Concurrent Multimodal Imaging Demonstrates That EEG-Based Excitation/Inhibition Balance Reflects Glutamate and GABA Balance

Abstract

Magnetic resonance spectroscopy (MRS)-derived excitation/inhibition (E/I) ratio is recognized as a valuable index of attentional modulation, the plasticity and stability of learning states, and disordered functioning. Inferring the E/I ratio from electroencephalography (EEG), however, offers greater accessibility and superior temporal resolution and therefore holds strong potential for advancing research on dynamic neural processes. Yet, the underlying neurochemical mechanisms contributing to changes in EEG-based E/I ratio remain unclear. In this study we used concurrent EEG and MRS to examine how the EEG-based E/I ratio correlates with the MRS-based E/I ratio in sleeping humans of both sexes (n = 15). The MRS-based E/I ratio was calculated as the ratio of Glx (glutamate + glutamine) to GABA + (GABA + coedited macromolecules) concentrations in early visual areas. We estimated 10 candidate EEG-based E/I indices using four algorithms across multiple spontaneous frequency bands from the occipital region. Uniquely, we quantified the associations between EEG- and MRS-based E/I ratios by separately analyzing between-subject and within-subject variations. We found that each EEG-based E/I algorithm showed reliable and positive associations with MRS-based E/I, particularly in the alpha band, which is known to play a key role in attentional modulation and in the plasticity and stability of learning states. These results highlight the potential of EEG-based E/I measures to serve as practical indices of neurochemical dynamics in early visual cortex.

Significance Statement

Although a balance between cortical excitation and inhibition is important for many functions, measurement of this balance has remained a challenge. Candidate methods to efficiently measure excitation/inhibition balance via electroencephalography (EEG) have been proposed, but validation of these methods using concurrent measurement of neurochemistry (i.e., using magnetic resonance spectroscopy) has been lacking. The results reported here show that several candidate EEG-based measures of excitation/inhibition balance are reliably and positively associated with the ratio of excitatory to inhibitory neurotransmitters. Such results provide a necessary foundation for advancing methods, basic research, and applied research using efficient EEG measures of excitation/inhibition balance to understand brain states.

Introduction

Brain functioning is largely regulated through a dynamically changing balance of excitatory and inhibitory (E/I) activities (Hensch et al., 1998; Shew et al., 2011). Such dynamics are closely linked to a variety of phenomena including neural plasticity (Hensch et al., 1998; Bavelier et al., 2010; Vogels et al., 2011), brain development (Hensch, 2004; Takesian and Hensch, 2013), disordered functioning (Bruining et al., 2020; Geertjens et al., 2022; van Nifterick et al., 2023), and cognition-related phenomena (Hellyer et al., 2016; Amil and Verschure, 2021; Yang et al., 2023). Reflecting the fact that E/I ratio varies on multiple spatial and temporal scales and is difficult to measure noninvasively (Ahmad et al., 2022; Craven et al., 2022), a variety of methods are used to assess E/I in humans. These include the use of electrophysiological (EEG) signatures and magnetic resonance spectroscopy (MRS). Each of these approaches holds great promise and yet faces serious limitations.

EEG-based approaches to quantifying E/I ratio (i.e., EEG-E/I) come from several theoretical and computational sources, including scale-free dynamics (Gao et al., 2017), criticality theory (Poil et al., 2012), and neural mass modeling (Yokoyama and Kitajo, 2023), and have been used to investigate phenomena ranging from vigilance (Yang et al., 2023) to Alzheimer’s disease (van Nifterick et al., 2023). Each approach has proposed mechanistic bases for its links with cortical excitation or inhibition, but empirical data linking brain regions’ neurochemistry to EEG measures has been lacking (Ahmad et al., 2022; Geertjens et al., 2022). In contrast, EEG-E/I measures have relative advantages include their wide availability, efficient implementation, and data from multiple brain regions simultaneously.

The use of MRS to study E/I (i.e., MRS-E/I) ratio presents a different and complementary set of trade-offs. Studies of E/I ratio using MRS have been conducted to study brain plasticity, including that involved in perceptual learning (Shibata et al., 2017; Tamaki et al., 2020), as well as the neural underpinnings of autism spectrum disorder (Rojas et al., 2015). While MRS-E/I has the distinct benefit of having a clear biological interpretation, namely, excitatory (glutamate) and inhibitory (gamma-aminobutyric acid; GABA) metabolite concentrations, it is limited to low temporal resolution and restricts the measurement only inside an MRI scanner, thereby precluding measurement in populations with contraindications or during certain behaviors, and is relatively expensive. Further, MRS estimates are noisy (Craven et al., 2022) and may not provide a direct ratio measure of E/I balance (Lea-Carnall et al., 2023).

