Atypical intrinsic neural timescales in temporal lobe epilepsy

Objective Temporal lobe epilepsy (TLE) is the most common drug-resistant epilepsy in adults. Here, we aimed to profile local neural function in TLE in vivo, building on prior evidence that has identified widespread structural alterations. Using multimodal MRI, we mapped intrinsic neural timescales (INT) at rest, examined associations to TLE-related structural compromise, and evaluated the clinical utility of INT. Methods We studied 46 TLE patients and 44 healthy controls from two independent sites, and mapped INT changes in patients relative to controls across hippocampal, subcortical, and neocortical regions. We examined region-specific associations to structural alterations and explored effects of age and epilepsy duration. A supervised machine learning paradigm assessed utility of INT for classifying patients-vs-controls and seizure focus lateralization. Results Relative to controls, TLE showed marked INT reductions across multiple regions bilaterally, indexing faster changing resting activity, with strongest effects in ipsilateral medial and lateral temporal regions, and sensorimotor cortices. Findings were consistent in each site and robust, albeit with reduced effect sizes, when correcting for structural alterations. TLE-related INT reductions increased with advancing disease duration, yet findings differed from aging effects seen in controls. Classifiers based on INT distinguished patients-vs-controls (balanced accuracy, 5-fold: 76±2.65%; cross-site, 72-83%) and lateralized the focus in TLE (balanced accuracy, 5-fold: 96±2.10%; cross-site, 95-97%) with high accuracies and generalization. Conclusions Our findings robustly demonstrate atypical macroscale function in TLE in a topography that extends beyond mesiotemporal epicenters. INT measurements can assist in TLE diagnosis, seizure focus lateralization, and monitoring of disease progression, which suggests clinical utility.


Introduction
Temporal lobe epilepsy (TLE) is a common drug-resistant epilepsy in adults. Traditionally considered as a 'focal' epilepsy with hippocampal pathology as its hallmark 1 , mounting histological and neuroimaging work indicates large-scale reorganization of brain structure. Structural and diffusion magnetic resonance imaging (MRI) studies, in particular, have revealed gray and white matter alterations across multiple cortical and subcortical areas 2, 3 , supporting the notion that TLE is a network disorder 4,5 .
Beyond the identified structural alterations, TLE has long been recognized to impact brain function 6,7 . Restingstate functional MRI (rs-fMRI) is increasingly used to interrogate neural function in health and disease, in light of its high spatial-resolution and full-brain coverage. Prior rs-fMRI studies in TLE identified atypical functional connectivity relative to controls, with altered inter-regional communication within circuits of temporal cortices, as well as between temporal and extra-temporal regions [8][9][10] . These findings are complemented by connectome-level investigations showing whole-brain functional reorganization 11,12 . In the literature focusing on functional imbalances in TLE, however, notably few studies have characterized intrinsic function at a regional level. Notably, and beyond the spatial localization of atypical function and connectivity, changes in functional signaling can also be characterized across temporal scales, a marker sensitive to neural processing 13 . One marker increasingly used in prior rs-fMRI and electrophysiology studies in healthy individuals is the measurement of intrinsic neural timescales (INT), which are sensitive to the shape of the long-range temporal autocorrelation of physiological signals 14,15 . As such, it helps to differentiate faster and slower processing. Moreover, studies have emphasized that INT vary topographically and recapitulate functional zones, with heteromodal association cortices implicated in higher order signal integration longer timescales than sensory/motor and unimodal cortices [14][15][16] . Notably, an increasing body of studies conducted in non-human primates as well as healthy humans has furthermore shown systematic associations between INT and region-to-region variations in brain morphology and microarchitecture 15,17 , suggesting that INT could serve as a local functional index anchored in foundational models of brain microcircuit organization. As imbalances in connectivity as well as microcircuit organization are recognized in TLE 10,18 , the systematic mapping of INT changes at a regional level may inform the development of patient-specific markers of atypical brain function.
Here, we mapped the topography of INT in TLE patients and examined alterations relative to healthy individuals. Aggregating data across two independent sites, we opted for a multimodal MRI acquisition and data processing paradigm data that allowed us to profile the topography of INT changes in TLE-vs-controls along hippocampal, neocortical, and subcortical regions. Moreover, integration of the rs-fMRI measures with structural and diffusion MRI data allowed us to explore whether INT alterations in patients relate to brain structure and fiber microstructure in TLE 10,19 . We also explored associations between INT and age and epilepsy duration, to provide a functional perspective on the notion that TLE may be associated with atypical aging and cumulative disease effects. Finally, we assessed whether brain-wide INT parametrization can inform patientvs-control classification and seizure focus lateralization, leveraging a supervised learning paradigm that implicated both cross-validation and cross-site generalizability assessments.

