Topographic divergence of atypical cortical asymmetry and atrophy patterns in temporal lobe epilepsy.

Temporal lobe epilepsy, a common drug-resistant epilepsy in adults, is primarily a limbic network disorder associated with predominant unilateral hippocampal pathology. Structural MRI has provided an in vivo window into whole-brain grey matter structural alterations in temporal lobe epilepsy relative to controls, by either mapping (i) atypical inter-hemispheric asymmetry; or (ii) regional atrophy. However, similarities and differences of both atypical asymmetry and regional atrophy measures have not been systematically investigated. Here, we addressed this gap using the multisite ENIGMA-Epilepsy dataset comprising MRI brain morphological measures in 732 temporal lobe epilepsy patients and 1418 healthy controls. We compared spatial distributions of grey matter asymmetry and atrophy in temporal lobe epilepsy, contextualized their topographies relative to spatial gradients in cortical microstructure and functional connectivity calculated using 207 healthy controls obtained from Human Connectome Project and an independent dataset containing 23 temporal lobe epilepsy patients and 53 healthy controls and examined clinical associations using machine learning. We identified a marked divergence in the spatial distribution of atypical inter-hemispheric asymmetry and regional atrophy mapping. The former revealed a temporo-limbic disease signature while the latter showed diffuse and bilateral patterns. Our findings were robust across individual sites and patients. Cortical atrophy was significantly correlated with disease duration and age at seizure onset, while degrees of asymmetry did not show a significant relationship to these clinical variables. Our findings highlight that the mapping of atypical inter-hemispheric asymmetry and regional atrophy tap into two complementary aspects of temporal lobe epilepsy-related pathology, with the former revealing primary substrates in ipsilateral limbic circuits and the latter capturing bilateral disease effects. These findings refine our notion of the neuropathology of temporal lobe epilepsy and may inform future discovery and validation of complementary MRI biomarkers in temporal lobe epilepsy.

Introduction

Temporal lobe epilepsy (TLE) is the most common drug-resistant epilepsy in adults. Its hallmark is pathology of mesiotemporal structures, notably the hippocampus, entorhinal cortex, amygdala and temporal pole.[1-5] The degree of atrophy in these regions correlates with the tendency to express epileptic activity.[6,7] Moreover, unilateral anteromesial resection leads to worthwhile improvement in approximately 90% of patients and long-term seizure freedom in more than 50%.[8-10]

MRI can identify the pathological substrate of TLE in vivo and lateralize and define the surgical target. Indeed, MRI has provided biomarker candidates for TLE diagnostics, prognostics and disease staging.[11-13] MRI analyses in TLE traditionally focus on manually or automatically delineating individual mesiotemporal structures, followed by (i) the analysis of inter-hemispheric grey matter asymmetry or (ii) the regional comparison of morphometric measures in patients relative to healthy controls. Studies focusing on mesiotemporal grey matter consistently reported atrophy and marked asymmetry, reaffirming that TLE is primarily a limbic disorder.[10,14-21]

With advancements and automation of image processing techniques, quantitative MRI analysis has been extended to the whole-brain level using volumetric analysis and voxel-based morphometry [20,22-27] as well as surface-based cortical thickness analysis.[28-32] These studies have mainly been cross-sectional regional comparisons between TLE and healthy controls, and explored patterns of inter-hemispheric asymmetry in TLE only sporadically. Whole-brain analyses often showed asymmetric mesiotemporal damage, and also revealed widespread and bilateral decreases in cortical grey matter outside the mesiotemporal lobe, with neither a limbic nor lateralized predominance.[31-35] Similar findings were confirmed by a multisite initiative aggregating and analysing brain morphometric measures in common epilepsies.[34,36]

Outside the mesiotemporal regions, the scarcity of asymmetry analyses precludes insights into how similar or different patterns of atypical structural asymmetry are relative to patterns of regional atrophy in TLE. Analysing both atrophy and asymmetry features could inform the development of individualized MRI biomarkers.[19,20,37,38] Moreover, comparing these patterns could clarify whether these reflect different disease processes. One emerging family of approaches stratifies cortical areas along spatial gradients of cortical microstructure and connectivity.[39-41] Cortical areas indeed show variable microstructural characteristics, often following sensory–fugal spatial gradients that relate to plasticity and neural excitability.[35,41-46] For example, paralimbic circuits differ from sensory networks by having an agranular architecture with only subtle laminar differentiation and relatively low myelin content, while sensorimotor areas have a marked layer IV and higher intracortical myelin.[35,47-50] Complementing these microstructural variations, recent work has shown gradients of functional connectivity running from sensorimotor networks towards heteromodal systems, notably the default mode network.[39] Contextualizing atrophy and atypical asymmetry patterns along these microstructural and functional connectivity gradients may shed light on potential anatomical determinants of cortical pathology in TLE.

