Globally elevated excitation–inhibition ratio in children with autism spectrum disorder and below-average intelligence

The participants were 49 boys with ‘non-syndromic’ ASD and 49 TD boys aged 6–15 years (see Additional file 1: Methods for recruitment and exclusion details). Each participant in the ASD group had a diagnosis of ASD confirmed by a child psychiatrist according to the DSM-5 criteria and interviews with children’s parents or caregivers. Parents/caregivers of all children were asked to fill in the Russian version of the Social Responsiveness Scale for children (SRS) [37].

IQ has been evaluated through standard scores on K-ABC subscales, as well as by calculating the Mental Processing Index (MPI) [38]. MPI assessment requires minimal use of verbal content, and this index is well suited both for TD children and for participants with atypical cognitive functioning, such as ASD [39]. The MPI includes four global scales: (1) the Sequential Processing scale, which assesses a child’s ability to arrange linearly or temporally related units of information in sequential or serial order, (2) the Simultaneous Processing scale, which involves abilities for holistic synthesis and integration of visual information, (3) the Learning scale, which evaluates a child’s ability to store and efficiently retrieve newly or previously learned information, (4) the Planning Scale, which focuses on Pattern Reasoning and Story Completion abilities. Scale scores are presented as age-adjusted standardized scores, normalized so that the mean is 100 and the standard deviation is 15, which is consistent with most cognitive assessment instruments. In the KABC-II, the term ‘average’ is used for MPI scores from 85 to 115, ‘above average’ for scores 116 to 130, ‘below average’ for scores 69 to 84, and ‘lower extreme’ for scores below 69. We used MPI scores to divide ASD participants into groups according to their cognitive abilities and to assess correlations with neurophysiological variables.

The study has been approved by the Ethical Committee of the Moscow State University of Psychology and Education. Verbal assent to participate in the experiment was obtained from all the subjects; parents/caregivers provided informed written consent.

For all the participants, the MEG data (see below) were available in the ‘eyes open’ (EO) condition, and for 45 TD and 38 ASD children the data were also available in the ‘eyes closed’ (EC) condition. The TD and ASD participants did not differ in age (Mann–Whitney U test, EO condition: N_TD = 49, N_ASD = 49, U = 1014.5, Z = 1.31, p = 0.19; EC condition: N_TD = 45, N_ASD = 38, U = 745.0, Z = 1.15, p = 0.25), but differed in the MPI scores (Mann–Whitney U test, p < 10^–10 for both conditions). The SRS scores were significantly higher in participants with ASD than in the control sample (Mann–Whitney U test, p < 10^–10 for both conditions).

The ASD participants were further subdivided into two groups based on their MPI IQ scores: average IQ (MPI > 85; ASD_>85) and below-average IQ (MPI < 85; ASD_<85). ASD children with IQ above 85 still had lower IQ than TD children (T test, p < 10^–8 for subjects in both EO and EC conditions). The SRS scores were significantly higher in ASD than TD participants (T test, p < 10^–10 for subjects in both EO and EC conditions, TD vs. ASD_>85 and TD vs. ASD_<85) and tended to be higher in ASD_<85 group than in ASD_>85 group (T test, EO: N_ASD>85 = 27, N_ASD<85 = 18, t₍₄₃₎ = 1.71, p = 0.09; EC: N_ASD>85 = 20, N_ASD<85 = 14, t₍₃₂₎ = 1.78, p = 0.08), but age did not differ in these three groups (Mann–Whitney U test, p’s > 0.3 for both conditions).

Detailed information about the experimental samples is summarized in Table 1. The number of participants of different age ranges is given in Additional file 1: Table S1.

Structural magnetic resonance (MR) scans (voxel size 1 mm × 1 mm × 1 mm) were performed on a General Electric Signa 1.5 T scanner. The T1 images were processed using the default FreeSurfer (v. 6.0.0) ‘recon-all’ pipeline. For the MEG analysis, the cortical surface was parcellated into 448 similar-size labels, as described by Khan et al. [40].

