Glymphatic dysfunction couples with cortical excitation–inhibition imbalance in epilepsy: Evidence from Rasmussen encephalitis

Abstract

Objective

The glymphatic system (GS) facilitates perivascular clearance of interstitial solutes and is modulated in part by neuronal activity. However, its relationship to cortical excitability in epilepsy remains unclear. We aim to clarify the mechanistic link between GS function and cortical excitation–inhibition (E–I) balance in patients with epilepsy.

Methods

We investigated this coupling in patients with Rasmussen encephalitis (RE), a rare epileptic disorder with unilateral cortical pathology. Using a hemispheric within‐subject design (N = 20), we compared the affected hemisphere (AH) and unaffected hemisphere (UH) within each patient, assessing glymphatic function via diffusion MRI (diffusion tensor imaging–analysis along the perivascular space [DTI‐ALPS] index) and E–I dynamics via resting‐state EEG spectral decomposition.

Results

The AH exhibited reduced DTI‐ALPS indices, elevated aperiodic exponents, and increased delta–theta oscillatory power. Across individuals, lower ALPS values in the AH correlated with higher aperiodic exponents (r = −.496, p = .026), but not with periodic EEG features. Principal component analysis (PCA) of channel‐level aperiodic asymmetry (exponent‐informed PCA) revealed a spatial pattern localized to atrophic cortical regions, which also showed strong correspondence with interhemispheric ALPS asymmetry (r = −.570, p = .009).

Significance

These findings demonstrate a spatially convergent link between glymphatic dysfunction, cortical activity with predominant inhibitory tone, and focal atrophy in the AH. Our results indicate that impaired perivascular clearance is correlated with altered cortical excitability, highlighting the potential role of the GS in maintaining neural stability in the epilepsy.

Cong Fu et al. demonstrate that glymphatic system dysfunction is linked to enhanced inhibitory cortical activity using diffusion MRI and EEG. These findings highlight a mechanistic link between perivascular fluid dynamics and neuronal activity, suggesting a role for glymphatic function in maintaining cortical stability in epilepsy.

Key points.

• Within‐subject analyses in Rasmussen encephalitis (RE) linked a lower diffusion tensor imaging–analysis along the perivascular space (DTI‐ALPS) index to an inhibition‐dominant aperiodic EEG profile.

• Its spatial colocalization with cortical atrophy further strengthened the biological plausibility of this association.

• Periodic oscillations did not explain DTI‐ALPS variance.

• Findings support a glymphatic–neuronal axis underlying cortical stability in epilepsy.

1. INTRODUCTION

The glymphatic system (GS) is a brain‐wide clearance pathway that removes metabolic waste by facilitating the exchange of interstitial fluid (ISF) surrounding neurons with cerebrospinal fluid (CSF). ¹ , ² Because neurons are highly sensitive to fluctuations in the extracellular environment, efficient clearance of extracellular solutes is essential for maintaining neural homeostasis. ³ The GS plays a central role in mediating the removal of pathological proteins, neurotransmitters, and ions from the brain parenchyma into the CSF and peripheral circulation, thereby supporting the balance of neuronal excitation and inhibition. ⁴ Synchronized neuronal action potentials have been shown to generate large‐amplitude, rhythmic ionic waves within the ISF in animal models, which significantly enhance glymphatic flow and molecular clearance. ⁵ Similarly, in rodent models of Alzheimer's disease, 40 Hz stimulation has been found to promote glymphatic clearance via neuronal synchronization. ⁶ , ⁷ Although these findings support a mechanistic link between neuronal activity and glymphatic function, whether such coupling exists in epilepsy remains unknown. To address this, the present study investigates the relationship between glymphatic function and brain electrical activity using advanced MRI and EEG in patients with glymphatic impairment.

Here, we employed patients with Rasmussen's encephalitis (RE) as an epileptic model to investigate hemispheric differences in glymphatic function and neuronal activity. RE is a rare, immune‐mediated neurological disorder characterized by progressive unilateral cortical atrophy and marked electrophysiological abnormalities on MRI and EEG, while the contralateral hemisphere remains relatively preserved. ⁸ , ⁹ , ¹⁰ Clinically, patients with RE present with focal motor seizures affecting the contralateral limbs. Notably, recent studies in epilepsy have shown that impaired glymphatic clearance is associated with solute accumulation and neuroinflammation, supporting the hypothesis that glymphatic dysfunction may contribute to excitation–inhibition (E–I) imbalance. ¹¹ , ¹² , ¹³ Taken together, these features make RE a compelling within‐subject model for controlled investigations of the association between glymphatic function and brain electrical activity.

To assess GS function in vivo, we used the diffusion tensor imaging–analysis along the perivascular space (DTI‐ALPS) index. ¹⁴ This metric leverages the orthogonal arrangement between medullary veins and major white matter fiber tracts in the periventricular region to assess diffusivity within the perivascular space, which serves as the primary pathway for glymphatic fluid transport. ¹⁵ At the level of the lateral ventricle body, the orientation of the perivascular space of medullary veins is orthogonal to major projection and association fibers, allowing diffusivity along the medullary veins to be analyzed independently from surrounding fiber tracts. Therefore, reduced ALPS values indicate impaired perivascular water movement, which is regarded as an indirect marker of glymphatic dysfunction. ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ In addition, we employed electrophysiological metrics encompassing both periodic (oscillatory) and aperiodic (1/f‐like) components. Periodic components in EEG reflect synchronized rhythmic activity within neural populations. In contrast, aperiodic components reflect asynchronous background activity and serve as a proxy for cortical E–I balance, where steeper 1/f slopes indicate increased inhibitory tone and flatter slopes reflect heightened excitability. ²⁰ , ²¹ , ²² , ²³

In this study, we employed RE as an in vivo human model of strictly unilateral pathology, designating the affected hemisphere (AH) as the region of interest (ROI) and using the relatively preserved contralateral (unaffected) hemisphere (UH) as an intra‐individual control to reduce intersubject variability. We quantified the DTI‐ALPS index and analyzed both periodic and aperiodic EEG components in the AH and UH. By integrating diffusion MRI and electrophysiological analyses, we aimed to elucidate the relationship between GS dysfunction and abnormal brain electrical activity—specifically neural oscillatory dynamics and E–I balance—and to determine whether impaired clearance mechanisms contribute to cortical instability in epilepsy.