We used the ENIGMA-Epilepsy dataset to map the topography of atypical inter-hemispheric asymmetry and regional atrophy in 732 individuals with TLE and 1418 healthy controls. Using a multisite mega-analysis, we systematically assessed the commonalities and divergences of these spatial patterns. We further contextualized findings with respect to microstructural and functional connectivity gradients, derived from parallel myelin-sensitive microstructural MRI and resting-state functional acquisitions,[39,41,51] obtained from both Human Connectome Project (HCP; 207 healthy controls) as well as a local cohort of TLE patients and controls (23 TLE and 53 healthy controls). We formulated the following hypotheses: (i) the spatial distribution of TLE-related cortical asymmetry and atrophy would differ, with the former being particularly temporo-limbic; and (ii) atypical asymmetry and atrophy maps would relate to cortical gradients, with the asymmetry map being more closely related to the primary temporo-limbic gradients derived from cortical microstructure. We also assessed whether inter-hemispheric asymmetry and regional atrophy mapping would show differential associations with clinical parameters, notably effects of disease duration and age of onset. In addition to benefiting from the high power of ENIGMA-Epilepsy, we validated the consistency of our findings at the level of single patients and individual sites.

Materials and methods

Participants

We analysed 2150 T1-weighted MRI datasets from 732 patients with TLE and confirmed/suspected mesiotemporal sclerosis (55% females, mean age ± SD = 38.56 ± 10.61 years; 391/341 left/right TLE) and 1418 healthy controls (55% females, mean age ± SD = 33.76 ± 10.54 years) obtained from 19 different sites via the Epilepsy Working Group of ENIGMA[34,36,52] (Table 1). Individuals with epilepsy were diagnosed by epilepsy specialists at each centre according to classifications of the International League Against Epilepsy.[53] TLE patients were diagnosed based on electroclinical and neuroimaging findings. Participants with a primary progressive disease (e.g. Rasmussen’s encephalitis), visible malformations of cortical development, or prior neurosurgery were excluded. For each site, local institutional review boards and ethics committees approved each included cohort study and written informed consent was provided according to local requirements.

Table 1

Demographic information of individuals with TLE and site-matched controls

Information	ENIGMA-Epilepsy		HCP (HC)	MICs
	TLE	HC		TLE	HC
n	732	1418	207	23	53
Age, years	38.56 ± 10.61	33.76 ± 10.54	28.73 ± 3.73	37.29 ± 11.96	30.84 ± 7.59
Sex, male: female	329:403	643:775	83:124	11:12	33:20
Age at onset, years	16.07 ± 12.27	N/A	N/A	21.59 ± 11.65	N/A
Side of focus, left/right	391/341	N/A	N/A	15/7 (1 bilateral)	N/A
Duration of illness, years	22.74 ± 14.06[a]	N/A	N/A	15.82 ± 12.45	N/A

Means and SDs are reported. HC = healthy control; N/A = not available.

Information available in 695 TLE patients.

Gradients were derived from two independent cohorts containing healthy controls and patients with TLE: (i) A sample of 207 unrelated healthy young adults (60% females, mean age ± SD = 28.73 ± 3.73 years) from the HCP[54]; and (ii) a sample of 53 healthy controls (38% females, mean age ± SD = 30.84 ± 7.59 years) and 23 TLE patients (52% females, mean age ± SD = 37.29 ± 11.96 years) from our local site at the MNI (microstructure-informed connectomics; MICs). All participants gave written and informed consent.

Data preprocessing

ENIGMA data

Participants underwent T1-weighted scans at each of the 19 centres, with acquisition protocols detailed elsewhere.[34] Imaging data were processed by each centre through the standard ENIGMA workflow described in Supplementary material and Supplementary Fig. 1.

HCP data

T1- and T2-weighted, as well as resting-state functional MRI (rs-fMRI) data, were obtained using a Siemens Skyra 3 T at Washington University.[54] The T1-weighted images were acquired using a magnetization-prepared rapid gradient echo sequence [repetition time (TR) = 2400 ms; echo time (TE) = 2.14 ms; inversion time (TI) = 1000 ms; flip angle = 8°; field of view (FOV) = 224 × 224 mm2; voxel size = 0.7 mm isotropic; and number of slices = 256]. T2-weighted data were obtained with a T2-SPACE sequence (TR = 3200 ms; TE = 565 ms; flip angle = variable; FOV = 224 × 224 mm2; voxel size = 0.7 mm isotropic; and number of slices = 256). The rs-fMRI data were collected using a gradient-echo echo-planar imaging sequence (TR = 720 ms; TE = 33.1 ms; flip angle = 52°; FOV = 208 × 180 mm2; voxel size = 2 mm isotropic; number of slices = 72; and number of volumes = 1200 per time series). During the rs-fMRI scan, participants were instructed to keep their eyes open, looking at a fixation cross. Two sessions of rs-fMRI data were acquired; each contained data of left-to-right and right-to-left phase-encoded directions, providing up to four time series per participant. HCP data underwent minimal preprocessing pipelines using FSL, FreeSurfer and Workbench,[55-57] briefly summarized in Supplementary material and Supplementary Fig. 1.

