect of the head template with the recording time 181 points, we used the 12-month template for all recordings. As a validation, the same analyses 182 were performed with age-matched templates and resulted in qualitatively similar conclusions. 183 The cortex for these templates has been parcellated using the Desikan-Killiany (43) scheme. 184 Sources were estimated using MNE-Python 0.23 (44), with the eLORETA inverse operator (45), 185 λ 2 =10 -4 , and with dipoles aligned perpendicular to the cortical mesh. Sources were averaged for 186 every brain region, using the "mean flip" mode from MNE-Python. 187 We initially computed functional connectivity for the broadband signals (5-100 Hz), as well as for 197 a few typical frequency bands: theta (5-8 Hz), alpha (8-12 Hz), beta (12-30 Hz), gamma (30-100 198 Hz). Given that our preliminary analyses did not indicate any noteworthy impact of frequency on 199 between-group differences in connectivity, we report only the broadband analyses. 200

Resting-state networks 201
To compare the functional connectivity estimated in this naturalistic video watching task within 202 the different resting-state networks, we labeled the brain regions as being part of the auditory, 203 default mode, dorsal attention, salience, or visual networks, or none of the above, following a 204 previously published classification (48). The functional connectivity for each of these networks 205 was computed as an average of the all-to-all connections between the regions that are part of 206 the corresponding networks. 207

Statistical analysis 208
Statistical analyses were run using pandas 1.1.4 for data manipulation, statsmodels 0.12.2 for 209 linear regressions, and seaborn 0.11.0 and matplotlib 3.4.0 for visualization. To improve the 210 normality of the connectivity measures, we transformed them using a logit equation (1). 211 This transformation changes the support of the connectivity measures from [0, 1] to [-inf, +inf] 213 and helps diminish the asymmetry of the distribution, particularly the heavy right tail we 214 observed in our empirical distributions. We further rejected the EEG recordings in which the 215 functional connectivity was considered a statistical outlier, defined as being either more than 1.5 216 inter-quartile intervals above the third quartile or below the first quartile (see supplementary 217 information for details). Table 2 lists the number of available recordings after artifact rejection. 218 We used mixed-effect multifactorial linear regressions to test the impact of biological sex, 223 diagnostic groups, site, and age on functional connectivity, using the subject as grouping 224 random-effect factor and the following model structure for the fixed effects: 225 logit(CON) ~ sex*group + site*age (2) 226 In a second time, we looked at correlations between ADOS calibrated severity scores and 227 functional connectivity, within the ELA group using: 228 Regression (3) using ADOS includes only the ELA subjects, whereas the regression using the 230 diagnostic group (2) used recordings from both groups. 231

Results 232
Below, we proceed with a detailed investigation of the connectivity in the EEG-IP dataset, 233 exploring categorical diagnostic group effects and the effect of ADOS scores as a dimension. 234 We also investigated whether these relationships were affected by multiple factors, including 235 age, biological sex, site, functional networks, and distance between communicating regions. 236 Overall, as described in detail below, we observed a tendency for underconnectivity associated 237 with familial risk and later ASD diagnosis, and we identified systematic differences between 238 male and female ELA infants. These differences might reflect the higher resilience of females to 239 ASD symptoms and support that a more pronounced alteration of the functional connectome is 240 necessary for the symptomatic expression of ASD in females. 241

Linear regressions 242
Both mixed-effect regression models described in (2,3) show a significant (2: p=2.4e-5; 3: 243 p=1.4e-3) and negative effect of age on functional connectivity. The regressions using the 244 diagnostic groups also revealed an interaction between the site and the age (p=0.03). All other 245 factors were not significant (p>0.1). Model (2) was run with 350 observations from 179 246 participants, whereas model (3) was run with 131 observations from 71 participants (only 247 including ELA infants with ADOS calibrated severity scores). In the following sections, we 248 further explore these results to look for more specific effects (e.g., specific to age, network, etc.) 249 that would not have been captured by these regression models. 250

Age effect 251
Our data seem to show a decrease in overall connectivity with age, with potentially a U shape 252 reaching a minimum value at 12 months. No other factor reliably modulated this effect. In this study, we used the EEG-IP database to examine whether functional connectivity between 319 EEG cortical sources during the first year of life is atypical in ELA infants later diagnosed with 320 ASD. It constitutes one of only a handful of studies on EEG functional connectivity in infants with 321 ASD (24,25,27,28,50,51). Furthermore, it uses methods that improve upon past studies, such 322 as using a robust functional connectivity metric and computing connectivity over cortical sources 323 using age-matched head templates. 324