There have been efforts to link electrophysiological measures and MRS data. These include multiple reports of associations between MRS-measured GABA concentrations with electrophysiological frequency bands (Edden et al., 2009; Balz et al., 2016). However, these studies measured E/I ratio in separate sessions, precluding detection of short-timescale covariations. Yet, such dynamic fluctuations in E/I ratio are essential for understanding processes such as plasticity, pharmacological response, and state-dependent cognition. Whether EEG-derived E/I indices reflect the neurochemical balance measured by MRS remains unresolved. This question is critical for determining whether the neurochemical basis of MRS-based E/I balance is shared by EEG-based measures of E/I balance.

To address this, we conducted a multimodal investigation using simultaneous EEG and MRS recordings in healthy humans to directly test whether EEG-E/I indices correlate with MRS-E/I ratios. We estimated multiple EEG-E/I indices from spontaneous activity across several frequency bands in the occipital region and compared them with Glx/GABA + ratios obtained via MRS from the early visual cortex (i.e., the ratio of glutamate + glutamine to GABA + coedited macromolecules). Importantly, we examined both between-subject and within-subject covariation across these modalities to assess whether EEG-E/I captures minute-to-minute neurochemical fluctuations or merely stable trait-level differences.

We found that EEG-E/I indices—particularly those derived from α-band activity—consistently and positively tracked MRS-E/I ratios in between-subjects comparisons. These findings suggest that EEG can provide widely accessible, biologically grounded indices of cortical excitatory/inhibitory balance, opening a path toward noninvasive, dynamic monitoring of E/I dynamics in both basic and clinical neuroscience.

Materials and Methods

Experimental design

We reanalyzed data from Study 1 of Tamaki et al. (2020), which consists of 18 humans of both sexes who slept for between 60 and 90 min following a perceptual learning task. Simultaneous polysomnography including EEG and MRS were recorded throughout the sleep session. See the original paper for full procedural details.

In the present study, we estimated the association between MRI-E/I and EEG-E/I using not only between-subjects variations but also within-subjects variations. While much previous work has focused on the relatively stable differences between individuals’ levels of excitation and inhibition (Edden et al., 2009; Cousijn et al., 2014; Balz et al., 2016), within-subjects variation in E/I ratio are central to many research domains such as vigilance (Yang et al., 2023) and learning-related neuroplasticity (Bavelier et al., 2010; Tamaki et al., 2020) because these topics are inherently concerned with intraindividual changes. Separating these results helps identify both the overall associations between MRS-E/I ratio and EEG-E/I ratio as well as their relative isomorphism in the context of different applications.

We segmented EEG and MRI by 5 min period (see below, MRS-based E/I ratio and EEG-based measures of E/I ratio). We then used Bayesian multivariate linear mixed-effects modeling and maximum-likelihood cross-validation to quantify the ability to predict MRS-based E/I using EEG-based estimates (see below, Statistical analysis). We separated these estimates into between-subjects and within-subjects analyses.

MRS-based E/I ratio

Our primary results used the originally published MRS data (Tamaki et al., 2020; Study 1), wherein a MEGA-PRESS sequence was used (1.25 s TR; 2.2 × 2.2 × 2.2 cm voxel, placed on midline in early visual cortices; 3 Tesla Siemens scanner; see Figure S3 for example spectrum and Table S1 for full MRS reporting table). Because existing MRS sequences such as MEGA-PRESS would have a greatly reduced signal quality if recorded from more than a single voxel at a time, we recorded spectra from only one voxel. We made two modifications to the MRS data analysis reported in the previous publication. First is the partial volume correction (PVC) to account for relative proportions of gray matter (i.e., to mitigate the influence of gray matter proportion on associations with EEG measures). Second, we separated them into 5 min sections of time instead of the originally published 10 min time sections (mean number of 5 min windows per participant = 13.1, SD = 3.6). As with the originally published work, we used LCModel software (Provencher, 1993) to identify Glx (Glutamate + Glutamine) and GABA + (GABA + coedited macromolecules) concentrations normalized to N-acetylaspartic acid (NAA). Due to a lack of isomorphism between E/I balance and a ratio of Glx to GABA+ (Lea-Carnall et al., 2023), we also report correlations with these metabolites separately in the Supplemental Information (normalized to NAA; Fig. S2). We then estimate MRS-based E/I ratio by dividing Glx by GABA+. We chose to continue using LCModel to estimate Glx and GABA+, rather than other candidate softwares, to maintain congruence with the original findings because reliable results in the original paper indicate that E/I from LCModel reflects true underlying cortical changes (Tamaki et al., 2020). That is, Tamaki et al. (2020) found functional correspondences of sleep-dependent change in Glx/GABA+, as estimated with LCModel, in the form of benefits to perceptual plasticity and increased resilience to retrograde interference. In addition, LCModel has been found to provide the most typical estimates of metabolite concentrations, as it shows the highest correlation of GABA estimates with the median of estimates from other software (Craven et al., 2022).

EEG-based measures of E/I ratio

For details of EEG preprocessing steps, such as MR artifact rejection, see Methods in Tamaki et al. (2020; Study 1). In all analyses we report EEG measures estimated from the closest seven electrodes (Oz, O1, O2, PO3, PO4, & PO7) to the MRS voxel, since the MRS voxel was placed over early visual areas (Tamaki et al., 2020).