MICs data

Data were acquired on a Siemens Prisma 3T scanner. Acquisition parameters were similar to the HCP dataset (T1-weighted: TR = 2300 ms; TE = 3.14 ms; TI = 900 ms; flip angle = 9°; FOV = 256 × 180 mm2; voxel size = 0.8 mm isotropic; and number of slices = 320; quantitative T1: same as T1-weighted except for TR = 5000 ms and TE = 2.9 ms; TI = 940 ms; flip angle 1 = 4°; flip angle 2 = 5°; rs-fMRI: TR = 600 ms; TE = 30 ms; flip angle = 52°; FOV = 240 × 240 mm2; voxel size = 3 mm isotropic; number of slices = 48; and number of volumes = 700). MICs data were preprocessed using micapipe (https://github.com/MICA-MNI/micapipe; last accessed February 5, 2022), which integrates AFNI, FSL, FreeSurfer, ANTs and Workbench.[55-59] Details are described in Supplementary material and Supplementary Fig. 1.

Atypical inter-hemispheric cortical asymmetry and regional atrophy

We calculated inter-hemispheric asymmetry of cortical thickness using the following formula: AI = (ipsi − contra) / |(ipsi + contra)/2|,[19,60,61] where AI is asymmetry index and ipsi and contra are the cortical thickness of ipsilateral and contralateral areas, respectively. The asymmetry index was z-normalized relative to site-matched pooled controls and sorted into ipsilateral/contralateral to the focus.[62] It was then harmonized across different sites by adjusting for age, sex and intracranial volume using ComBat, a batch-effect correction tool that uses a Bayesian framework to improve the stability of the parameter estimates.[63,64] We compared the harmonized asymmetry index between individuals with TLE and controls using a general linear model implemented in SurfStat.[65] Multiple comparisons across brain regions were corrected using the FDR procedure.[66] In addition to parcel-wise analysis, we stratified asymmetry measures according to seven intrinsic functional communities (visual, somatomotor, dorsal attention, ventral attention, limbic, frontoparietal and default mode)[67] and lobes (frontal, parietal, temporal, occipital, cingulate and insular cortex). In addition to the atypical asymmetry index, we assessed cortical atrophy in TLE patients relative to controls. Cortical thickness measures were z-normalized, flipped hemispheres of right TLE and harmonized as for the asymmetry index. We compared the harmonized cortical thickness between the groups and the findings were multiple-comparison corrected using FDR[66] as well as stratified according to functional communities and lobes.

Association to gradients of cortical microstructure and function

We assessed topographic underpinnings of TLE-related asymmetry and atrophy through spatial correlation analysis with microstructural and functional gradients, the principal eigenvectors explaining spatial shifts in microstructural similarity and functional connectivity.[39,41] Gradients were defined using two alternative datasets, either based on both the HCP (i.e. healthy controls) or based on the MICs (i.e. healthy controls and TLE patients), using BrainSpace (https://github.com/MICA-MNI/BrainSpace; last accessed 5 February 2022).[51] Specifically, we calculated a parcel-to-parcel affinity matrix for each feature using a normalized angle kernel considering the top 10% entries for each parcel. As in prior work,[39,45,51,68-74] we opted for diffusion map embedding,[75] a non-linear technique that is robust to noise and computationally efficient.[76,77] It is controlled by two parameters, α and t, where α controls the influence of the density of sampling points on the manifold (α = 0, maximal influence; α = 1, no influence) and t scales eigenvalues of the diffusion operator. The parameters were set as α = 0.5 and t = 0 to retain the global relations between data points in the embedded space, following prior applications.[39,41,45,51,69,78] We examined associations of the estimated gradients with cortical asymmetry in a single hemisphere and atrophy patterns in both hemispheres via linear correlations, where significance was determined using 1000 non-parametric spin tests that account for spatial autocorrelation[79] implemented in the ENIGMA Toolbox.[80]