Participants
We studied people with TLE and healthy individuals who were aggregated from two independent sites: a) the Considering the TLE patients, we observed no between-site differences with respect to sex ( 2 = 0.12, p = 0.73), the proportion of patients with a left-/right-sided seizure focus ( 2 = 1.97, p = 0.16), and epilepsy duration (t = 0.35; p = 0.73). On the other hand, there were differences in age (t = -2.05, p = 0.05) and age at seizure onset (t = -2.95, p = 0.005).  29 . Preprocessing included discarding the first five volumes, reorientation, skull stripping, motion and distortion correction, removal of nuisance signals and artefacts using an ICA-FIX classifier, and regression of automatically identified, high-motion timepoints. A boundary-based registration mapped functional timeseries to each participant's cortical surface 30 . In parallel, subject-specific subcortical and hippocampal segmentations were non-linearly registered to each participant's native fMRI space.
Functional data were mapped to participant-specific cortical surfaces, then registered to the hemisphere-matched Conte69 surface template, and smoothed using a 10mm full-width-at-half-maximum kernel.
Smoothed data were downsampled to a template with 10,000 vertices for computational efficiency.
Hippocampal and subcortical timeseries were averaged within the corresponding co-registered segmentations.

Intrinsic neural timescale maps
We estimated INT values at each cortical vertex and in each subcortical region and the hippocampus following prior work 31,32 . The autocorrelation function (ACF) of the rs-fMRI timeseries for a given measurement point was calculated using the following formula: For a given region, denotes the rs-fMRI signal, $ is the mean signal across timepoints, is the time lag (time bin = TR), and is the number of timepoints. Then, INT was calculated as the sum of values in the initial positive period: where TR is the repetition time of the fMRI signal and N is the lag directly preceding the first negative ACF value. Here, multiplying the obtained sum of ACF values by TR aimed to adjust for differences in the temporal resolution of the rs-fMRI signal. The procedure was repeated for all vertices, subcortical structures and bilateral hippocampi, yielding a whole-brain INT map. As for main analyses, INT maps in patients were z-normalized relative to controls across two sites (MICA-MICs & EpiC) and sorted into ipsilateral/contralateral to the seizure focus.

Statistical analysis
i) Case-control analysis. Statistical analysis was carried out using SurfStat (https://mica-mni.github.io/surfstat/) 33 for MATLAB [R2021b, The Mathworks Inc.]. As in previous work 10, 25 , we fitted surface-based linear models to compare INT values between TLE and controls, controlling for age, sex, and site: For a given region i, INTi is the INT measure, Age and Sex are terms controlling for age and sex, respectively.
Site is the term controlling for site (i.e., EpiC and MICA-MICs) and Group is the group factor (i.e., TLE and controls). We also separately evaluated consistency of findings within each site, repeating the above group comparisons while controlling for age and sex: ii) Control for cortical morphology and microstructure. To assess INT alterations in TLE patients independent from grey matter morphological and SWM microstructural changes, we examined case-control INT differences while additionally correcting for cortical thickness and SWM diffusion features (i.e., FA, MD) at each vertex separately 10,19 . For subcortical structures and the hippocampus, we controlled for the corresponding volume and average diffusion parameters in the surrounding shell 18,25 .
iii) Age-and disease duration-related effects. To study interactions between diagnostic group and age on INT, we first built linear models that included an age and group main effect term, and an age×group interaction term. We also separately evaluated effects of age on INT within TLE and controls. To infer cumulative disease factors in TLE, we assessed effects of disease duration and age of seizure onset on INT.
iv) Sensitivity analyses. We performed several sensitivity analyses to demonstrate robustness and consistency of our main findings. First, we repeated the INT analysis while controlling for head motion, after calculating mean framewise displacement (FD) of each participant during the rs-fMRI scan 34 . We also repeated the INT analysis while additionally regressing out the global mean signal.
ⅴ) Correction for multiple comparisons. Findings were corrected for multiple comparisons by controlling the family-wise error (FWE) at pFWE < 0.05 for the cortex using random field theory and false discovery rate (FDR) at pFDR < 0.05 for subcortical structures and the hippocampus, respectively.