We assessed four approaches to quantifying electrophysiological activity from EEG. Three of these approaches were calculated separately for three canonical frequency bands (see below). There was a total of 10 EEG-E/I measures to test associations with MRS-E/I. The first approach was overall amplitude. While the overall amplitude is not a measure of E/I per se, associations between MRS-E/I within specific frequency bands have been observed previously (e.g., theta and low beta; Tamaki et al., 2020). This is why we included the overall amplitude as one approach. The second approach reflected scale-free dynamics reflected by the 1/f power slope (Gao et al., 2017), and the third was an extension of this scale-free perspective that derives from criticality theory (Bruining et al., 2020). The last approach was a set of estimates utilizing a neural mass model of excitatory and inhibitory dynamics (Yokoyama and Kitajo, 2023). We show the detailed information for each approach below.

Frequency band amplitude

As one of the most basic measures identifiable within EEG signals, we included simple amplitude (i.e., square root of power) measures from several frequency bands. We used finite impulse response (FIR) bandpass filtering in EEGLAB (Delorme and Makeig, 2004) to isolate theta (5–8 Hz), alpha (8–12 Hz), and low beta (12–20 Hz) bands, estimated the amplitude in 5 s windows, and then calculated the trimmed mean amplitude for each 5 min measurement period (i.e., for robustness to outliers, after removing the 5% lowest and 5% highest amplitude windows, calculating the arithmetic mean of the amplitudes within the remaining windows). These three frequency bands were examined due to previous work finding links between MRS-based E/I ratio and theta as well as low beta power (Tamaki et al., 2020), and the use of alpha amplitude envelopes in the implementation of the criticality-based E/I ratio (Bruining et al., 2020). While previous findings have also indicated that higher gamma frequencies may be linked to E/I balance, we chose not to examine these frequencies because they are much more susceptible to noise when measured with EEG (in contrast to MEG; Cousijn et al., 2014) and would also be partially captured in our 1/f measurements described below.

While there have been some indications of links between frequency bands and E/I ratio (Porjesz et al., 2002; Duncan et al., 2014; Balz et al., 2016), we include amplitude here primarily as a baseline approach against which to assess the more algorithmically complex approaches to estimating EEG-based E/I ratio. That is, as an algorithmically simple derivation from resting-state EEG, we consider amplitudes to be valuable measures to assess overall data reliability (e.g., see Supplement Across-channel correlations; we would expect amplitude measures to have high cross-channel correlations and a lack of such correlations would indicate poor data quality), test the divergence of EEG-E/I measures from amplitudes (see Supplemental Information Associations among EEG measures), as well as provide parsimonious reflections of MRS measures if the relevant correlations were to be found (Edden et al., 2009; Tamaki et al., 2020).

EEG-E/I from 1/f slope (“1/f”)

The next approach to derive E/I ratio from EEG signals is based on the widespread observation that neural activity, in temporal, spatial, and frequency domains, displays scale-free dynamics (power-law, or “1/f” when in the frequency domain; Linkenkaer-Hansen et al., 2001; He, 2014; Hahn et al., 2017). Computational derivations of the slopes of these 1/f distributions have been linked to cortical E/I ratio due to the relative timescales of excitatory and inhibitory interneurons’ firing rates (Gao et al., 2017; Lendner et al., 2020). As with previous work (Gao, Peterson, and Voytek, 2017; Yokoyama and Kitajo, 2023), we estimated the 1/f slope coefficient between 30 and 50 Hz (bandpass filtered) for each 5 min window of interest, which avoids contamination of the 1/f slope with deviations from this slope associated with alpha and beta frequencies. Such deviations dominate the power–frequency slope in lower frequencies. We implemented this approach with Python scripts available in conjunction with Yokoyama and Kitajo (2023). Recently developed analytical methods are available which have been used to estimate the 1/f slope coefficient using a wider band of frequencies by also accounting for periodic components of the power–frequency spectrum (i.e., deviations from the 1/f slope; Donoghue et al., 2020). We chose not to use these methods due to their relative complexity, the need for hyperparameter choices to model periodic components (Donoghue et al., 2022), and the possible collinearities between estimated coefficients. Due to the shared underlying theory, we expect that results using the 1/f slope from the 30–50 Hz window should be similar to results if using a more complex algorithm. That is, we expect that lower (i.e., more negative and steeper) 1/f slope to be more inhibitory-dominant and higher (i.e., less negative and flatter) 1/f slope to be more excitatory-dominant.

Unlike the other approaches, 1/f slope does not consider power or amplitude within discrete frequency bands, but instead the change in power across a wide range of frequencies. The following approach, based in criticality theory, is an elaboration of these principles of scale-free dynamics.

