The association of structural connectome efficiency with cognition in children with epilepsy

OBJECTIVE
Cognitive impairment is common in children with epilepsy (CWE), but understanding the underlying pathological processes is challenging. We aimed to investigate the association of structural brain network organisation with cognition.


METHODS
This was a retrospective cohort study of CWE without structural brain abnormalities, comparing whole brain network characteristics between those with cognitive impairment and those with intact cognition. We created structural whole-brain connectomes from anatomical and diffusion tensor magnetic resonance imaging using the number of streamlines and tract-averaged fractional anisotropy. We assessed the differences in average path length and global network efficiency between children with cognitive impairment and those without,using multivariable analyses to account for possible clinical group differences.


RESULTS
Twenty-eight CWE and cognitive impairment had lower whole brain network global efficiency compared with 34 children with intact cognition (0.54, standard deviation (SD):0.003 vs. 0.56, SD:0.002, p < 0.001), which is equivalent to longer normalized network average path lengths (1.14, SD:0.05 vs. 1.10, SD:0.02, p = 0.003). In multivariable logistic regression cognitive impairment was not significantly associated with age of onset, duration of epilepsy, or number of antiseizure medications, but was independently associated with daily seizures (p = 0.04) and normalized average path length (p = 0.007).


CONCLUSIONS
Higher structural network average path length and lower global network efficiency may be imaging biomarkers of cognitive impairment in epilepsy. Understanding what leads to changes in structural connectivity could aid identification of modifiable risk factors for cognitive impairment. These findings are only applicable to the specific cohort studied, and further confirmation in other cohorts is required.


Introduction
Neurobehavioral and cognitive dysfunction are frequent in children with epilepsy (CWE), occurring in around 25-40% [1,2].Behavioural and cognitive issues can negatively impact quality of life and can be associated with increased health and social care needs [3][4][5].Cognitive dysfunction in epilepsy has been associated with an early age of onset [6,7], a longer duration of epilepsy [8], aetiology [5][6][7], increased seizure frequency [8], use of antiseizure medications (ASMs) [8], lower socioeconomic status [1], and the presence of structural brain abnormalities [9,10].A complex interplay of genetics, environment, epileptogenic pathology, the seizure disorder itself, and its drug treatment is proposed to underlie cognitive impairment in CWE [11].E-mail address: julie.woodfield@ed.ac.uk (J.Woodfield). 1 Member of ERN EpiCARE.Epilepsy is increasingly viewed as a disorder of brain network function [12], with local and global disruption of functional and structural brain networks reported, even in focal-onset epilepsies [12][13][14].Structural brain networks, or connectomes, can be created using cortical and subcortical anatomical regions as network nodes and the white matter tracts connecting these regions as network edges [12].Graph theory can then be applied to study mathematical properties of networks, such as average path length between nodes and the inverse of this, the global efficiency of the network [12].Disruptions in structural networks in adult temporal lobe epilepsy have been linked to neuropsychological deficits [15].In addition, structural brain network efficiency has been linked with cognitive development and intelligence in normally developing children and adolescents [16][17][18].However, only a few studies have investigated whether structural brain network topology is associated with the cognitive dysfunction seen in childhood epilepsy [10,19,20].Alterations in structural network modularity, clustering, and path length have been reported in CWE and cognitive dysfunction [19,20], but other studies have shown no association between cognition and network measures [10].Previous studies of structural connectomes in groups with CWE and cognitive impairment were carried out in those with mild cognitive dysfunction, with the impaired groups studied achieving mean intelligence quotient (IQ) scores above 92 points [10,20].
It is not known whether CWE and significant cognitive impairment that impacts daily life also have global structural network changes or whether the network changes seen in the small number of CWE and cognitive dysfunction studies so far are reproducible.Based on findings from normally developing children and adolescents [16][17][18], and studies showing structural network alterations in those with epilepsy and mild cognitive impairment [19,20], we hypothesised that those with epilepsy and significant cognitive impairment would have lower global efficiency or the equivalent of higher network average path lengths in networks modelled from structural magnetic resonance imaging (MRI) compared with those with normal cognitive function.We therefore aimed to investigate whether the global network efficiency of networks derived from structural MRI is associated with cognitive dysfunction in childhood epilepsy, particularly for those with severe cognitive and behavioural issues not included in prior studies.

Material and methods
This was a retrospective cohort study of CWE comparing the structural MRI connectomes between those with cognitive impairment and those with intact cognition.This manuscript was prepared in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) checklist [21].