Consistency mapping across sites and individuals

We assessed the robustness of our findings within a probabilistic framework at the single site and subject level. The consistency across sites was measured by calculating linear correlations between epilepsy-related asymmetry and atrophy findings and gradients for each site. For individual-level consistency, we counted how many participants are comprised within a specific threshold (i.e. z < −2). The counts were divided by the number of participants to obtain a probability map. Thus, the consistency probability indicates that the top N% patients showed extreme asymmetry or cortical atrophy measures in a given region. The consistency probability was correlated with microstructural and functional gradients, with 1000 non-parametric spin tests.[79,80]

Associations with clinical variables

We associated clinical variables of duration and onset of epilepsy with atypical asymmetry index and cortical atrophy using supervised machine learning. We utilized 5-fold nested cross-validation[76,81-83] with least absolute shrinkage and selection operator (LASSO) regression.[84] We split the dataset into training (4/5) and test (1/5) partitions, and each training partition was further split into inner training and testing folds using another 5-fold cross-validation. Within the inner fold, LASSO finds a set of non-redundant features (i.e. atypical asymmetry index or cortical atrophy of brain regions) that could explain the dependent variable (i.e. disease duration or onset age). Using a linear regression, we predicted the clinical variables of inner fold test data using the features of the selected brain regions. The model with best accuracy (i.e. minimum mean absolute error, MAE) across the inner folds was applied to the test partition of the outer fold and the clinical variables of outer fold test data were predicted. We repeated this procedure 100 times with different training and test partitions to avoid subject selection bias. We assessed the prediction accuracy by calculating linear correlations between the actual and predicted clinical variables with their 95% confidence interval across 100 repetitions, as well as MAE. The significance of the correlation between actual and predicted values was assessed using 1000 permutation tests by randomly shuffling participant indices. A null distribution was constructed, and it was deemed significant if the real correlation value did not belong to 95% of the distribution (two-tailed P < 0.05). We compared our model with the baseline model (i.e. predicted clinical variable = mean(training set clinical variable)) and improved prediction performance of our model was assessed using Meng’s z-test.[85] To assess whether the frequency of the selected brain regions derived from LASSO regression across cross-validations and repetitions is related to microstructural and functional gradients, we calculated spatial correlations between cortex-wide probability distributions and each of the gradients. Significance was assessed using 1000 spin tests.[79,80] As a post-hoc analysis, we correlated the cortical features of the highly probable regions (selected probability > 0.5) and clinical variables, and the significance was calculated based on 1000 permutation tests by randomly shuffling participant indices.

Sensitivity analysis

Left and right TLE

To assess whether left and right TLE show consistent results, we repeated assessing atypical cortical asymmetry and atrophy and correlating the effects with gradients for separate left and right TLE subgroups. We furthermore assessed cortical atrophy in individuals with left and right TLE for each hemisphere. We conducted paired t-tests to compare cortical atrophy between hemispheres within left or right TLE, and two-sample t-tests between left and right TLE. Multiple comparisons were corrected using FDR.[66]

Different density of connectivity matrix

In our main analysis, we estimated microstructural and functional gradients using connectivity matrices with 10% density. We repeated generating the gradients from connectivity matrices with different densities (20%, 30%, 40%, 50%) and correlated with atypical cortical asymmetry and atrophy patterns.

Gradients generated using local dataset

We generated microstructural and functional gradients using a combined dataset of healthy individuals and patients with TLE, to assess consistency of the topographic relationships between TLE-related asymmetry/atrophy and cortical gradients.

Volumetric analysis

We additionally assessed atypical inter-hemispheric asymmetry and regional atrophy patterns of six subcortical regions (amygdala, caudate, nucleus accumbens, pallidum, putamen, thalamus), as well as the hippocampus, defined using the Desikan–Killiany atlas.[86] We estimated the volume of each region, calculated asymmetry index,[19,60,61]z-normalized both asymmetry index and volume of TLE patients relative to controls, flipped hemispheres in right TLE patients and harmonized data across different sites by adjusting for age, sex and intracranial volume using ComBat.[63,64] We compared asymmetry and regional volume between individuals with TLE and controls using a general linear model.[65] Next, we assessed the robustness of atypical asymmetry and atrophy by calculating consistency probability. Lastly, we performed the prediction analysis by considering both cortical thickness and subcortical/hippocampal volume measures using LASSO regression [84] with five fold nested cross-validation.[76,81-83] The prediction procedure was repeated 100 times with different training and test datasets and the performance was measured using linear correlations between the actual and predicted clinical variables with their 95% confidence interval, as well as MAE. We compared our model with the baseline model, and assessed improvement of the prediction performance using Meng’s z-test.[85]

Data availability

The data that support the findings of this study are available on request from the corresponding author. The data are not all publicly available in a repository as they contain information that could compromise the privacy of research participants.