2. MATERIALS AND METHODS

2.1. Participants

We conducted a retrospective case‐note review of 20 patients with RE who underwent presurgical multidisciplinary evaluation at the Department of Pediatrics, Peking University First Hospital. The inclusion criteria followed RE patients diagnosed according to the diagnostic criteria Part B of Bien et al. ⁹ Twenty RE patients with high‐quality MRI data were included for further study. The clinical characteristics of patients in the RE group are shown in Table S1. All participants or their guardians signed the informed consent form.

2.2. MRI acquisition

High‐resolution T1‐weighted images were acquired on a 3.0T MRI system (Philips Achieva) using a multi‐channel head coil. Imaging was performed in sagittal orientation with a magnetization‐prepared rapid gradient‐echo sequence. The main parameters were as follows: repetition time (TR) = 7.938 ms, echo time (TE) = 3.698 ms, flip angle = 8°, field of view (FOV) = 256 × 256 mm², voxel size = 1.0 × 1.0 × 1.0 mm³. We acquired diffusion‐weighted imaging data also using a 3 T Philips Achieva MRI system. Imaging was performed using a spin‐echo two‐dimensional (2D) EPI (echo planar image) sequence: b = 0 and 1000 s/mm² with 32 directions, TR = 8744 ms; TE = 65.3 ms, flip angle = 90°, FOV = 256 × 256 mm², voxel size of 2.0 × 2.0 × 2.0 mm³. Given that all participants were children, when sedation was required for MRI acquisition, oral chloral hydrate solution (.5 mL/kg; maximum 10 mL) was administered ~30 min before scanning. General anesthesia was not used.

2.3. Voxel‐based morphometry analysis

Gray matter volume (GMV) and total intracranial volume (TIV) of each hemisphere were quantified using the fully automated Computational Anatomy Toolbox (CAT) integrated within Statistical Parametric Mapping 12 (SPM12). Cortical segmentation was performed using the Unified Segmentation approach, followed by normalization with Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (DARTEL). An 8‐mm volume‐based smoothing kernel was applied to the data. The quality of the processed data was assessed visually using the retrospective quality assurance protocol provided by CAT12. GM ROIs of each side was masked using the LPBA40 (LONI Probabilistic Brain Atlas) with 28 regions for each hemisphere (Figure 1).

FIGURE 1.

Multimodal analysis pipeline. (A) GMV estimation. Individual anatomic MRI scans were subjected to tissue segmentation and spatial registration to quantify GMV across hemispheres. VBM was used to visualize regional atrophy differences between the AH and UH. (B) DTI‐ALPS calculation. ROIs were manually placed on color‐coded FA maps to extract diffusion metrics along association and projection fibers. The DTI‐ALPS index was calculated based on directional diffusivity aligned with perivascular spaces. (C) Periodic and aperiodic EEG analysis. EEG signals were recorded from symmetrical scalp electrodes over the AH and UH. PSD was decomposed into periodic (oscillatory) and aperiodic (1/f‐like) components. AH, affected hemisphere; DTI‐ALPS, diffusion tensor imaging–analysis along the perivascular space; E–I, excitation–inhibition; FA, fractional anisotropy; GMV, gray matter volume; PSD, power spectral density; ROI, region of interest; UH, unaffected hemisphere; VBM, voxel‐based morphometry.

2.4. Quantification of DTI‐ALPS

The automated DTI‐ALPS index was used to evaluate the glymphatic function. FSL commands from FMRIB Software Library version 6.0.7.6 (FSL; Oxford Centre for Functional MRI of the Brain, Oxford, UK) were applied. The original DTI data were corrected for eddy‐current–induced distortions and head motion (FSL eddy). No reverse phase‐encoded images were acquired; therefore susceptibility‐induced distortion correction (FSL topup) was not performed. The outputs were then inputted into the DTI fitting tool to create fractional anisotropy (FA) maps and tensor images with six directions for each subject. FA map was linear co‐registered to the Johns Hopkins University–International Consortium for Brain Mapping‐fractional anisotropy‐1mm (JHU‐ICBM‐FA‐1mm) template, and the transformation matrix was applied to six diffusivity maps. For ROI placement, the uppermost layer of the lateral ventricle body was selected based on the JHU‐ICBM‐FA‐1mm template in the Montreal Neurological Institute (MNI)152 space, and a 5‐mm diameter ROI was placed on the bilateral projection fibers and association fibers, which correspond to the superior corona radiata (SCR) and the superior longitudinal fasciculus (SLF). The MNI coordinates of the centers of the bilateral ROIs for the projection fibers were: 24, −12, 24 and −28, −12, 24, respectively. For the association fibers, the MNI coordinates of the bilateral ROIs were: 36, −12, 24 and −40, −12, 24, respectively. We visually inspected all FA maps for ROIs co‐registered failures. The diffusivity values of D _xx, D _yy, and D _zz of bilateral SLF and SCR were automatically outputted for the DTI‐ALPS calculation. The DTI‐ALPS index of the AH and UH was calculated separately using the following Equation (1):

$$\text{ALPS index}=\frac{\text{mean}\left(D_{\text{xxproj}},D_{\text{xxassoc}}\right)}{\text{mean}\left(D_{\text{yyproj}},D_{\text{zzassoc}}\right)}$$ (1)

where D _xxproj is the diffusivity along the x‐axis in the projection fiber, D _xxassoc is the diffusivity along the x‐axis in the association fiber, D _yyproj is the diffusivity along the y‐axis in the projection fiber, and D _zzassoc is the diffusivity along the z‐axis in the association fiber.