MEG was recorded with a 306-channel MEG system (Vectorview, Elekta-Neuromag). The ‘resting state’ recordings were always performed in the beginning of the recording session followed by experimental tasks analyzed elsewhere [41, 42]. The participants were instructed to sit calmly, first with eyes closed, and then, they were instructed to open their eyes. No other instructions were given. Two to four minutes of MEG were recorded in each experimental condition. Participants’ behavior was constantly monitored through a video camera positioned inside the MEG room. Trigger marks and experimenter’s comments were further used to select MEG intervals for analysis.

The signal was initially sampled at 1000 Hz with a 0.03–300 Hz band-pass filter. Head position was continuously monitored and corrected to each subject’s initial head position using MaxFilter software (v. 2.2). The temporal signal-space separation (tSSS) method [43] with correlation limit 0.9 was applied to compensate for correlated environmental noises.

Further steps of data preprocessing were performed with the MNE-python toolbox (v. 0.22.0) [44]. The raw data were down-sampled to 500 Hz. The signal-space projections method [45] was applied to the raw data to suppress biological artifacts (eye blinks and cardiac artifacts). The raw signal was subdivided into 1 s non-overlapping epochs. Epochs contaminated with bursts of myogenic artifacts were automatically detected using MNE-python function ‘annotate_muscle_zscore’ with default parameters (ch_type =  ‘mag,’ threshold = 5.0, min_length_good = 0.2, filter_freq = [110, 140]) and excluded from analysis. From this point on, only gradiometers were used for analysis. Those epochs where the head origin deviated from the initial position by more than 20 mm were excluded from analysis. The percent of dropped epochs was higher in the ASD than in TD children in the EO condition (Mann–Whitney U = 940.0, p = 0.03) and tended to be higher in the EC condition (Mann–Whitney U = 733.0, p = 0.10). To preclude group differences in timing of the analyzed epochs relative to the condition onset, we randomly excluded 10% of the ‘good’ epochs in the TD group and then have chosen the first 60 epochs per condition and subject in both ASD and TD groups for further analysis. As a result, there was no significant difference in epoch times relative to the condition onset time between the three experimental groups for both EO (Kruskal–Wallis one-way analysis of variance (ANOVA), H(2, N = 98) = 2.73, p = 0.25) or EC (Kruskal–Wallis one-way ANOVA, H(2, N = 83) = 2.38, p = 0.30) conditions. The mean distance of the head origin from the initial position did not differ between the groups (Kruskal–Wallis one-way ANOVA: EO, H(2, N = 98) = 2.58, p = 0.27, mean distance in TD/ASD_>85/ASD_<85 was 4.1/4.8/5.3 mm; EC, H(2, N = 83) = 0.58, p = 0.75, mean distance in TD/ASD_>85/ASD_<85 was 2.5/2.8/2.6 mm).

To estimate background noise, MEG data were recorded in the absence of a subject (empty room), directly before each subject’s MEG recording. The empty room MEG recordings were spatially filtered using the tSSS method [43] and were further used to estimate the background noise in the frequency range of interest, as well as to construct the noise covariance for the standardized low-resolution brain electric tomography (sLoreta) analysis [46], as described below.

Following previous studies [27, 29, 32], we sought to estimate the slope of aperiodic activity at the high-frequency part of the spectrum. To choose the optimal high-frequency range for the aperiodic slope analysis, we first estimated the power spectral density (PSD) in the sensor space using Welch’s method. Figure 1 shows an averaged over all participants power spectrum of a centrally located gradiometer (‘total power’) and the grand averaged power spectrum of the signal recorded by this gradiometer in the empty room.

The overall ‘total power’ spectrum is characterized by presence of peaks corresponding to the typical neuro-oscillation bands (alpha, beta). For frequencies above approximately 35 Hz, peaks are not obvious and the 1/f-like trend is apparent. We therefore set the lower bound for our aperiodic activity slope estimate to 35 Hz. On the other hand, a knee appears in the spectrum for frequencies above approximately 50 Hz where the ‘total power’ approaches that of the empty room spectrum (i.e., where power is approximately constant as a function of frequency). In order to also avoid potential effects of the 50 Hz power line cycle and the notch filter used for its removal, we set the upper bound for our aperiodic activity slope estimate to 45 Hz. In the 35–45 Hz range, the grand averaged over all subjects in EO condition spectrum displays a clear 1/f-like characteristic (i.e., it is well approximated by a line on the log–log scale; Fig. 1).