Although there are data-sharing restrictions imposed by (i) ethical review boards of the participating sites, and consent documents; (ii) national and trans-national data sharing law, such as General Data Protection Regulation (GDPR); and (iii) institutional processes, some of which require a signed Material Transfer Agreements (MTA) for limited and predefined data use, we welcome sharing data with researchers, requiring only that they submit an analysis plan for a secondary project to the leading team of the Working Group (http://enigma.ini.usc.edu; last accessed 5 February 2022). Once this analysis plan is approved, access to the relevant data will be provided contingent on data availability and local PI approval and compliance with all supervening regulations. If applicable, distribution of analysis protocols to sites will be facilitated.

Results

Atypical inter-hemispheric asymmetry patterns differ from regional cortical atrophy in TLE

We found significant deviations in inter-hemispheric asymmetry in TLE relative to controls, especially in lateral and medial temporal cortex, as well as precuneus, with ipsilateral regions being atypically smaller than contralateral regions (P < 0.05; Fig. 1A). Stratifying effects according to intrinsic functional communities,[67] highest deviations in asymmetry were observed in the limbic network followed by default mode and somatomotor networks (Fig. 1B). Lobar analysis identified most marked degrees of atypical asymmetry in the temporal lobes. Asymmetry patterns of TLE were markedly different from regional differences in bilateral cortical thickness. Indeed, comparing cortical thickness between TLE and healthy controls showed widespread and bilateral cortical thickness reductions in TLE, with strongest effects in precentral, paracentral and superior temporal regions (P < 0.05; Fig. 1A). Findings were distributed across somatomotor, dorsal attention and visual networks (Fig. 1B). Similarly, lobar stratification pointed to multilobar effects, most marked in frontal, parietal and occipital lobes in both hemispheres. Notably, spatial correlations between atypical asymmetry and atrophy patterns in a single hemisphere were very low and did not surpass null models with similar autocorrelation (r = 0.05, P = 0.27).[79]

Figure 1

Topography of atypical cortical asymmetry and atrophy patterns in TLE. (A) Atypical inter-hemispheric asymmetry of cortical thickness and regional cortical atrophy between individuals with TLE relative to controls, calculated using ENIGMA-Epilepsy dataset. Blue regions indicate significant ipsilateral versus contralateral cortical thickness asymmetry/atrophy in TLE relative to controls. Patient hemispheres are sorted into ipsilateral/contralateral to the seizure focus. (B) Effects (i.e. asymmetry index and cortical thickness) are stratified according to seven intrinsic functional communities[67] and major lobes. (C) Associations between epilepsy-related findings and microstructural/functional gradients calculated using HCP dataset. Cortex-wide microstructural profile similarity matrix and scree plot describing connectome variance after identification of principal eigenvectors are shown. The first principal eigenvector (microstructural gradient) is shown on the cortical surface. Spatial correlations between the principal microstructural gradient and TLE-related effects (i.e. atypical cortical asymmetry and atrophy) are reported with scatter plots. (D) Identical analysis to C but based on functional gradients. Cing = cingulate; DAN = dorsal attention network; DMN = default mode network; FPN = frontoparietal control network; Front = frontal; Ins = insular; LBN = limbic network; Occ = occipital; Par = parietal; SMN = somatomotor network; Temp = temporal; VAN = ventral attention network; VN = visual network.

A diverging topographic landscape of TLE-related atypical asymmetry and atrophy

Next, we assessed spatial associations of epilepsy-related findings with microstructural and functional gradients. The microstructural gradient depicted a continuous differentiation of cortical features between sensory and limbic areas, and was negatively correlated with atypical asymmetry index (r = −0.13, P = 0.03), reflecting elevated atypical asymmetry in temporo-limbic cortices in TLE (Fig. 1C). On the other hand, it was positively and markedly correlated with regional atrophy in TLE (r = 0.72; P < 0.001; Fig. 1C). The difference between these two correlations was significant (P < 0.001; Meng’s z-test),[85] indicating a dissociation of atypical asymmetry and atrophy patterns with respect to the primary microstructural gradient. The functional gradient differentiated primary sensory from transmodal regions, and did not show a significant association with atypical inter-hemispheric asymmetry in TLE (r = −0.10, P = 0.12), but a low-to-moderate positive association with regional atrophy (r = 0.31, P < 0.001; Fig. 1D).