2.5. EEG recording and preprocessing

EEG signals were recorded using a 64‐channel Nicolet system (10–20 standard) at 512 or 1024 Hz, with impedance <5 kΩ. Online referencing was applied to the electrode near central midline (Cz). For analyses, a subset of 19 standard scalp electrodes was selected: Fp1/2, F3/4, C3/4, P3/4, O1/2, F7/8, T3/4, T5/6, Fz, Cz, Pz. Interhemispheric differences were calculated for eight homologous pairs to capture asymmetry.

Five‐minute segments of artifact‐free EEG during sleep Stages II/III were manually extracted and processed using the EEGLAB toolbox. EEG data were re‐referenced to the average of all electrodes and downsampled to 256 Hz. To preserve spectral integrity, bad channels were not interpolated; instead, epochs containing persistent channel artifacts were strictly excluded from analysis. Signals were filtered using a zero‐phase Hamming‐windowed Sinc FIR bandpass filter (.5–80 Hz) and a 50 Hz notch filter (stopband 49–51 Hz). Artifacts were managed in two steps. First, independent component analysis (ICA) was performed to identify and remove cardiac artifacts, muscle activity, and slow eye movements. On average, 1–3 components were removed per subject. Second, manual artifact rejection of body movements was conducted to ensure optimal signal quality for analysis.

2.6. Analysis of EEG periodic and aperiodic components

The power spectral density (PSD) was estimated using the Brainstorm toolbox (Welch's method: 10s window, 50% overlap) across .5–80 Hz. The Fitting Oscillations and One‐Over‐F (FOOOF) toolbox¹ was utilized to decompose the PSD into background and oscillatory components. The total PSD was modeled as shown in Equation (2):

$$\mathrm{PSD}=L+\sum _{n=0}^N{G}_n$$ (2)

where L is the aperiodic “background” signal, which was model using a “fixed” mode, as shown in Equation (3):

$$L=b-\log \left(k+F^x\right)$$ (3)

where b is the broadband offset and x is the exponent. In this study, F represents the vector of input frequencies, which was restricted to .5–40 Hz. The parameter k, known as the “knee” parameter, was set to 0. With the knee parameter set to zero, the aperiodic component follows a linear trend in log–log space, such that the exponent quantitatively reflects the steepness of the power decay across frequencies: a higher exponent corresponds to a steeper spectral profile, indicating increased inhibitory tone. ²³

For the periodic component (G), peaks were fitted as Gaussians, as shown in Equation (4):

$$G_n=a*\exp \left(\frac{-{\left(F-c\right)}^2}{2w^2}\right)$$ (4)

with an amplitude, c center frequency, w the bandwidth of the Gaussian G, and F the input frequencies. The oscillatory power was then analyzed within 5 canonical frequency bands: delta (.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–80 Hz).

Specific FOOOF settings were: peak width limits = .5–12 Hz; max peaks = 3; min peak height = 3 dB; and detection threshold = 2 standard deviations (SD). All fits were visually inspected to ensure physiological validity.

2.7. Principal component analysis of aperiodic EEG differences

Principal component analysis (PCA) was performed to reduce the dimensionality of the interhemispheric aperiodic exponent differences (Exp‐diff) across EEG channels and to detect the spatial relationship between Exp‐diff and ALPS‐diff. The analysis included eight symmetrical EEG channel pairs: Fp1/2‐diff, F3/4‐diff, C3/4‐diff, P3/4‐diff, F7/8‐diff, T3/4‐diff, T5/6‐diff, and O1/2‐diff. Prior to PCA, all Exp‐diffs were standardized (z‐transformed). The suitability of the dataset for PCA was confirmed using the Kaiser–Meyer–Olkin (KMO) measure of sampling adequacy and Bartlett's test of sphericity. The overall significance of the PCA model was evaluated using a chi‐square test, with a threshold of p < .05. PCA was conducted using the JASP toolbox (Version 0.19.2), and components were retained based on eigenvalues greater than 1, resulting in three principal components (PCs). To enhance interpretability, factor loadings were extracted using the PC extraction method and subjected to orthogonal Varimax rotation. For each subject, PC scores were computed as the weighted linear combination of standardized Exp‐diff variables and their corresponding component loadings as shown in Equation (5):

$$\mathrm{PC}{k}^{\left(i\right)}=\sum j=1^p{L}_{\mathit{kj}}·Z_j^{\left(i\right)}$$ (5)

$p$: the total number of variables (Exp‐diff across eight symmetrical EEG channel pairs); $L_{\mathit{kj}}$, the loading of the jth variable on the kth principal component (after Varimax rotation); $Z_j^{\left(i\right)}$, and the z‐transformed value of the jth variable for subject i. The resulting Exp‐informed PC scores were used in subsequent correlation analyses with the DTI‐ALPS metric. To assess the robustness of the PCA solution given the modest sample size, we performed a leave‐one‐out (jackknife) sensitivity analysis. Specifically, the PCA procedure (z‐scoring, extraction of three components, and orthogonal Varimax rotation) was repeated after excluding each participant in turn using a custom Python script that replicated the main analysis settings. Components were matched to the full‐sample solution based on maximal loading similarity (allowing for sign flips), and the stability of component loadings and the PC2‐ALPS‐diff association was summarized across iterations.