EEG-E/I from criticality theory (“fE/I”)

A recently developed algorithm for identifying E/I ratio from EEG data is grounded in the computational biology framework of criticality, which characterizes the nature of many biological systems as existing at the cusp of state transitions (Bak, 1999; Poil et al., 2012; Márton et al., 2014). One application of this theoretical approach is to neural criticality, which proposes that neural activity exists approximately at a balanced ratio between excitation and inhibition to maximize the flexibility of macroscale neural computations (Shew et al., 2011; Cocchi et al., 2017; Avramiea et al., 2022; Hengen and Shew, 2025). To estimate this type of EEG-based E/I ratio (Juarez Martinez et al., 2019; Ahmad et al., 2022), we used an algorithm which is based in criticality theory and has been validated in computational, pharmacological (Bruining et al., 2020), and animal models (Kat et al., 2024); the developers of this algorithm label it “functional E/I ratio” (fE/I). We do not know of any previous work linking fE/I to MRS measures.

We applied this approach to the same three bandpass-filtered frequency bands we considered when examining simple amplitude measures. In brief, this algorithm first determines the degree of criticality, which is inferred from the relationship between the window size of an analyzed signal and the variability in its cumulative detrended amplitude profile. Second, the degree of criticality is linked to the signal amplitude; in inhibitory-dominant regimes, these two values should be positively correlated, while in excitation-dominant regimes, they should be anticorrelated (Bruining et al., 2020; Diachenko et al., 2024). Using code published in conjunction with Bruining et al. (2020), fE/I values were estimated for each of the above-defined three frequency bands (theta, alpha, and low beta). As in the publication establishing the method, estimation of fE/I used a window size of 5 s with an interwindow overlap of 0.8 (Bruining et al., 2020). We also tested the effects of increasing the window sizes to 15 s and found minimal changes and here report only 5 s window sizes for simplicity. For each 5 min measurement period, we also estimated the detrended fluctuation analysis (DFA) coefficient and excluded measurement periods with DFA below 0.6, as recommended by the developers of this approach (Bruining et al., 2020; Diachenko et al., 2024).

EEG-E/I from neural mass model (“NMM-E/I”)

A fourth approach comes from a body of work instantiating neural mass models, that is, computational models of neural circuitry such as cortical columns; such models can be scaled up and can simulate complex dynamics and characterize dynamic and nonlinear features of neural activity (Jansen and Rit, 1995; Cooray et al., 2023). Yokoyama and Kitajo (2023) extended earlier neural mass models (Jansen and Rit, 1995), which had been developed for purposes such as modeling epileptic dynamics, with a key insight: such models can include estimates of the relative activity of circuits’ excitatory and inhibitory neural populations. Following this, and so they could compare to the 1/f approach described above, Yokoyama and Kitajo (2023) showed that the E/I ratio derived from such models could capture the fluctuating activity of humans’ electrophysiological activity during sleep. This approach to estimating E/I ratio is quite new, and to our knowledge it has not been widely assessed. Thus, we estimated this NMM-E/I using Python code available in conjunction with Yokoyama and Kitajo (2023) for each 5 min measurement period and for each of the canonical frequency bands described above (i.e., theta, alpha, and low beta). We also used the algorithm’s hyperparameters from Yokoyama and Kitajo (2023). Unlike the fE/I algorithm, the NMM-E/I ratio algorithm does not involve separation into 5 s segments. We also tested the NMM-E/I on a single wider set of frequencies, more similar to the approach implemented in Yokoyama and Kitajo (2023), on bandpass-filtered data between 5 and 20 Hz (i.e., the lower and upper limits of the three frequency bands considered in the above analyses). The resulting NMM-E/I values did not improve correspondence to MRS-E/I relative to the values estimated from distinct canonical frequency bands (see Supplementary Information Alternative estimation of NMM-E/I).

Statistical analysis

For both within-subjects and between-subjects analyses, we used two separate statistical approaches, Bayesian multivariate linear mixed-effects modeling and maximum-likelihood cross-validation, which were complementary to one another. In the first approach, in order to best estimate the effect size (i.e., correlation coefficient) between each EEG measure and MRS-based E/I ratio, we fit a Bayesian multivariate linear mixed-effects model using brms in R (Bürkner, 2017). Using this approach we specified each of the measures as variables to predict (i.e., “response” variables). Each response variable was fit with an overall intercept, a random-effects by-participant intercept, and a fixed-effect autoregressive term with an order of one to account for nonindependence of temporally proximal measurements. Within the multivariate model, the correlations across the response variables of the by-participant intercepts were estimated (in brms “Wilkenson” notation the formula for each E/I measure was response ∼ 1 + ar(time, order = 1, groupBy = subjectID) + (1|subjectR|subjectID); for schematic, see Fig. S1). In this way we estimated simultaneously the between-subjects effects, in the form of the correlations between different response variables’ by-participant intercepts, as well as within-subjects effects, in the form of the residual correlations remaining after accounting for the autoregressive and by-subject terms. We fit an alternative model without the autoregressive term, due to the possibility that this term would suppress observed within-subjects effects, but there was no qualitative difference when compared with a model including the autoregressive term. We report the model with the autoregressive term because this model accounts for all known sources of nonindependence in each measurement (i.e., including univariate temporally correlated noise), thereby leading to the correlation matrix of residuals to be a putatively iid measure of within-subjects covariation. We report the means of the posterior distributions of these two correlations, both between-subjects and within-subjects, as the best empirical estimate of the linear bivariate links between each measure. In addition, this model included an estimation of the remainder of the correlation matrix (e.g., between all EEG-E/I measures as well as estimates of Glx and GABA+); we report these correlations to determine the extent to which our EEG measures are isomorphic and are potentially more reflective of Glx or GABA+ variations (Fig. S2). The fitted model object, including data, model specification, and all posterior samples, is available at osf.io/ucgj6/.