Underconnectivity in infants with ASD 325
Our observations provide some insights into the developmental origins of underconnectivity in 326 children and adults with ASD. They do not clearly support an underconnectivity hypothesis in 327 very early childhood during the pre-diagnostic period, and correlations between ADOS and 328 functional connectivity, when significant, tended to be of negative sign (i.e., larger ADOS were 329 associated with less functional connectivity; see Figure 5 and Supplementary Figures 3 and 5). 330 Nevertheless, our results are inconclusive with respect to a relationship between EEG functional 331 connectivity and autism in infants. Any potential effect is likely to be of relatively small sizes or is 332 difficult to capture through population averages due to sample heterogeneity. Such inconclusive 333 results are compatible with recent reports from the large sample study LEAP (52). 334

335
Beyond generalized group differences in the all-to-all connectome, we also looked for different 336 subsets of connections (e.g., belonging to specific resting-state networks) to investigate the 337 possibility of a more specific neural effect. Our investigation failed to reveal a systematic, 338 reliable, and reproducible pattern across our two data subsets. This situation may be due to a 339 few factors. First, averages may not contrast groups if the effect of ASD on functional 340 underconnectivity is inconsistent across subjects (e.g., an heterogeneous mixture of over and 341 underconnectivity may end up to show a normal level of connectivity at the group level). Further, 342 infant EEG data is inherently noisy, and the experimenters have little control over the infant's 343 behavior due to incontrollable factors such as tiredness and fussiness of the infants, which is 344 likely to cause variable brain activations within and between subjects during EEG acquisitions. 345 Lastly, we noted what looks like a significant degree of source leakage. Similar to volume 346 conduction between scalp channels, source leakage generates zero-lag correlations between 347 brain regions. It is hard to know what portion of unlagged synchrony is due to genuine zero-lag 348 connectivity known to exist even between distant brain regions (53,54) and what part is due to 349 source leakage and unresolved challenges associated with the under-determined estimation of 350 cortical sources from scalp signals. Regardless of its cause, such source leakage blurs regional 351 specificity by increasing the apparent similarity of brain activity across regions. Notwithstanding 352 these potential issues and the possibility that we missed some effects, these analyses were 353 thorough and extensive, and it is likely that any group effect, if present, would be of small size. 354

Effect of biological sex on the relationship between ASD and functional connectivity 355
In our analyses, we observed different connectivity profiles between elevated-risk females and 356 males, with females at 12 months showing a negative association between functional 357 connectivity and social affect, as measured by ADOS calibrated severity scores. These 358 conclusions are coherent with previous findings that females diagnosed with ASD show a larger 359 deviation from neurotypical functional connectivity (55) and suggest that a larger difference 360 compared to neurotypical functional connectivity is required for females to start showing ASD 361 social symptoms. These differences could reflect early protective brain mechanisms that afford 362 resiliency against ASD social symptoms in spite of altered connectivity in elevated-risk females, 363 which may contribute to the lower prevalence of ASD in that population. The interplay of 364 affected dimensions (RRBs versus social affect) and biological sex may also explain disparities 365 in diagnosis prevalence. 366 Given the preponderance of autism in males, biological sex has emerged as a potential 367 protective mechanism that mitigates risk (56). With respect to social affect symptoms, females 368 appear to be more resilient, requiring higher genetic loading to reach ASD diagnostic threshold 369 (57). Greater social cognitive abilities in females might contribute to such resilience and may be 370 reflected in anatomical brain differences such as a comparatively thinner cortical sheet in 371 several brain regions in females (58). A corresponding increased loading of functional 372 connectivity required for ASD diagnosis was found in resting-state fMRI (55), with increased 373 local connectivity in somatomotor and limbic networks and decreased local connectivity in 374 default mode networks. Females with ASD have also been found to show increased connectivity 375 compared to males (18) as well as typically developing females (19). Compared to males, 376 higher functional connectivity in females has been observed for whole-brain connectivity and in 377 functionally distinct networks, including the default mode and the central executive networks 378 (15,16). In their study, Olson et al. (2020) found that reduced functional connectivity was 379 associated with early ASD symptoms, specifically in females. Taken together, the functional 380 connectivity profile in ASD suggests that higher connectivity in females is linked to lower 381 severity of ASD symptoms, although there might be some regional specificity, at least in resting-382 state fMRI (55). Our findings in infants could reflect early neurodevelopmental markers of these 383 divergent biological sex trajectories that perhaps contribute to reduced prevalence in females 384 with higher functional connectivity. Such "markers" should be understood and studied for the 385 insight that they can provide about underlying mechanisms rather than as potential clinically 386 relevant biomarkers since the effect size of these relationships is likely to be relatively small, not 387 making them plausible candidates for classification on their own 1 . 388 Lastly, we note that in this paper we follow the World Health Organization definitions for sex and 389 gender, thus when discussing biological sex differences, we refer to differences currently 390 thought to be influenced by biological and genetic properties. Still, we acknowledge that it is 391 hard to completely disentangle the effects of biological sex and gender socialization in human 392 development, particularly considering that gender socialization begins at birth and may influence 393 neurobiology (59-61). We would also like to acknowledge that autistic individuals may be less 394 likely to identify with their sex assignment from birth compared to neurotypical individuals 395 (62,63), and while it is not possible to assess gender identity in infancy, we nonetheless 396 1 Although these relationships do not have an effect size sufficiently strong to consider them as clinically relevant biomarkers, we cannot dismiss the possibility that they might eventually be useful in combination with other markers to better subtype ASD. However, much work would be needed before a robust classification from a constellation of weak markers significantly improves diagnostic practices currently based on simpler observational measures. encourage future autism studies to consider both gender and biological sex factors when 397 possible. 398