2.8. Quantification and statistical analysis

The right AH was flipped to left side to increase statistical power in this smaller cohort. Data are presented numerically as N (%) or mean ± SD. The Shapiro–Wilk test was used for normality test. We compared the DTI‐ALPS index and aperiodic exponent of intra‐subject two hemispheres using two‐tailed paired‐samples t‐test. Correlation analyses were performed to examine the relationship between DTI‐ALPS and aperiodic features. We assessed the association of the ALPS index difference of both sides with aperiodic exponent difference of both sides, and three PCs using a correlation analysis. p‐Value <.05 was considered statistically significant. All statistical analyses were conducted using JASP toolbox. Voxel‐based linear models compared AH atrophy profiles to UH using SPM, correcting for multiple comparisons using the false discovery rate (FDR) procedure. We regard FDR‐corrected p < .05 as statistically significant. In addition to voxel‐wise testing, we conducted region‐level analyses (27 predefined anatomic regions) by extracting AH and UH volumes and performing paired t‐tests with FDR (Benjamini–Hochberg) correction. These results were reported in the Supplementary Materials.

3. RESULTS

3.1. Patient information

A total of 20 patients diagnosed with unilateral RE were included in the study. Among them, 11 patients (55.0%) presented with left hemispheric involvement, whereas 9 patients (45.0%) had right hemispheric involvement. Regarding sex distribution, 5 patients (25.0%) were male and 15 patients (75.0%) were female. The mean age ± SD at onset was 4.45 ± 2.67 years, and the average disease duration was 21.10 ± 19.80 months. At inclusion, the mean age of patients was 6.20 ± 2.90 years. Patients detail characteristics were presented in the Table S1.

3.2. Interhemispheric cortical atrophy (AH vs UH)

As shown in Table 1, we compared the inter‐hemispheric GMV; the results presented that GMV was significantly lower in the AH comparing with the UH using the Wilcoxon signed‐rank test (Z = −3.920, p < .001). For the detection of GMV pattern in each side, the t‐value map was shown from one sample t‐test (compared to zero) in Figure 2A. Compared to the UH, our findings showed widespread cortical atrophy in the AH group. Whole‐brain voxel‐based morphometry (VBM) detected areas of cortical atrophy, a significant local minima observed in the inferior frontal gyrus with the greatest cluster size (t = −28.932, p < .01). Other atrophy regions also found in the anterior lobe of the cerebellum (t = −10.004, p < .01), the precentral gyrus (t = −7.771, p < .01), and the superior parietal lobule (t = −8.167, p < .01). The significance level was set at p < .05, FDR‐corrected, with the cluster size >200 voxels. The full VBM results including the findings for cortical atrophy were displayed in Table 2 and Figure 2.

TABLE 1.

Comparison between AH and UH in GMV, DTI‐ALPS, and EEG features.

	AH (N = 20)	UH (N = 20)	Group statistic	p Value
GMV (mL)	223.997 [202.334, 245.660]	254.540 [247.130, 261.950]	−3.920	<.001*
DTI‐ALPS index	1.213 (.196)	1.373 (.191)	−6.384	<.001*
D xxassoc (×10−3 mm2/s)	.9780 [.8932, 1.063]	.9385 [.8790, .9980]	1.232	.231
D xxproj (×10−3 mm2/s)	.7375 [.6021, .8729]	.6920 [.6211, .7629]	1.027	.313
D zzassoc (×10−3 mm2/s)	.6775 [.5795, .7755]	.6000 [.5244, .6756]	3.024	.001*
D yyproj (×10−3 mm2/s)	.7085 [.5855, .8315]	.6110 [.5324, .6896]	3.061	.001*
Aperiodic exponent
Mean	2.258 (.527)	2.187 (.550)	2.207	.040*
Fp1/2	2.231 (.549)	2.197 (.599)	1.003	.329
F7/8	2.235 (.585)	2.168 (.636)	1.075	.296
F3/4	2.280 (.501)	2.163 (.586)	1.857	.079
C3/4	2.258 (.562)	2.162 (.583)	2.634	.016*
T3/4	2.239 (.538)	2.206 (.554)	.684	.502
T5/6	2.239 (.547)	2.175 (.520)	1.222	.237
P3/4	2.258 (.544)	2.181 (.548)	1.857	.079
O1/2	2.325 (.496)	2.241 (.496)	2.188	.041*
PSD
Delta (× 10−3)	3 [3.0, 65]	2 [1.15, 28.5]	2.912	.002*
Theta (×10−4)	5.414 [4.184, 6.644]	2.715 [2.315, 3.115]	1.979	.048*
Alpha (×10−4)	1.412 [.902, 1.922]	1.397 [.931, 1.863]	1.157	.261
Beta (×10−5)	1.949 [1.258, 26.40]	1.571 [1.27, 1.87]	1.232	.231
Gamma (×10−6)	1.515 [1.148, 1.882]	1.275 [.513, 2.037]	−.411	.701

Note: For normally distributed data, a paired Student's t‐test was applied. For data that did not follow a normal distribution, the Wilcoxon signed‐rank test was used as a non‐parametric alternative to the paired t‐test. Values are presented as mean (SD) or median [interquartile range].

Abbreviations: AH, affected hemisphere; ALPS, along the perivascular space; D _xx, diffusivity along the perivascular space, signifying left to right directional diffusion; D _yy, diffusivity along projection fibers, indicating superior to inferior directional diffusion; D _zz, diffusivity along association fibers, depicting anterior to posterior directional diffusion; GMV, gray matter volume; PSD, power spectral density; UH, unaffected hemisphere.

^* p < .05.

FIGURE 2.