In the second statistical approach, we used maximum-likelihood cross-validation as tests of the reliability of linear associations between EEG-based and MRS-based E/I ratio and to mitigate the possibility of overfitting our data. This approach involved iteratively and randomly partitioning the data into “training” and “test” sets, fitting a maximum-likelihood linear regression model predicting MRS-based E/I ratio to the training set, and assessing the R-squared (i.e., proportional reduction in sum of squared error; out-of-sample delta R-squared $\mathrm{\Delta }{R}_{\mathrm{oos}}^2$) on the test set (Dale et al., 2021; Scott et al., 2023). Between-subjects and within-subjects analyses were necessarily implemented somewhat differently. Between-subjects cross-validation involved first calculating the subject-level means for each measure, and then using simple random partitioning of this aggregated dataset into 80%/20% subsets, with data being fit with a bivariate robust linear model to 80% of the data and $\mathrm{\Delta }{R}_{\mathrm{oos}}^2$ assessed on the remaining 20%. The null model in this case was an intercept-only linear regression. In contrast, within-subjects cross-validation involved randomly selecting a single observation (i.e., 5 min time period) from each participant to keep as a test set, fitting a linear model including a by-participant covariate (i.e., to control for participant-level means in MRS-E/I ratio) to the remaining training set and then assessing the proportional change in out-of-sample squared error $(\mathrm{\Delta }{R}_{\mathrm{oos}}^2)$ when compared with a model with only the participant covariate as a predictor. This is the predictive value of the EEG-based measure. To control for between-subjects variations in the predictor EEG-based measures, for each of these within-subjects cross-validation models, the EEG-E/I measure was first normalized within participants, after using a Yeo–Johnson transformation to minimize univariate skew, and then z-scored. Here we report the median $\mathrm{\Delta }{R}_{\mathrm{oos}}^2$ of 500 resampled models for each EEG-E/I measure.

Results

We first assessed data quality and excluded data coming from any 5 min window in which at either fE/I ratio or MRS-E/I ratio demonstrated high noise relative to signal. MRS data quality was quantified using the ratio of the uncertainty of metabolites’ estimated concentration relative to the corresponding estimate; as in the originally published work (Tamaki et al., 2020), a 20% level of the Cramer–Rao lower bound was used for MRS exclusion. EEG measures including amplitude, 1/f, and NMM-E/I ratio do not have established cutoff thresholds for exclusion. Thus, we assessed EEG quality using the DFA exponent coefficient criterion which Bruining et al. (2020) describe in conjunction with the fE/I method. This involves estimating the DFA coefficient and excluding windows in which the coefficient is less than 0.6. While the original sample included 18 participants’ datasets, three participants had only three or fewer remaining MRS values and were excluded from further analysis. This left 15 participants data, with a mean amount of remaining data being 65.5 min.

Between-subjects associations between EEG-E/I and MRS-E/I

Interindividual differences in certain measures out of 10 estimates of EEG-E/I were reliably correlated with MRS-E/I (Fig. 1, Table 1). Each of the candidate algorithms produced E/I estimates with a positive cross-validated predictive variance explained in at least one frequency band (i.e., $\mathrm{\Delta }{R}_{\mathrm{oos}}^2>0$; Table 1). Here, we consider each of the links with positive $\mathrm{\Delta }{R}_{\mathrm{oos}}^2$ values to be reliable. Among 10 EEG-EI measures, the criticality-based fE/I approach derived from alpha oscillatory activity was most closely linked to MRS-E/I across participants, followed by 1/f slope, then NMM-E/I derived from low beta activity, and then fE/I derived from theta band (see Fig. S4 for scatterplots). All other measures, including overall amplitudes, were not reliably predictive of interindividual differences in MRS-E/I, as their $\mathrm{\Delta }{R}_{\mathrm{oos}}^2$ were not positive. The lack of association with overall amplitudes provides a justification for all of the candidate algorithms’ approaches to estimating E/I ratio; further, the strong performance of criticality-based α-band E/I supports the developer’s original findings using this algorithm in alpha frequencies (Bruining et al., 2020; Diachenko et al., 2024).

Figure 1.