Consistency across sites and individuals

We confirmed the above topographic divergence across individual sites (Fig. 2A) by correlating microstructural and functional gradients with atypical asymmetry and regional atrophy in TLE for each site separately (Fig. 2B). These follow-up analyses confirmed our main findings (Fig. 1C and D) that showed dissociation between atypical asymmetry and atrophy patterns. Multisite analyses were expanded by assessing consistency at the level of individual patients (Fig. 2C). Prevalent atypical asymmetry was confirmed in somatomotor and limbic regions (Fig. 2D) and spatial patterns revealed significant associations only with the microstructural gradient (r = 0.13/−0.02, P = 0.04/0.42 for microstructural/functional gradients; Fig. 2E). The consistency probability of regional cortical atrophy showed higher consistency in sensory, precuneus and temporal regions, and it showed significant negative correlations with both gradients (r = −0.29/−0.30, P < 0.001/<0.001), supporting patient-level consistency.

Figure 2

Consistency of atypical cortical asymmetry and atrophy. (A) World map of data acquisition sites. (B) Spatial correlations between topographic gradients and atypical cortical asymmetry/atrophy patterns of all sites. (C) Schema describing the computation of patient-wise consistency probability. The number of patients with large deviations of cortical features (i.e. atypical inter-hemispheric asymmetry or regional cortical atrophy) was counted. (D) Consistency probability of atypical cortical asymmetry and atrophy. (E) Spatial correlations between consistency probability and topographic gradients.

Utilizing supervised machine learning, we probed associations of both atypical inter-hemispheric asymmetry and regional atrophy with disease duration and age at seizure onset. While cortical atrophy significantly predicted the clinical variables outperforming the baseline model (disease duration: mean ± SD r = 0.26 ± 0.02, MAE = 11.38 ± 0.10, Meng’s z-test P < 0.001; age at seizure onset: r = 0.17 ± 0.02, MAE = 9.91 ± 0.08, Meng’s z-test P = 0.01), atypical asymmetry did not (disease duration: Meng’s z-test P = 0.27; age at seizure onset: Meng’s z-test P = 0.20; Fig. 3A,D). Considering cortical atrophy, sensorimotor, medial/lateral temporal, and precuneus were frequently selected across cross-validations as salient features for the prediction for disease duration (Fig. 3A), and sensorimotor and limbic regions for age at seizure onset (Fig. 3D). As in the main analyses, we observed significant associations of the selected probability with connectome gradients (disease duration: r = −0.27/−0.34 P = 0.002/<0.001 for microstructural/functional gradients; age at seizure onset: r = −0.25/0.03 P < 0.001/0.35; Fig. 3B and E). Associations in highly probable regions (selected probability > 0.5) were negative, i.e. disease duration/age at seizure onset associated with cortical thickness reductions (r = −0.30/−0.21, permutation test P < 0.001/<0.001; Fig. 3C and F).

Figure 3

Associations between cortical features and clinical variables. (A) Probability of regions being selected across 5-fold nested cross-validation and 100 repetitions for predicting duration of epilepsy using atypical asymmetry index (left) and regional atrophy (right). Correlations between actual and predicted values of epilepsy duration are reported in the scatter plots. Black lines indicate the mean correlation and grey lines represent the 95% CI for 100 iterations with different training/test datasets. (B) Linear correlations between gradients and selected probability. (C) Spatial correlations between duration of epilepsy and atypical asymmetry index (left), as well as cortical atrophy (right) in highly probable (selected probability > 0.5) regions. (D–F) Identical analysis to A–C, but with respect to age at seizure onset. MAE = mean absolute error.

Sensitivity analyses

Several analyses supported robustness of our main findings.

We repeated the above analyses in left and right TLE separately. While the degree of asymmetry was stronger in left than right TLE, findings were overall consistent (Supplementary Fig. 2). In the temporal lobe, while both left and right TLE patients showed more pronounced ipsilateral cortical atrophy, ipsilateral atrophy in left TLE was more marked than in right TLE, while contralateral atrophy was stronger in the latter subgroup (Supplementary Fig. 2E and F).

We repeated our analyses by varying the thresholds of microstructural and functional connectivity matrices across different densities (20%, 30%, 40%, 50%). Gradients and their associations with inter-hemispheric asymmetry, as well as regional atrophy, remained consistent (Supplementary Fig. 3).

We also repeated the analysis after building gradients using a different dataset comprising both healthy individuals and patients with TLE. Microstructural and functional gradients were highly similar to those from the HCP dataset, and topographic associations between TLE-related asymmetry/atrophy and cortical gradients remained consistent when using gradients based on a combined dataset of healthy individuals and patients with TLE (Supplementary Fig. 4).