Hemispheric comparisons of GMV, glymphatic function, and EEG spectral metrics. (A) GMV comparison. Violin plots show that GMV was significantly lower in the AH compared to the UH (Z = −3.920, p < .001). Upper panels showed t‐value maps from one‐sample t tests against zero, illustrating the spatial distribution of atrophy within each hemisphere. Lower panels presented paired t test maps directly contrasting GMV between AH and UH. (B) DTI‐ALPS index comparison. The AH showed significantly lower DTI‐ALPS values than the UH (t = −6.384, p < .001). A representative subject's FA map with four overlaid ROIs used for ALPS calculation was shown (slice at z = 100). (C) Aperiodic exponent comparison. The AH showed a significantly higher exponent than the UH (t = 2.207, p = .040). Group‐averaged aperiodic curves (Exp‐AH = 2.258, Exp‐UH = 2.187) were plotted below for reference. (D) Periodic power comparison. PSD analysis revealed significantly higher delta band power (Z = 2.912, p = .002) and theta band power (Z = 1.979, p = .048) in the AH relative to the UH using paired Wilcoxon signed‐rank tests. No significant differences were observed for alpha, beta, or gamma bands. Right panel shows interhemispheric differences expressed as log‐transformed ratios, Δlog10(power) = log10(AH) − log10(UH), for each frequency band. Individual subject values are shown as jittered points, with black circles indicating the median difference and vertical bars representing the bootstrap‐derived 95% confidence intervals. AH, affected hemisphere; DTI‐ALPS, diffusion tensor imaging‐analysis along the perivascular space; GMV, gray matter volume; UH, unaffected hemisphere.

TABLE 2.

GMV comparison between AH and UH.

Regions	MNI coordinates			Cluster size	Peak t values
	X	Y	Z
Inferior frontal gyrus	64.5	13.5	15	8770	−28.932
Cerebellum anterior lobe	4.5	−46.5	−33	940	−10.004
Precentral gyrus	39	−9	37.5	412	−7.771
Superior parietal lobule	24	−81	54	217	−8.167

3.3. DTI‐ALPS index results

The descriptive statistics for the DTI‐ALPS metrics are presented in Table 1. For the DTI‐ALPS of affected side (AH‐DTI‐ALPS), the mean value was 1.213 ± .196, ranging from .891 to 1.574. Similarly, the DTI‐ALPS index of unaffected side (UH‐DTI‐ALPS) had a higher mean of 1.373 ± .191 and a range from .977 to 1.646. Significant differences were found between the AH and UH for the DTI‐ALPS index (t ₍₁₉₎ = −6.384, p < .001) using a paired‐samples t‐test. The calculated inter‐hemispheric differentials, DTI‐ALPS (ALPS‐Diff) had a mean values of −0.157 ± 0.115.

Among the diffusion parameters, significant differences were observed in the diffusivities along y‐axis in the projection fibers and z‐axis in the association fibers. Specifically, D _yyproj was significantly elevated in the AH compared to the UH (Z = 3.061, p = .001). Similarly, D _zzassoc was significantly higher in the AH compared to the UH (Z = 3.024, p = .001). No statistically significant differences were observed for D _xxassoc (p = .231) or D _xxproj (p = .313) between hemispheres.

3.4. Periodic and aperiodic metrics

For the periodic EEG signal results, significant differences in PSD were observed between the AH and UH (Figure 2D). Delta band power was significantly higher in the AH (Z = 2.912, p = .002). Theta band power was also significantly increased in the AH (Z = 1.979, p = .048). However, no significant differences were found in alpha, beta, and gamma band power between hemispheres (p > .05 for all). Our results indicate that inter‐hemispheric asymmetry was most pronounced in the delta (delta‐diff: 9.875 × 10⁻⁵ [−1.901 × 10⁻³, 2.099 × 10⁻³]) and theta bands (theta‐diff: 2.162 × 10⁻⁶ [−2.05 × 10⁻⁴, 2.09 × 10⁻⁴]), followed by alpha‐diff (3.258 × 10⁻⁷ [−1.259 × 10⁻⁵, 1.324 × 10⁻⁵]), beta‐diff (9.725 × 10⁻⁸ [−2.949 × 10⁻⁶, 3.144 × 10⁻⁶]), whereas differences in higher‐frequency bands such as gamma (gamma‐diff: −1.605 × 10⁻⁹ [−1.359 × 10⁻⁷, 1.328 × 10⁻⁷]) was minimal or near zero. The topographic visualizations were presented in Figure S1.

For the aperiodic EEG metrics, exponent of the affected (Exp‐AH) and unaffected (Exp‐UH) hemispheres yielded mean values of 2.258 ± .527 and 2.187 ± .550, respectively. The comparison between Exp‐AH and Exp‐UH also revealed a significant difference (t ₍₁₉₎ = 2.207, p = .040). The calculated inter‐hemispheric differentials, aperiodic exponent (Exp‐diff), had mean values of .072 ± .145.

Using the same approach, we additionally analyzed data from 43 patients with other hemispheric disorders to validate the conclusions regarding electrical activity in this study. Details are provided in the Results of the Supplementary Methods and Figure S2.

3.5. DTI‐ALPS index correlation with periodic and aperiodic EEG metrics

From the ipsilateral hemisphere analysis (AH or UH), there was no remarkable correlation between DTI‐ALPS and EEG metrics, including periodic and aperiodic EEG signals. For the periodic EEG metrics, PSD of all five frequent bands presented lower correlation with DTI‐ALPS index. As shown in Figures 3A and S3, the DTI‐ALPS index from the affected hemisphere (ALPS‐AH) exhibited positive correlations with high‐frequency EEG bands, particularly gamma (r = .275, p = .240), whereas lower‐frequency bands such as delta and theta showed weaker or negative correlations (r < .2). In contrast, the DTI‐ALPS index from the unaffected hemisphere (ALPS‐UH) was positively correlated across all frequency bands, with the strongest associations observed in the alpha (r = .369, p = .109) and gamma (r = .328, p = .158) in Figure 3B. However, no significant correlation was found (p > .05). For the periodic exponent, there was a weak positive correlation with the DTI‐ALPS index (r = .045, p = .849) in AH. Similarly, the exponent in the UH showed a slightly stronger but still weak positive correlation with ALPS‐UH (r = .024, p = .920).