Between-subject associations between EEG-E/I measures and MRS-E/I. Correlation of by-participant random intercepts from the Bayesian multivariate mixed-effects model is plotted for each EEG measure’s correlation with MRS-based ratio of estimated Glx to GABA+. Numerical values are shown in Table 1. The values are sorted by decreasing $\mathrm{\Delta }{R}_{\mathrm{oos}}^2$ (i.e., measures at the top have the highest out-of-sample variance explained). $\mathrm{\Delta }{R}_{\mathrm{oos}}^2$ is shown in Table 1. Dashed line separates variables explaining reliable out-of-sample variance with those that do not; we consider all associations above the dashed line to be reliable. For scatterplots of reliable correlations, see Supplemental Information.

Table 1.

Between-subject associations between EEG-E/I measures and MRS-E/I

EEG measure	$\mathrm{\Delta }{R}_{\mathrm{oos}}^2$	Correlation (r)
1/f slope	0.131	0.211
Amplitude: alpha	<0	0.081
Amplitude: low beta	<0	−0.054
Amplitude: theta	<0	−0.132
fE/I: alpha	0.190	0.258
fE/I: low beta	<0	−0.09
fE/I: theta	0.112	0.129
nmmEI: alpha	<0	0.106
nmmEI: low beta	0.103	0.066
nmmEI: theta	<0	−0.049

Each EEG measure’s association with MRS-based E/I ratio (i.e., Glx/GABA+) is shown in the form of cross-validated out-of-sample variance explained $(\mathrm{\Delta }{R}_{\mathrm{oos}}^2)$ and effect size extracted from a Bayesian multivariate mixed-effects model (r). The effect sizes (r) are also plotted in Figure 1, sorted by $\mathrm{\Delta }{R}_{\mathrm{oos}}^2$. All $\mathrm{\Delta }{R}_{\mathrm{oos}}^2$ values above zero are considered to be reliable, with values below indicated by <0.

Within-subjects associations between EEG measures and MRS-based E/I

After controlling for participant-level effects, several approaches to EEG-E/I were reliably associated with MRS-E/I (i.e., $\mathrm{\Delta }{R}_{\mathrm{oos}}^2>0$; Fig. 2, Table 2). While this wide range of EEG-E/I approaches providing within-subjects out-of-sample predictiveness for MRS-E/I is promising, a closer investigation of the effects tempers this promise somewhat. First, although out-of-sample variances explained were positive, they were quite small, with most measures having $\mathrm{\Delta }{R}_{\mathrm{oos}}^2<0.01$ (i.e., <1% of out-of-sample variance explained). Second, of the EEG measures with positive $\mathrm{\Delta }{R}_{\mathrm{oos}}^2$ values, only theta amplitude had a positive correlation with MRS-E/I (r = 0.08). These results are in striking contrast to the between-subjects findings, where a relatively large amount of variance was explained with positive relations between EEG-E/I and MRS-E/I measures.

Figure 2.

Within-subject associations between EEG-E/I measures and MRS-E/I. Correlation of by-participant random intercepts from the Bayesian multivariate mixed-effects model, indicating each EEG measure’s correlation with MRS-based ratio of estimated Glx to GABA+. Each panel is sorted by decreasing $\mathrm{\Delta }{R}_{\mathrm{oos}}^2$ (i.e., measures at the top have the highest out-of-sample variance explained). Dashed line indicates whether variables explained out-of-sample variance or did not; we consider all associations above the dashed line to be reliable.

Table 2.

Within-subject associations between EEG-E/I measures and MRS-E/I

EEG measure	$\mathrm{\Delta }{R}_{\mathrm{oos}}^2$	Correlation (r)
1/f slope	<0	−0.079
Amplitude: alpha	0.021	−0.118
Amplitude: low beta	0.013	−0.097
Amplitude: theta	0.005	0.081
fE/I: alpha	0.004	−0.112
fE/I: low beta	0.001	−0.066
fE/I: theta	<0	0.051
nmmEI: alpha	0.009	−0.120
nmmEI: low beta	<0	0.009
nmmEI: theta	<0	−0.036

Each EEG measure’s association with MRS-based E/I ratio (i.e., Glx/GABA+) is shown in the form of cross-validated out-of-sample variance explained $(\mathrm{\Delta }{R}_{\mathrm{oos}}^2)$ and effect size extracted from a Bayesian multivariate mixed-effects model (r). Effect sizes (r) are plotted in Figure 2, sorted by $\mathrm{\Delta }{R}_{\mathrm{oos}}^2$. All $\mathrm{\Delta }{R}_{\mathrm{oos}}^2$ values above zero are considered to be reliable, with values below indicated by <0.