We performed source localization analysis using linearly constrained minimum variance (LCMV) beamformer spatial filters [47] that significantly reduce the contribution of biological artifacts (e.g., myogenic activity) into the brain activity estimated at the source level [48]. In order to gauge the effect of residual myogenic contamination on the group differences in spectral slope, we also compared the LCMV results with those obtained using sLoreta [46]. In contrast to beamforming methods, the methods based on the least-squares minimum norm algorithms such as sLoreta tend to overfit non-brain signals [49]. Therefore, the contribution of myogenic artifacts to the activity estimated at cortical sources using sLoreta is expected to be greater than in the case of LCMV beamformer (see Additional file 1: Results: ‘Contribution of myogenic artifacts to the group differences in spectral slope’).

MEG data were co-registered with individual’s MR image of the head, and the forward model was created using a one-layer boundary element model (BEM) and surface-based source space (4096 vertices per hemisphere). The raw data were filtered between 30 and 140 Hz, and the data covariance matrix was estimated individually for each condition. The LCMV beamformer filter with a regularization parameter of 0.1 was calculated for the source orientation that maximizes power. The same filter was applied to each 1 s data epoch. We then calculated the neural activity index (NAI) that represents the signal normalized to the spatially inhomogeneous noise [47]. In each source vertex, the spectral power was estimated using Welch’s method. The power spectra at each vertex were normalized by the maximum value at the corresponding spectra and then averaged within each of 448 labels. For each label, we then fitted a linear regression line to the logarithm of the power spectrum between 35 and 45 Hz vs. the logarithm of the frequency (‘polyfit’ function, Python library ‘NumPy,’ v. 1.20.2). The linear term coefficient was used as an estimate of the 35–45 Hz spectral slope (hereinafter referred to as the ‘spectral slope’).

We also estimated the aperiodic slope in the 2–45 Hz range using the ‘FOOOF’ parameterization algorithm [50]. A detailed description and discussion of these results are given in Additional file 1: Results (‘Spectral slope estimated with FOOOF’).

While beamformers provide good estimates of spectral power changes (e.g., when contrasting conditions) [51], they might provide suboptimal estimates of the absolute power. Therefore, to quantify the periodic spectral power, we performed source localization with sLoreta that has zero localization error [46]. For this analysis, MEG and empty room signals were band-passed in the 1–47 Hz range. The noise covariance matrix was derived from one minute of ‘MaxFiltered’ empty room data. The sLoreta inverse solution was estimated with the following parameters: semi-orthogonal orientation of the dipole source to the cortical surface (parameter loose set to 0.4), depth weighting parameter set to 0.8, and signal-to-noise ratio (SNR) set to 1. The inverse operator was applied to each one-second data epoch, and the power spectra were obtained at the source level using Welch’s method. To separate the periodic from aperiodic spectral power in the alpha and beta bands, we applied the FOOOF algorithm [50] to the average power spectra of each label in the 2–45 Hz range. The periodic component was then estimated as: [original spectra—aperiodic fit]. The mean alpha and beta powers were assessed as an average spectral power of the periodic component in the 7–13 Hz and 14–25 Hz bands, respectively. This approach gives a reasonable approximation on the power of periodic component in alpha and beta ranges (see ‘Results’ section).

To ensure that our results were not biased by head size, we calculated individual ‘sensitivity maps’ (‘mne.sensitivity_map’ function in MNE-python), which estimate how well each source is sampled by a sensor array (i.e., the planar gradiometers). The mean sensitivity in a label was then used as a nuisance variable in our regression analysis.

To estimate the rank-order stability of the spectral slopes, we calculated intraclass correlation coefficients (ICC) between conditions (EO/EC). ICC was estimated for a fixed set of ‘raters’ (EO, EC) and for the average of all ‘ratings’ (mean spectral slopes) [52].

We calculated Spearman correlation coefficients to estimate the relationship between spectral slope and IQ, SRS, and age, because we expected a monotonic, but not necessarily strictly linear relationships. To control for nuisance variables (background noise, sensitivity, spectral power, age), we calculated partial Spearman correlations.