FIGURE 3.

Associations between DTI‐ALPS index and EEG spectral metrics. (A) Affected hemisphere (AH). Left: scatter plot of the diffusion tensor imaging–analysis along the perivascular space index in the affected hemisphere (ALPS‐AH) vs the aperiodic exponent (Exp‐AH), showing a weak, non‐significant correlation (r = .05, p = .849). Shaded band indicates the 95% confidence interval around the linear fit. Right: heatmap of Pearson's correlation coefficients between ALPS‐AH and EEG metrics (mean Exp and five frequent bands). (B) Unaffected hemisphere (UH). Left: scatter plot of ALPS‐UH vs Exp‐UH (r = .02, p = .920), with 95% confidence interval. Right: heatmap of correlation coefficients between ALPS‐UH and the five periodic band powers and mean Exp in the UH. (C) Interhemispheric differences (diff). Left: scatter plot of the difference in ALPS index vs the difference in exponent, revealing a significant negative correlation (r = −.496, p = .026). Right: heatmap of correlations between ALPS‐diff and the differences in each EEG metric. The DTI‐ALPS index and the aperiodic exponent (Exp) are dimensionless (unitless) measures; therefore, no physical units are shown on the axes. p‐Values are reported without multiple‐comparison correction.

Notably, the interhemispheric difference in DTI‐ALPS index (ALPS‐diff) showed a significant negative correlation with aperiodic exponent (Exp‐diff) (r = −.483, p = .026), and to a lesser extent with delta‐diff (r = −.172, p = .442) and theta‐diff (r = −.133, p = .548) bands in Figure 3C.

3.6. DTI‐ALPS index correlation with Exp‐informed PC scores

In addition to analyzing hemisphere‐averaged exponent (Exp) values, we also examined the exponent at each individual channel in Figure 4A. Among individual channels, a significantly higher aperiodic exponent was observed at the C3 channels in the AH (Z = 2.634, p = .016). Furthermore, the O1/2 channels showed a marginally significant difference (Z = 2.188, p = .041), whereas other channels results were far from significance (p > .05). Among all regions, F3/4 showed the largest hemispheric difference (F3/4‐diff: .117 ± .283), followed by C3/4‐diff (.097 ± .164), O1/2‐diff (.084 ± .172), P3/4‐diff (.077 ± .187), F7/8‐diff (.067 ± .281), and T5/6‐diff (.064 ± .234), whereas the smallest asymmetries were observed in T3/4 (.033 ± .218) and Fp1/2 (.034 ± .153).

FIGURE 4.

DTI‐ALPS index correlation with Exp‐informed PC scores. (A) Violin plots show Exp values at each EEG channel for AH and UH. A significantly elevated Exp was found in channel C3 (Z = 2.634, p = .016), with marginal significance in O1/2 (Z = 2.188, p = .041). The overlaid lines indicate the mean of Exp in each hemisphere. The bottom line plot shows Exp‐diff, with F3/4 showing the largest asymmetry (.117 ± .283). (B) Results of PCA on Exp‐diff values across EEG channels. The left panel shows a scree plot with the actual eigenvalues (solid line with dots) and simulated eigenvalues from parallel analysis (dotted line with triangles). PCA retained three PCs, as the actual eigenvalues exceeded the simulated ones. The right panel shows the rotated component loadings: PC1 is dominated by T5/6, O1/2, and T3/4; PC2 by Fp1/2, F7/8, and F3/4; and PC3 by C3/4 and P3/4. (C) We performed PCA on the aperiodic exponent differences across EEG channels (Exp‐diff) and then tested their correlation with ALPS‐diff. A significant negative correlation was observed for PC2 (r = −.57, p = .009). No significant associations were found for PC1 (r = −.30, p = .196) or PC3 (r = .10, p = .663). Shaded bands indicate 95% confidence intervals for linear regression. (D) The spatial distribution of PC2‐related EEG channels (Fp1/2, F7/8, F3/4) overlaid on VBM statistical results. The heatmap indicated the GMV differences between the AH and UH. The crosshair marked the voxel with the peak hemispheric GMV difference (t = −28.932, MNI: 64.5, 13.5, 15). ALPS‐diff, hemispheric difference in ALPS index; Exp, mean aperiodic exponent; Exp‐diff, hemispheric difference in exponent; n.s., not significant; PC, principal component.

Our results indicate that PCA successfully reduced the dimensionality of the EEG features while preserving the most relevant variance, with each PC reflecting unique patterns of variation among the features (Figure 4B). Specifically, the KMO test yielded an overall measure of sampling adequacy of .600. Bartlett's test of sphericity was significant (χ ² = 119.971, p < .001). Three PCs were extracted, explaining a cumulative variance of 84.3%. The rotated solution revealed distinct loading patterns that PC1 was primarily associated with T5/6‐diff (.969), O1/2‐diff (.895), and T3/4‐diff (.808), reflecting posterior‐temporal channels; PC2 captured Fp1/2‐diff (.911), F7/8‐diff (.834), and F3/4‐diff (.747), representing frontal regions; and PC3 was defined by C3/4‐diff (.895) and P3/4‐diff (.692), corresponding to central‐parietal channels. The chi‐square test for model significance yielded a significant result (χ ² = 25.289, p < .001).