Supplemental analyses

EEG-E/I reliability assessment

We implemented a cross-channel correlation approach to estimate a heuristic reliability of our EEG data and EEG-based E/I measures. Due to basic principles of volume conduction (Burle et al., 2015), as well as our intention to measure relatively large-scale variations in physiological activity, we expected highly homogeneous estimates of EEG measures across our measured posterior channels. A lack of such a correlation for all EEG measures would indicate poor EEG data quality, whereas a lack of correlation within a single EEG measure may indicate computational instability of that measure. Because each of our EEG-based methods involved first estimating a value for each channel, we were able to estimate the cross-channel correlations for our analyzed posterior electrodes within each method. We tested these correlations within subjects by finding the pairwise cross-channel partial rank correlations in estimated E/I values, partialling out participant-level intercepts. We then report the median cross-channel partial rank correlation (Fig. 3A). While we do not have a strict reliability criterion for these correlation measures, we believe that the relative cross-channel correlations should provide a general sense of the computational stability of the measures; amplitude, fE/I, and 1/f measures demonstrated acceptable reliability.

Figure 3.

A, Median rank correlations between analyzed posterior channels’ estimates of EEG-E/I. Most estimates were highly correlated across channels, although NMM-based estimates had relatively low cross-channel reliability. B, Ratios of between-subjects variances to within-subjects variances. Most measures had less than a 3:1 ratio of these variances, with MRS-based E/I having between-subjects variances approximately 1.45 times as large as within-subjects variances.

Between-subjects variability relative to within-subjects variability

In light of differences in between-subjects relative to within-subjects correspondences between EEG-based and MRS-based measures of E/I, we sought to quantify the relative between-subjects and within-subjects variances for each measure. To determine between-subjects variance of a given measure, we first found each participant’s median value for each measure and then calculated the variances of these median values. To determine within-subjects variances, we first calculated each participant’s variance for each measure and then found the median of these variances. The ratio of between-subjects variance to within-subjects variances is reported in Figure 3B, with no clear bias toward larger between-subjects variance or within-subjects variance.

Discussion

Our results show that EEG can be used to estimate E/I ratio in a manner that is compatible with variations in E/I ratio estimated from MRS. In particular, the functional EEG-E/I algorithm derived from criticality theory was most closely associated with between-subjects differences in MRS-based E/I ratio. The reliance of this algorithm on temporal variations in alpha power corroborates the work of the developers of the fE/I algorithm (Bruining et al., 2020; Diachenko et al., 2024). Beyond this criticality-based EEG-based E/I ratio measure, each of the other EEG-based E/I ratio approaches were also positively associated with between-subjects variations in MRS-based E/I ratio. These results showed broad support for each of the approaches, which were themselves not highly correlated (Fig. S2). A lack of isomorphism between the EEG-based measures means that future work may benefit from combining these algorithms to best characterize cortical E/I.

The relative contribution of estimates from separate frequencies is of particular interest. While they are not yet fully understood, the heterogeneous generative mechanisms underlying variations in theta, alpha, and beta oscillations provides researchers and clinicians a rich and multidimensional signal from which to infer shifts in cortical states (Scheeringa and Fries, 2019; Shafiei et al., 2023). Indeed, independent frequency bands have been previously associated with glutamate and/or GABA levels (Edden et al., 2009; Tamaki et al., 2020), and we found that a broad spectrum of frequencies showed links to MRS-based E/I ratio when analyzed with different algorithms in the current work (i.e., theta and alpha analyzed with fE/I; low beta analyzed with a neural mass model; gamma when calculating a 1/f slope). We believe that combining amplitude and E/I estimates from several frequency bands may provide additional precision. Future studies are necessary to attempt such a combinatorial search which would be outside the scope of our limited dataset.

In contrast to between-subjects patterns of E/I ratio, within-subjects associations between MRS-based E/I ratio and EEG-based E/I ratio were much weaker. The reasons for the weakness of these associations cannot be known with certainty. However, it is likely to arise partly to a decrease in signal strength and an increase in noise when compared with between-subject comparisons. That is, estimating EEG-based E/I ratio as well as MRS-based E/I ratio on windows of only 5 min is likely to lead to relatively noisy estimates. While shorter measurement windows measuring GABA with MRS have been used in the past (Frank et al., 2022), longer windows (e.g., 10 min) are more standard. Unfortunately, our available data is sufficiently limited that estimating within-subjects associations between EEG-based E/I ratio and MRS-based E/I ratio appears to be largely hampered by the reduced signal-to-noise ratio. The MR context of these estimates may also have itself reduced data quality in estimating EEG-based E/I. Thus, it would be ideal to collect much longer simultaneous EEG and MRS measurements, where enough numbers of 10 min window could be collected; however, this would itself constrain the measurements of E/I balance changes occurring on rapid timescales (see below).

There may be alternative sources of divergence between EEG-based and MRS-based E/I ratios. On the one hand, the neural mechanisms giving rise to each of our measured signals may be distinct. While MRS measured bulk concentrations of metabolites in a relatively deep and spatially precise portion of posterior cortex, our EEG measurements were reflecting temporal variations in the activity of relatively superficial and widespread cortical neurons. Similarly, the causal determinants of oscillatory activities, and their variations over time, are not yet known with temporal or molecular precision. It may be the case that distinct cell types, neuromodulators, or long-range connections influence EEG signals which would not be reflected in the measurement of Glx and GABA+ in a single voxel within posterior cortices (Ahmad et al., 2022). The relations between EEG-based E/I ratios and MRS-based E/I ratio may also be nonlinear, such that different EEG analysis algorithms could capture different aspects of underlying MRS-based E/I ratio; larger datasets would be needed to determine whether an optimal and possibly nonlinear combination of EEG-based measures could better account for MRS-based E/I ratio.