Spatial permutation tests
Statistical significance of spatial correlations between cortical maps was assessed using spin permutation tests that accounted for spatial autocorrelation 35 , implemented in BrainSpace (https://brainspace.readthedocs.io/) 36 .
Specifically, spin tests generated null models (1,000 repetitions) of brain maps by randomly rotating the maps and reassigning each vertex with the values of its nearest neighbors 35 . The significance of the real correlation coefficient was estimated as its position in the null distributions. A similar framework assessed significance of spatial correlations between subcortical maps with the exception that subcortical labels were randomly shuffled while preserving the pairwise three-dimensional Euclidean distance, using the ENIGMA Toolbox (https://enigma-toolbox.readthedocs.io/) 37 .

Classification and lateralization analysis
To explore whether INT could discriminate patients from controls and lateralize the seizure focus in patients, we utilized support vector machines implemented in LIBSVM (https://www.csie.ntu.edu.tw/~cjlin/libsvm/) 38 .
To reduce overfitting, we parcellated cortex-wide INT maps using the Schafer-300 atlas 39 , and subsequently generated inter-hemispheric asymmetry maps, computed as AI = ipsi -contra 40 , where ipsi and contra was the INT of ipsilateral and contralateral areas. The latter procedure was also applied to subcortical regions and the hippocampus, which yielded 7 additional features. Two models were thus evaluated: Considering that informative features differed slightly from fold to fold, we obtained feature weights by averaging all folds' feature weights and then normalized them across parcels (i.e., | ! |/ ∑ | ! | with i indicating i th parcel). In the classification analysis, we z-normalized asymmetry maps in patients with respect to controls and sorted feature data into ipsilateral/contralateral to the focus, and regressed out age, sex, and site. In the lateralization analysis, we did not z-normalize asymmetry maps. Here, we first generated a general linear model for the training set to reduce the effects of age, sex, and site, and applied the estimated parameters to the testing set. The residuals served as algorithm inputs. We assessed classification and lateralization algorithms in two different scenarios. First, we used 5-fold cross-validation with 100 iterations across both datasets combined.
Participants across both datasets were randomly split into 5 folds. Classifiers were trained on 4 folds and tested iteratively on the one held-out until all had served as a testing set; this procedure was repeated 100 times so that we can obtain different training and test sets. AUC and balanced accuracy were measured for each repetition and averaged across them. Statistical significance was assessed using 1,000 permutation tests with a threshold of p<0.05. Briefly, the features were unchanged, and labels were randomly shuffled for 1,000 times and split into the training and testing sets. The p value was calculated by dividing the shuffled times by the number that was equal to or higher than the real balanced accuracy or AUC. Second, we also evaluated crosssite generalizability. Here, we trained the classification/lateralization algorithms on one dataset using leaveone-subject-out cross validation and tested them on the other dataset.

Group differences in INT
We  Each dot represents an individual vertex or subcortical/hippocampal structure. ***, p < 0.001.

Effects of cortical morphology and microstructure
We repeated INT analyses after additionally controlling for MRI-based measures of grey matter morphology and SWM microstructure. As for grey matter changes, TLE showed bilateral cortical and subcortical atrophy relative to controls, with strongest effects in paracentral, dorsomedial, and ventromedial prefrontal regions (pFWE < 0.05) as well as the ipsilateral hippocampus (pFDR < 0.05) (Figure 2A), in keeping with prior findings 10,12,25 . After controlling for grey matter changes, patterns of INT reductions were similar to our original findings (Figure 2A; neocortex: r = 0.998, pspin < 0.001; subcortex and hippocampus: r = 0.933, pshuf < 0.001). Effect sizes for INT differences were however reduced both at the neocortical level (Cohen's d = -0.47±0.09, 2% reduction) and at the level of subcortical and hippocampal regions (Cohen's d = -0.47±0.10, 16% reduction).
Effect size reductions after controlling for gray matter atrophy were particularly strong in the ipsilateral hippocampus (48% reduction).