To further assess the spatial relationship between Exp‐diff, ALPS‐diff, and cortical atrophy, we conducted correlation analyses using Exp‐informed PC scores derived from PC1, PC2, and PC3 (Figure 4C). Among the three components, PC2 scores showed a significant negative correlation with ALPS‐diff (r = −.57, p = .009), whereas PC1 (r = −.30, p = .196) and PC3 (r = .10, p = .663) were not significantly associated. Notably, the EEG channels contributing most strongly to PC2—Fp1/2, F7/8, and F3/4—corresponded to regions exhibiting the largest GMV differences between the AH and UH. As shown in Figure 4D and detailed in Table 2, the voxel with the largest t‐value (t = −28.932) was located at MNI coordinates X = 64.5, Y = 13.5, Z = 15, which falls within the inferior frontal gyrus, as defined by the LPBA40 anatomic atlas.

Robustness was further supported by a leave‐one‐out sensitivity analysis, which showed that both the rotated component structure and the PC2–ALPS‐diff association remained stable across all iterations (see Supplementary Results).

4. DISCUSSION

In the present study, we assessed the in vivo function of the GS utilizing a non‐invasive, diffusion‐based MRI metric (DTI‐ALPS) and explored its association with E–I imbalance in patients with RE comparing the AH and the UH groups. Our findings demonstrated that the AH group exhibited marked glymphatic dysfunction and a pronounced E–I imbalance relative to the UH group (Figure 2). Furthermore, increased glymphatic impairment was accompanied by a more marked inhibition activity (Figure 3). Finally, we observed a convergent spatial pattern linking E–I imbalance with localized cortical atrophy and impaired glymphatic clearance (Figure 4).

RE provides a unique neurobiological framework for investigating GS dysfunction and neural activity imbalance due to its highly lateralized pathology. Clinically characterized by drug‐resistant focal seizures and progressive neurological deterioration, RE is marked by chronic inflammation confined to a single cerebral hemisphere, leading to pronounced hemispheric atrophy and cognitive decline. ¹⁰ In this study, we applied RE as a disease model to explore GS function using the DTI‐ALPS index that is sensitive to hemispheric differences in perivascular diffusivity. ¹⁴ The marked asymmetry in DTI‐ALPS indices observed between the AH and UH further validates the utility of this approach. Electrophysiologically, RE is characterized by unilateral slow‐wave activity and epileptiform discharges on EEG, mirroring the lateralized seizure focus. ⁹ These features align with MRI findings that consistently reveal focal inflammatory and atrophic changes in the temporal and frontal cortices of the affected hemisphere. ⁸ , ²⁴ Such convergent anatomic and functional asymmetries reinforce the suitability of RE as an internal contrast model, facilitating robust assessment of glymphatic electrophysiological coupling in focal cortical pathology.

Retrospective analyses indicate that GS function is reduced in patients with epilepsy compared to control subjects, ¹¹ and the ipsilateral DTI‐ALPS index was significantly lower than the contralateral DTI‐ALPS index, ¹³ which supported our findings. Moreover, GS function can be fully restored following anterior temporal lobectomy in patients with temporal lobe epilepsy, highlighting the critical role of the GS in epilepsy pathogenesis. ²⁵ We included the EEG data from the non–rapid eye movement (NREM) phase; since glymphatic clearance is primarily active during sleep, any disruptions in sleep quality due to neurological disorders could negatively impact glymphatic function. ² There were group differences in the interhemispheric DTI‐ALPS in patients with focal cortical dysplasia (FCD). ¹³ One explanation for the relationship between FCD and glymphatic dysfunction is that FCD type II is particularly associated with an increased risk of sleep‐related epilepsy, ²⁶ potentially exacerbating glymphatic dysfunction due to the exclusive or predominant occurrence of seizures during sleep.

Aberrant E–I patterns are implicated in numerous neurological disorders, such as epilepsy. ²³ Our RE model was also an epileptic model that showed hemispheric epileptiform discharges and focal motor seizures. Jiang et al., ²⁷ using interictal stereotactic EEG recordings from patients with epilepsy, reported that the seizure‐onset zone (SOZ) exhibited a more negative 1/f power slope than non‐SOZ regions, indicating greater E–I imbalance. Duma et al. ²⁸ found that patients with temporal lobe epilepsy showed significantly larger exponents in bilateral frontal and temporal areas, consistent with an inhibition‐dominant E–I profile. Similarly, Yang et al., ²⁹ using intracranial depth electrodes, observed that the 1/f slope of non‐oscillatory local field potentials was lower during ictal compared to interictal states. These findings are consistent with our results, in which the SOZ (AU group) showed a higher aperiodic exponent relative to the non‐SOZ (HU group) during the interictal period, reflecting more inhibitory activity. This supports the interictal suppression hypothesis. ³⁰ , ³¹

Emerging evidence and our findings jointly support a bidirectional link between glymphatic dysfunction and E–I imbalance in the human epileptic cortex. In RE, the AH showed both reduced DTI‐ALPS index and elevated aperiodic exponent, suggesting glymphatic impairment alongside increased inhibitory dominance. It is notable that the inter‐hemispheric differences in these metrics were significantly correlated, indicating spatial coupling between impaired waste clearance and altered cortical excitability. Previous study has found that synchronized neuronal activity facilitates glymphatic flow by driving rhythmic ionic waves through the interstitial space.⁵ Disruption of this synchrony—commonly seen in epileptic tissue—reduces clearance efficiency. Conversely, impaired glymphatic clearance may disrupt neuronal homeostasis. The accumulation of phosphorylated tau (p‐tau) and serum proteins in perivascular regions has been implicated in altered neuronal synchrony, leading to increased excitability and axonal injury. ³² , ³³ , ³⁴ , ³⁵ In addition, impaired ISF drainage can result in glutamate accumulation, which is neurotoxic and has been shown to suppress inhibitory neurotransmitter synthesis, thereby promoting cortical hyperexcitability. ³⁶ Disrupted clearance of activity‐regulating ions such as potassium and calcium may further contribute to pathological excitability, particularly in epilepsy. ⁴ Together, these findings underscore a dynamic and reciprocal relationship between neuronal activity and GS function, with implications for neural stability and disease susceptibility.