Volumetry of subcortical regions and the hippocampus

We also studied the volume of subcortical structures as well as the hippocampus. While atypical asymmetry and atrophy patterns both supported marked ipsilateral hippocampal effects (P < 0.05; Supplementary Fig. 5A), spatial correlations between atypical inter-hemispheric asymmetry and regional atrophy were moderate and not significant (r = 0.51, P = 0.06). As for the cortical thickness-based results, these findings were consistent across individual subjects (Supplementary Fig. 5B). When we considered both cortical thickness and subcortical/hippocampal volume, we were able to confirm our initial results in that atrophy but not atypical asymmetry related to age at seizure onset, while disease duration was significantly associated with both measures, outperforming the baseline model (atypical asymmetry: P = 0.004 for disease duration, P = 0.14 for age at seizure onset; atrophy: P < 0.001 for both disease duration and age at seizure onset; Meng’s z-test; Supplementary Fig. 5C).

Discussion

Together with the multisite ENIGMA-Epilepsy initiative,[34,36,52,87] we investigated patterns of atypical inter-hemispheric asymmetry of cortical thickness and cross-sectional regional atrophy in a large sample of TLE patients and healthy controls. In particular, we studied whether (i) the spatial distribution of atypical inter-hemispheric asymmetry differed from patterns of regional atrophy in TLE relative to controls; (ii) these patterns follow different topographic principles of cortical organization, particularly with respect to microstructural and functional gradients; and (iii) these effects showed a differential association to effects of epilepsy duration and age of onset. We found that atypical inter-hemispheric asymmetry analysis and regional atrophy mapping provide complementary insights into the pathology of TLE in vivo, with atypical asymmetry showing an ipsilateral limbic signature, while cross-sectional cortical thickness mapping indicated widespread and bilateral atrophy in TLE. Atypical asymmetry and atrophy patterns of the cortex were also differentially associated with microstructural and functional gradients representing core axes of cortical organization,[39-41,44,88] supporting a topographic divergence of these two characterizations of TLE-related pathology. Findings were consistent across different sites and participants, corroborating generalizability. While cortical atrophy was correlated with disease duration and age at seizure onset, atypical asymmetry did not show an association to these variables. Collectively, our study underscores complementarity of atypical asymmetry and atrophy mapping for in vivo pathology mapping, which will be relevant for future imaging biomarker discovery and validation efforts.

In managing TLE patients, preoperative lateralization of temporal lobe pathology is key to define the surgical target and often relies on the qualitative visual assessment of inter-hemispheric asymmetry. Quantitative imaging analyses in clinical and research settings can be geared towards the identification of asymmetry, and several prior studies have systematically investigated between-hemisphere differences in grey matter morphological measures in TLE. Most of these studies have focused on the asymmetry of the hippocampus and adjacent mesiotemporal structures, suggesting marked limbic structural asymmetry in TLE.[19,21,37,89] Asymmetry analysis has several benefits, including the ability to use a given patient as their own baseline while controlling for corresponding measures in controls. However, the field lacks systematic analyses of asymmetry, particularly outside the mesiotemporal region. There have been no quantitative comparisons of atypical inter-hemispheric asymmetry with maps of cross-sectional regional atrophy mapping, in which measures in patients with TLE are compared to groups of healthy controls. When carried out in structures of the limbic system, atrophy mapping also reveals structural compromise in TLE compared to healthy controls, with variable degrees of asymmetry ranging from relatively ipsilateral to rather bilateral depending on the TLE subgroups.[10,15,19] The advent of automated morphometric analysis has resulted in a predominance of studies focusing on cross-sectional regional thickness comparisons, and relatively few large-scale analyses have assessed the topography of atypical cortical thickness asymmetry patterns in TLE.[31-35,38] Notably, although atypical asymmetry and atrophy are sometimes used interchangeably in the neuroimaging literature of TLE as in vivo indices of pathology, our findings pointed to differences in the topography of atypical cortical asymmetry and regional patterns of cross-sectional atrophy in TLE. Spatial correlation analysis confirmed this, failing to identify an association between these spatial patterns after accounting for spatial autocorrelations. Atypical asymmetry in TLE followed a more specific paralimbic topography with maximal effects in the mesiotemporal lobe, in line with the classical conceptualizations of TLE as a limbic network disorder.[1-3,90] On the other hand, in line with prior single site analyses[28-31] and recent ENIGMA-Epilepsy studies,[34,36,52] regional cortical atrophy mapping confirmed ipsilateral mesiotemporal atrophy in TLE, as well as widespread and bilateral effects outside paralimbic cortical areas. Findings were consistent in both left and right TLE patients. Thus, and despite both left and right TLE groups potentially showing different structural compromise,[21,34,62,91-94] findings overall suggest a similar divergence of atrophy and asymmetry patterns irrespective of seizure focus lateralization.