The Benjamini–Yekutieli false discovery rate (FDR) correction for multiple comparisons was applied to the p values when tests were calculated for multiple cortical labels. The threshold 0.05 was applied to both the univariate and FDR-adjusted p values.

To become a clinically useful signature of E–I balance, the spectral slope must not only be sensitive to changes in the neural E–I ratio, but also reflect trait-like individual differences that remain stable across different functional states. In awake humans, the E–I balance is shifted toward higher inhibition during states associated with high alpha power [53‐55]. We therefore expected that if the spectral slope in children with and without ASD is sensitive to changes in the E–I ratio, then it would be steeper (more negative) in the EC condition (characterized by stronger alpha activity), as compared with the EO condition. We also proposed that if the spectral slope represents a stable individual characteristic, it should have high rank-order stability between the EO and EC conditions.

Figure 2A, B shows the cortical distribution of the spectral slopes in the three groups of children. Using ICC, we estimated the rank-order stability of the spectral slopes of the separate cortical regions (448 labels) between the EC and EO conditions. We also estimated the rank-order stability of the spectral slope averaged over all cortical labels (hereafter, the mean slope). The rank-order stability of the mean slope was good (ICC₍₈₂₎ = 0.87, p < 0.000001) [52] and exceeded that in the individual cortical labels, where it ranged from 0.06 to 0.79 (Fig. 2C). Considering the highest rank-order stability of the mean spectral slope, we used this measure for further analysis.

A paired t test revealed a highly significant and strong decrease in the mean slope from the EO to EC condition in the combined group of subjects (T₍₈₂₎ = 5.81, p < 0.000001, Cohen d = 0.64), as well as in the separate groups (TD: T₍₄₄₎ = 3.67, p = 0.0006, Cohen d = 0.55; ASD_>85: T₍₂₂₎ = 3.38, p = 0.0003, Cohen d = 0.71; ASD_<85: T₍₁₄₎ = 3.31, p = 0.005, Cohen d = 0.85; Fig. 2D).

The mean spectral slope values for the three groups of participants and the two experimental conditions are shown in Fig. 3A. The general linear model analysis with factors Group, Condition, and Age revealed a main effect of Age (F_(1,79) = 12.76, p = 0.0006, η² = 0.14) indicating a general flattening of the spectral slope with brain maturation (Fig. 3B). The highly significant main effect of Condition (F_(2,79) = 38.3, p < 1e−10, η² = 0.33) did not interact with Group (Condition × Group: F_(2,79) = 2.4, p = 0.1, η² = 0.06) or Age (Condition × Age: F_(2,79) = 2.6, p = 0.1, η² = 0.03), thus suggesting that the strong changes in spectral slope associated with opening/closing the eyes are comparable across age and in groups from different diagnostic categories. As expected, we found a significant main effect of Group (F_(2,79) = 6.95, p = 0.0017, η² = 0.15): The slope was flatter in the ASD_<85 than in either TD (F_(1,79) = 7.2, p = 0.009) or ASD_>85 (F_(1,79) = 13.8, p = 0.0004) groups. Unexpectedly and by contrast, though, the slope for the ASD_>85 group tended to be steeper than that in the TD group (F_(1,79) = 2.80, p = 0.098).

We repeated the analysis with factors Group and Age for the EO condition for which the data were available for a greater number of participants. Again, there was a significant effect of Age (F_(1,94) = 18.1, p = 0.00005, η² = 0.16) and Group (F_(2,94) = 8.4, p = 0.00045, η² = 0.15), which was explained by a flatter slope in the ASD_<85 group than in the rest of participants.

To test whether there is a link between IQ and the mean spectral slope in TD children, and whether it differs from that in children with ASD, we calculated correlations between the spectral slope and IQ in the TD and combined ASD groups. The ASD vs. TD difference in partial Spearman correlation coefficients (when adjusting for age) was significant for the EO condition (N_TD = 49, N_ASD = 49, Z = 1.67, p = 0.047), and the same tendency was observed for the EC condition (N_TD = 45, N_ASD = 38, Z = 1.41, p = 0.079). While children with ASD demonstrated reliable negative correlations between the mean slope coefficient and IQ, no such correlations were found in the TD group (Fig. 3C).