Limitations 399
This study has been limited mostly due issues related to small sample size inherent in sibling 400 studies. In a previous systematic review, we have shown that studies on the impact of ASD on 401 EEG and MEG functional connectivity often report contradictory results, probably due to many 402 confounding factors across studies (e.g., differences in inclusion/exclusion criteria, in 403 connectivity metrics, in frequency bands, in participant demographic characteristics) and 404 methodological difficulties in estimating reliably the cerebral sources of EEG/MEG activity and 405 the functional connectivity between them (17). This review also showed that small sample sizes 406 are often used in studies of functional connectivity in autism. Histograms of sample sizes used 407 in these studies show that samples of 10 ASD subjects or less are not uncommon (24%), and 408 most studies (74%) have ASD groups of no more than 25 subjects (see Supplementary Figure  409 6). 410 Our study compares relatively well, with an average size for the ASD group of 22 participants 2 411 across time points (sites combined). This is particularly true considering the prospective nature 412 of this study, i.e., although a comparatively large number of infants enter the study at young 413 ages, only a fraction of them are later diagnosed with ASD. For that reason, the non-autistic 414 groups are significantly larger than the sample of participants with ASD. Actually, our dataset 415 constitutes the largest infant sample and the second-largest sample overall among the 416 functional connectivity studies included in our previous review. 417 Nevertheless, our analyses have been limited by relatively small sample sizes. The first reason 418 for that is that infant EEG recordings are comparatively noisier than adult EEG recordings 419 because it is harder to control sources of physiological artifacts (e.g., EMG and EOG 420 contamination, movements, etc.) and to control the behavior or the attention of infants. Thus, a 421 larger proportion of subjects end up discarded due to poor recording quality, and the data that In summary, our analyses revealed relatively small and unspecific effects of ASD risk on EEG 437 functional connectivity, potentially due to heterogeneity in how functional connectivity 438 abnormalities present themselves in different subjects. It nevertheless showed sex-specific 439 differences in how functional connectivity correlates with later autism severity scores. We 440 obtained these results using a recently published connectivity measure (CIPLV) that solves 441 many issues previously observed for similar measures. Further, we benefited from newly 442 released structural head templates for infants to perform connectivity analysis between EEG 443 sources rather than between EEG scalp signals, resulting in connectivity measures that can 444 more easily be associated with brain regions and that are less likely to be confounded by known 445 issues such as volume conduction and common sources. Our observations of ADOS scores 446 negatively correlating with functional connectivity for social affect in females and RRBs in males 447 might indicate that a different level of loading on functional connectivity is required to impact 448 ADOS scores depending on the sex, which might highlight sex-specific resilience to social affect 449 and RRBs symptoms in ASD. Average logit-transformed CIPLV connectivity. Displayed as a function of the age (x-axis), the site (columns), biological sex (row), and the diagnostic outcome groups (color). Whiskers represent the bootstrapped 95% con dence intervals. These 257 plots are for all-to-all connectivity averaged by recording. Functional connectivity (CIPLV) plots. Displayed for the three diagnostic groups (upper three graphs) and corresponding plots for the difference between the TLA and the two subgroups of ELA infants (lower two panels). The three top graphs and the two bottom graphs have been plotted using the same colormaps to allow fair comparisons. The left (right) side of these plots corresponds to the left (right) hemisphere. The order of the regions in each hemisphere is the same and is shown on the left side of the gure, from posterior (bottom) to anterior (top) regions. These plots only show the 100 region pairs with the strongest connectivity (top three panels) and the 100 pairs with the strongest between-group differences in connectivity (bottom two panels).

Figure 3
Logit-transformed CIPLV values within the different resting-state networks. Displayed per network (different panels), time points (x-axis), and diagnostic groups (color). Whiskers represent the bootstrapped 95% con dence interval.

Figure 4
Average logit-transformed CIPLV connectivity as a function of the distance. Displayed between regions (xaxis), age (rows), site (columns), and group (color). To smooth these lines, distances are split into 20 bins each covering 5% of the distribution. Shaded regions show 95% bootstrapped con dence intervals.

Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download. Supportinginformation.pdf