Our findings were further contextualized by quantifying the alignment of asymmetry and atrophy patterns along microstructural and functional gradients.[39-41] Cortical microstructural gradients place sensorimotor cortices with strong laminar differentiation and high myelin content at one end and paralimbic regions with subtle myelination, low laminar differentiation and increased synaptic densities at the other end.[41,72] Microstructural gradients, preserved across species,[41,72,95] follow canonical models of sensory–fugal cortical hierarchies[88] and capture inter-regional variations in heritability and plasticity.[96] While also starting at sensorimotor systems, the principal functional gradient radiates towards transmodal networks, such as the default mode and frontoparietal systems, and not the paralimbic cortices.[39] This divergence between microstructural and functional gradients may relate to less tethering of phylogenetically more recent association networks, such as the default mode network, from underlying signalling molecules[97] and may more closely reflect macroscale functional organization.[67] Spatial correlation analyses supported the dissociation of atypical cortical asymmetry and atrophy patterns with respect to microstructural gradients, where we observed increasing degrees of asymmetry towards the temporo-limbic anchor of the microstructural gradient, while atrophy patterns increased towards primary sensorimotor and unimodal association areas. While confirming stronger effects towards sensorimotor anchors in the case of atrophy patterns, functional gradient associations were less conclusive about atypical asymmetry, indirectly underscoring the paralimbic pattern of the latter. Furthermore, these findings may indicate that cortical morphological changes are better captured by microstructural than by functional hierarchies, a finding echoing prior associations between intracortical cellular-synaptic factors and measures of cortical thickness.[98-101] Both cortical thickness and microstructural gradients were derived from structural MRI and thickness and intracortical myelin are largely related to cortical cytoarchitecture,[45,102] which could indeed explain a strong correlation between atrophy and the microstructural gradient. It is, nevertheless, important to point out that cortical thickness and intracortical microstructural measures were calculated using different approaches: (i) cortical thickness was measured as the distance between pial and white matter cortical interfaces from T1-weighted data, tapping into overall cortical morphology; while (ii) the microstructural gradient was derived from depth-dependent intracortical intensity profiles based on the ratio of T1- and T2-weighted data from the HCP dataset (and from quantitative T1 relaxometry data in the case of the local MICs dataset). Prior research has furthermore shown that cortical morphology as indexed by cortical thickness measures as well as internal cortical microstructure reflects likely complementary aspects of healthy and diseased brain organization. For example, age-related effects on cortical thinning and myelination do not occur in parallel, but rather in a different spatial distribution during typical development[45,102] as well as lifespan.[103-106] In TLE, it has been shown that changes in intracortical microstructure based on quantitative T1 relaxometry occur above and beyond changes in MRI-based cortical thickness, suggesting that potentially different biological and pathological processes drive changes in morphology and microstructure in the condition.[107]

Big data initiatives such as ENIGMA-Epilepsy offer increased sensitivity to identify disease-related patterns of structural compromise. Extending from initial meta-analysis efforts,[34] it is furthermore possible to assess consistency of findings at the single site and individual patient levels. Here, we observed that the dissociation between atypical cortical asymmetry and atrophy remained consistent when we considered individual sites separately, and to some degree also at the level of individual participants. Using machine learning, we associated cortex-wide morphological data with clinical variables and showed inter-individual differences in cortical atrophy associated with disease duration and age at seizure onset. Associations were primarily driven by primary regions in sensorimotor cortex, together with temporal cortices and the precuneus. Unlike cortical thickness, atypical asymmetry patterns were not significantly associated with these clinical variables. These divergent clinical associations suggest that atypical inter-hemispheric asymmetry and regional cortical atrophy potentially reflect different TLE pathological processes, with asymmetry being more specifically related to an initial insult of the limbic circuitry. Alternatively, patterns of TLE-related atrophy in widespread and bilateral cortical territories had apparent progressive effects. The latter finding is consistent with prior cross-sectional, longitudinal and meta-analytic findings assessing disease progression effects in TLE.[29,32,91,108-111] This effect may relate to ongoing seizures, as supported by prior data showing associations to seizure frequency,[91,110,111] as well as from anti-epileptic drug treatment.[112,113] Moreover, drug-resistant patients are at increased risk for mood disorders and psychosocial challenges,[114] which may furthermore adversely impact brain health.

We found that measures of atypical asymmetry and atrophy provide complementary windows into structural compromise in TLE, a finding also supported by the differential relationships to cortical topographic gradients and diverging associations to clinical parameters. Our findings advance our understanding of large-scale pathology in TLE and may direct future discovery and validation of clinically useful neuroimaging biomarkers.

Click here for additional data file.