We also tested for the correlation between the mean slope and SRS scores in the ASD groups. In children with ASD, there was a tendency for correlation between MPI IQ and SRS scores: N = 45, Spearman R = − 0.28, p = − 0.06). However, the correlations between the mean slope and SRS scores were not significant (EO: N_ASD = 45, R = 0.14, p = 0.34; EC: N_ASD = 34, R = 0.29, p = 0.1).

To ensure that the link between the mean spectral slope and IQ is not driven by brain regions wherein source estimates are most susceptible to myogenic contamination, we calculated partial Spearman correlation coefficients between MPI IQ and spectral slopes in the separate cortical labels, while controlling for age and sensitivity (Fig. 4A; see Methods for details). For the EO condition, the largest group of significant correlations (p < 0.05, FDR corrected) overlapped the supramarginal area and temporoparietal junction of the right hemisphere (Fig. 4B, see Additional file 1: Table S2 for the full list of the cortical labels that survived the FDR correction). In general, the spatial distribution of the correlations is evidence against a potential ‘myogenic explanation’ of their origin. Similar negative correlations were observed for the EC condition (Fig. 4A), but none of them survived correction for multiple comparisons (Fig. 4B).

It has been argued that changes in the spectral slope of brain activity are secondary to the changes in periodic activity (e.g., power of alpha or beta rhythms) [56]. The power of periodic activity varies with functional and behavioral state, which in turn may depend on a subject’s diagnosis and IQ. To ensure that the group differences in the spectral slope were not explained by differences in the periodic activity, we estimated the periodic spectral power in the alpha and beta bands.

Figure 5 shows cortical distributions of alpha and beta periodic power in the three groups of participants. In accord with the results of Muthukumaraswamy and colleagues [56], a stronger periodic alpha and beta power was associated with a steeper, more negative spectral slope in both TD and ASD participants and in both experimental conditions (Table 2). However, unlike the slope, the grand averaged periodic power did not differ between the groups (ANOVA Group effect for the alpha band, EO: F_(2,94) = 1.89, p = 0.16, η² = 0.038; EC: F_(2,80) = 1.29, p = 0.28, η² = 0.031; ANOVA Group effect for the beta band, EO: F_(2,94) = 1.4, p = 0.26, η² = 0.028; EC: F_(2,80) = 2.1, p = 0.13, η² = 0.049).

To further ensure that the link between spectral slope and IQ in the ASD group did not stem from inter-individual differences in periodic activity, we calculated partial Spearman correlations between the mean spectral slope and IQ while controlling for the mean alpha and beta power. Partialling out the periodic power did not significantly affect the correlation between the mean spectral slope and MPI IQ in children with ASD (EO: N = 49, R_partial = − 0.43, p = 0.003; EC: N = 38, R_partial = − 0.35, p = 0.04).

An important limitation of our study is the high heterogeneity of our ‘idiopathic’ ASD sample that included subjects with different etiologies of ASD. To confirm the validity of the spectral slope biomarker, it would be important to investigate participants with monogenic disorders and animal models that are characterized by predominant shifts of the E–I balance toward excitation [9‐11] or inhibition [5, 6, 78]. Another caveat is that putative abnormalities of the E–I ratio inferred here by the 1/f slope may not be specific to autism, but rather play a general role in the pathophysiology of intellectual disability. Future research should explore this question. It is also important to note that our study included a relatively small number of subjects, and they were all boys. Since the E–I balance may be differently affected in males and females with ASD [35], the possibility of generalizing our results to girls requires additional research.

There are also a few potential improvements that could increase the reliability of measurement of the 1/f spectral slope and promote its practical application. Firstly, while our participants were not engaged in any specific activity, one should consider using a modified experimental paradigm that allows for longer periods of artifact free data in children [92]. Secondly, the accuracy of the slope estimation could be improved by increasing the upper limit of the frequency range beyond 45 Hz, e.g., when the power line cycle is 60 Hz and/or a better SNR is available. Thirdly, the new and rapidly developing technology of ‘on-scalp’ MEG [93‐95] may help to obtain equally high-quality data in all parts of the head, including the frontal areas, which are typically positioned furthest away from the sensors in ‘conventional’ MEG systems.