Our findings further support the existing evidence linking brain atrophy to dysfunction in the GS, particularly as reflected by reduced DTI‐ALPS index. Li et al. recently demonstrated that reduced automated DTI‐ALPS indices are significantly correlated with widespread brain parenchymal and hippocampal atrophy in Parkinson's disease. ¹⁵ Furthermore, Hsiao et al. reported that a higher DTI‐ALPS index indicated a better cortical reserve in regions coincided with the default mode network. ³⁷ And a higher DTI‐ALPS index in older people could reflect a greater cortical volume reserve. This intrinsic relationship highlights the critical involvement of the GS in neurodegenerative pathophysiology. Mechanistically, impaired glymphatic clearance is hypothesized to facilitate the accumulation of neurotoxic substances such as soluble amyloid beta (Aβ) and tau oligomers, which exacerbate neuronal damage and subsequently drive brain atrophy and neurodegeneration. ³⁸

The latest study reported that E–I balance is associated with age‐related variability in GMV. ³⁹ They observed a significant positive correlation with slope and GMV in rostral middle frontal gyrus, and a significant negative correlation with inferior temporal cortex. Mechanistically, the observed relationship between GMV and aperiodic slopes in inferior frontal gyrus can be interpreted through the lens of synaptic pruning and cortical maturation, whereas sensorimotor cortical development—as indexed by E–I balance—stabilizes during young adulthood. It is well known that temporal–frontal is the most common areas to be affected in patients' brain of RE. ⁹ In the atrophic cortical regions of RE, typical histopathological findings include neuronal loss and glial proliferation, ⁴⁰ which may disrupt the normal neurological activity leading to a steepening of aperiodic slopes. Future research should further examine relationships between aperiodic activity and brain structures.

4.1. Limitations of the study

Several limitations of the present study should be acknowledged. First, although the UH showed higher GMV than the AH, our study did not include an age‐matched healthy control group to determine whether UH volumes were within the normal range or reflected true compensatory hypertrophy. Therefore, the finding should be interpreted as a relative interhemispheric difference (UH > AH) rather than definitive evidence of compensatory increase. This will be addressed in future studies with appropriate control cohorts. Second, the spatial resolution of EEG was constrained by the limited number of recording channels, which restricted our ability to perform accurate source localization of epileptiform activity. Employing high‐density EEG or integrating EEG with advanced imaging modalities such as magnetoencephalography (MEG) or simultaneous EEG–functional MRI (fMRI) could enhance spatial resolution and improve localization accuracy in future investigations. Third, sedation was used transiently during MRI acquisition to minimize motion and was not intended as an experimental manipulation. Although we did not directly measure CSF dynamics and instead relied on the diffusion‐based ALPS index, anesthesia/sedation may still modulate glymphatic‐related processes and potentially influence ALPS estimates; however, such effects would be expected to act globally and thus affect both hemispheres similarly in our interhemispheric comparisons. Fourth, although hemispheric disorders encompass a variety of etiologies, our study focused exclusively on patients with RE as the disease model. Therefore, future research should explore whether similar glymphatic alterations and E–I imbalances are present across other hemispheric epilepsies to enhance the generalizability and clinical relevance of the results. Finally, the ALPS index may be confounded by structural alterations such as atrophy or demyelination, and future studies should integrate ALPS with complementary approaches (e.g., volumetric quantification of perivascular spaces, contrast‐enhanced MRI, or advanced diffusion models) to better disentangle glymphatic dysfunction from structural pathology.

5. CONCLUSION

Our findings establish a previously unrecognized coupling between glymphatic dysfunction and E–I imbalance in patients with epilepsy. These results reveal a glymphatic–neuronal interaction axis, supporting a model wherein impaired perivascular clearance contributes directly to inhibitory‐shifted cortical states. Given the emerging clinical significance of glymphatic dysfunction across neurological diseases, targeting glymphatic pathways may offer novel therapeutic opportunities to restore cortical excitability balance in epilepsy and related disorders.

AUTHOR CONTRIBUTIONS

C.F. and Y.Y. designed the methodology, performed the analysis, and wrote the original draft of the manuscript. G.L. and T.J. supervised and provided oversight of the research project. C.F., Y.Y., and P.G. conceptualized the research goals and provided methodological insights. P.G., K.L., and C.T. provided datasets for research. X.W., C.T., Y.J., Q.L., L.C., and all co‐authors listed above discussed the results, provided insights, and contributed to reviewing and editing of the manuscript.

CONFLICT OF INTEREST STATEMENT

The authors declare that they have no conflict of interest. We confirm that we have read the Journal's position on issues involved in ethical publication and affirm that this report is consistent with those guidelines.

Supporting information

Data S1.

Fu C, Yang Y, Gong P, Lv K, Liu Y, Tang C, et al. Glymphatic dysfunction couples with cortical excitation–inhibition imbalance in epilepsy: Evidence from Rasmussen encephalitis. Epilepsia. 2026;67:3563–3576. 10.1002/epi.70228

DATA AVAILABILITY STATEMENT

Anonymized data underlying these results will be made available upon request to any qualified investigator. Please email Guoming Luan.