Our results also reinforce the fact that the phenomena we are considering may be varying over time on fundamentally different timescales. One of the most notable functions of variations in excitation and inhibition are their changes over developmental time, that is, years or even decades (Hensch, 2004; Takesian and Hensch, 2013), while in contrast, cognition-related modulation of E/I would necessarily operate on subminute timescales (Hellyer et al., 2016). The methods we considered provide clear tendencies toward longer or shorter timescales, with MRS-based measures needing a minimum of several minutes, NMM-based measures providing estimates which are nominally relevant on the subsecond level, and the other methods falling somewhere in between. It may be possible, then, that these measures simply do not reflect underlying cortical changes that operate on similar timescales, leading to a lack of evident within-subject relations between them.

Robust links between EEG-based E/I ratio and MRS-based E/I ratio at a between-subjects level provides validation for EEG-based measures to be used in certain areas of research for which they have developed, namely, in investigating special populations such as individuals with ASD (Bruining et al., 2020), Alzheimer’s disease (van Nifterick et al., 2023), or monogenetic neurodevelopmental disorders (Geertjens et al., 2022). Our results indicate that the between-subject differences in E/I revealed by contrasting these groups with healthy controls are likely to have common mechanistic bases whether they are measured using EEG or MRS. In contrast, for research which relies heavily on both within-subject and between-subject contrasts, such as the study of neuroplasticity (Shibata et al., 2017) or vigilance (Yang et al., 2023), EEG-based E/I ratio measures may not allow the assumption of mechanistic isomorphism.

While alignments or divergences between EEG-based E/I ratios and MRS-based measures provide evidence regarding a common mechanistic basis, it is also worth emphasizing that neither EEG-based nor MRS-based estimates are themselves pure measures of excitation and inhibition. MRS at 3 Tesla involves averaging over a relatively large voxel and over several minutes of time, and thus the resulting spectra are imprecise in each dimension. More importantly, however, the resulting estimates of metabolites may be noisy and they reflect quantities of metabolites that may not act as neurotransmitters at any given time. Lastly, the very definition of excitation/inhibition balance differs depending on the discipline and application, and the voxel-aggregated relative concentration of glutamate to GABA across all cellular compartments would not be expected to be a direct measure of all theoretically meaningful forms of E/I ratio. This is especially the case, for example, in light of recent findings that measured GABA may increase in response to increases in either excitation or inhibition (Lea-Carnall et al., 2023); our observed correlations with the separated metabolites may thus be more directly interpretable than a ratio and are reported in the Supplemental Information. Thus, while we believe that our results provide important evidence for a shared mechanism influencing both MRS-based and EEG-based E/I ratios, we caution the reader against presupposing that either method could provide a “ground truth.”

While our results indicate that several recently developed algorithms for measuring E/I using EEG are indeed partially reflecting variations in the cortical ratio of glutamate to GABA, several limitations are present in our findings. First, the size of our dataset does not allow for precise estimates of effect sizes or compelling rejection of candidate EEG-based E/I ratio measures, and larger datasets would be needed to allow better combinatorial identification of EEG-based measures indexing underlying neurochemical changes. Second, the study of sleeping humans may lead to idiosyncratic results when contrasted to wakefulness. As an initial investigation, sleep is a valuable period to investigate because changes in vigilance state are likely to induce measurable variations in cortical E/I balance; however, if such results are to be extrapolated to phenomena occurring in awake humans (e.g., in relation to cognitive performance), there may not be direct generalizability of our results. Additionally, we unfortunately lack the quantity of data necessary to perform analyses separated by vigilance state. Each of these limitations motivates future comparisons using concurrent multimodal imaging in awake humans and with larger datasets.

Overall, although MRS-based E/I ratio is a relatively direct measure of actual molecules of interest within the brain, this measurement technique is resource intensive, spatially and temporally imprecise and limited to a single voxel per acquisition, and provides a quantification of spectra that is itself noisy. Indeed, the estimation of metabolite concentrations has no consensus standard and results from a single MR sequence may vary depending on the estimation algorithm (e.g., inter-algorithm correlations between GABA concentration estimates ranging from below 0.1 to above 0.8; Craven et al., 2022). Thus, a neurochemically validated EEG-based E/I ratio would be invaluable for future research. We tested whether EEG-based E/I ratio was correlated with MRS-based E/I ratio and found that they are reasonably well correlated with each other through two independent statistical approaches in early visual areas, with several limitations in our findings. Future studies are necessary to further overcome these limitations and test whether the reasonable association could be found in other brain regions.