Effects of age and disease duration on INT
Globally, we observed significant negative correlations between age and mean cortical and subcortical INT in people with TLE after controlling for sex and site (Figure 3A;

Classification and lateralization performance
We leveraged supervised statistical learning to distinguish patients from controls and lateralize the seizure focus in TLE. In the classification analyses, the combination of cortical and subcortical INT achieved a balanced accuracy of 76±2.65% (p = 0.001) and AUC of 0.82±0.02 (p = 0.001) ( Figure 4A and Table 2).
Selected features mapped onto temporo-polar, parahippocampal, medial orbitofrontal, and paracentral regions ( Figure 4A). In the lateralization analysis, the overall balanced accuracy was 96±2.10% (p = 0.001) and the AUC was 0.99±0.01 (p = 0.001), based on the combination of cortical and subcortical INT. Informative features were found in the paralimbic cortex, encompassing both temporal pole and insula ( Figure 4B).
Classifiers generalized well from one site to the other (Figure 4C and 4D, Table 2). Moreover, classifiers based on cortical features alone showed similar performance to that of whole brain models ( Figure S6 and Table S1).  Core to our work was the quantification of INT from resting-state fMRI (rs-fMRI), a recently developed measure that taps into the temporal autocorrelation of neural signals 14,15 . In our controls, this measure differentiated sensory and motor systems from heteromodal association cortices, which had longer timescales, in agreement with previous work in healthy adults [14][15][16] and developing cohorts 31 . Differences in timescales have previously been related to functional hierarchy of the human cortex 15 . In particular, shorter INT in sensory and motor systems have been associated to their more specialized, externally-oriented functional roles, which may facilitate neural responsiveness to changing environmental contexts. On the other hand, longer INT in heteromodal regions may contribute to their role in multisensory integration as well as abstract, higher-order processing 15,16 . In typical development, prior studies have shown changes in INT with aging, in particular a lengthening of intrinsic timescales in association and limbic cortices 41 , which may parallel ongoing changes in the microstructural and connectional differentiation in these regions throughout childhood and adolescence 42,43 . To identify regional patterns of functional imbalances in TLE, we fit neocortical, hippocampal, and subcortical models that additionally controlled for effects of age, sex, and site. This provided robust evidence for marked INT reductions in TLE, in a cortical-subcortical territory encompassing the hippocampus, and thalamus, alongside with lateral temporal, parieto-occipital, and paracentral regions 12 . Effects were overall bilateral, yet more extensive ipsilaterally. This was particularly observed in temporal regions, where INT extended anteriorly to temporopolar areas. In our work, findings of INT reductions were invariably preserved when additionally correcting for head motion and the global signal, suggesting that the results reflected real group differences and not spurious results driven by motion artefacts or imaging processing parameters.
independently acquired samples, thus being robust to differences in acquisition protocols and some between site differences in clinical and socio-demographic inclusion.
Prior rs-fMRI studies investigated atypical inter-regional connectivity, and showed atypical functional interactions within mesiotemporal regions 44,45 , between mesial and lateral temporal regions 7 , and in networks including both temporal and extratemporal regions 9, 44 . In addition, several studies suggest alterations in signal amplitude 19 or regional signaling homogeneity of neural signals 46  Leveraging supervised machine learning, we demonstrated that classifiers informed by INT could distinguish TLE from controls and lateralize the seizure focus with high accuracy. These findings complement earlier efforts in epilepsy classification and focus lateralization using a wide range of features from structural MRI, diffusion MRI, and rs-fMRI connectivity analysis [63][64][65] . Notably, we specifically assessed cross-site prediction and confirmed adequate generalization, contrasting most of the prior that evaluated functional classifiers within single sites. Informative features included mesiotemporal areas but also spanned beyond these, likely reflecting the known system-level involvement of TLE 4, 12 . Collectively, these findings highlight a promising utility of intrinsic time scale parametrization for personalized diagnostics in TLE, and motivate future biomarker discovery efforts that combine different structural and connectional measures with biologically meaningful indices of brain function, such as INT.  Figure S1. Between-group differences in regional intrinsic neural timescales (INT) carried out separately in the EpiC and MICA-MICs site, after correcting for age and sex. Figure S2. Between-group differences in regional intrinsic neural timescales (INT) after additionally regressing out head motion (top) or the global mean signal (bottom). Findings have been corrected for multiple comparisons at a family-wise error (FWE) rate < 0.05 for cortices, and at a false discovery rate (FDR) < 0.05 for subcortices and the hippocampus.