Description of the cohort
We retrospectively identified consecutive CWE-either focal or generalised-who had undergone imaging with an epilepsyspecific MRI brain protocol at the University Medical Centre Utrecht and had no structural MRI abnormalities.Clinical and demographic details of patients scanned using this protocol were recorded in a research database at the time of imaging.We included those children aged between one and 17 years with a diagnosis of epilepsy documented by their treating paediatric neurologist.We excluded those under one year old (due to poor greywhite differentiation making MRI segmentation unreliable) and children with no documentation of cognitive function or MRI not available for analysis.Clinical and demographic data were extracted from the research database and the electronic patient records and included: age at time of MRI, sex, age at time of first seizure, seizure frequency at time of MRI, ASMs being taken at time of MRI, seizure onset type (focal or generalised), electroencephalography (EEG) findings, epilepsy aetiology, and classification.

Cognitive impairment
Electronic health records were searched to establish if the child had a diagnosis of intellectual disability documented by their neuropsychologist, paediatrician, or paediatric neurologist.If the child had a documented diagnosis of an intellectual disability or if they attended special education due to significant educational needs, they were assigned to the group with cognitive impairment.If the child had documentation that there were no concerns about their cognitive development or educational attainment, they were assigned to the group with intact cognition.Children with special educational needs only due to a physical disability with normal cognitive and educational development were assigned to the group with intact cognition.Children with no documentation regarding intellectual disability, no neuropsychology assessment, or no documentation of educational attainment were excluded.Group assignment (impaired cognition or not) was independently determined from each child's medical record by two doctors on two separate occasions and discussed and agreed upon by consensus with a third doctor.
Neuropsychological and cognitive assessments were undertaken as clinically indicated, so not all patients underwent formal neuropsychological testing.When neuropsychological testing was carried out, it was by a paediatric neuropsychologist using ageappropriate instruments.We extracted total IQ, or developmental quotient (DQ) scores.Both IQ and DQ scores were combined and used in analyses as raw scores, as all tests used had a population-normal mean of 100 points and a standard deviation of 15 points.We did not make any new diagnoses of cognitive impairment using the scores as all children who had undergone neuropsychological testing had a comprehensive neuropsychological assessment and diagnoses made as appropriate.When patients scored lower than the lowest score possible on the test or were not able to be assigned an IQ score, they were excluded from the analyses of IQ scores but were still included in the group with cognitive impairment.

Structural network construction
All participants underwent a three-dimensional (3D) volume T1-weighted MRI sequence and a diffusion tensor imaging (DTI) sequence on a 1.5 T Philips Achieva scanner with a SENSE 8 channel head coil.The 3D volume T1-weighted scan was a fast field echo sequence (repetition time (TR): 25 ms, echo time (TE): 4.6 ms, flip angle: 30 • , FOV: 256 Â 256 mm [2], slice thickness: 1 mm, acquisition time: 2 min 11 s).The DTI sequence consisted of one single-shot spin-echo EPI b = 0 s/mm 2 acquisition followed by EPI scans with 15 different diffusion directions at a b value of 800 s/mm 2 (TR: 1,382 ms, TE: 60 ms, flip angle = 90 • , field of view (FOV) = 128 Â 128 mm 2 , slice thickness = 2 mm, acquisition time: 2 min 12 s).Participants with excessive artefacts or structural abnormalities were excluded.The 3D T1-weighted sequence was processed using Freesurfer (https://surfer.nmr.mgh.harvard.edu/).Processing included extracting the brain using a watershed and surface deformation procedure [22], volumetric segmentation of deep grey matter and subcortical white matter [23], definition of tissue boundaries using intensity normalisation and automated topology correction [24], then surface-based registration by inflation of the cortical surface and parcellation into cortical units matching sulcal and gyral folding patterns using the Desikan-Killiany atlas [25,26].All segmentations and parcellations were manually checked for quality and accuracy at each stage, and manual adjustments were undertaken as required.
Motion and eddy current distortions were corrected in the DTI sequences by aligning all 15 diffusion-weighted scans to the b = 0 s/mm 2 acquisition [27].Diffusion tensors were estimated at each voxel, and the main diffusion direction was selected as that voxel's principal eigenvector using robust estimation of tensors by outlier rejection [28].
The fractional anisotropy (FA) at each voxel was calculated, and deterministic fibre tractography by means of fibre assignment by the continuous tracking (FACT) algorithm was used to reconstruct whole brain white matter pathways [29].Eight seed points were placed in each white matter voxel; and the stopping criteria used were an FA of less than 0.1 in the voxel, a turning angle more than 45 degrees, or exit of the streamline from the brain mask volume.Linear registration of the DTI sequence to the 3D T1-weighted sequence was performed to align the segmentation with the streamlines.
Whole brain networks were created in matrix laboratory (MATLAB) using the 68 cortical regions of the Desikan-Killiany atlas, 14 subcortical regions, and three brain stem regions as nodes (see the Supplementary information for a full list of the regions used).An FA-weighted network adjacency matrix was formed by averaging the FA values in all voxels traversed by the streamlines connecting each pair of nodes.We also created a raw number of streamlines (NoS) adjacency matrix by counting the streamlines connecting each pair of nodes.Binary networks were created from both the FA and NoS adjacency matrices by setting all present connections to a weight of one.

Network measures
The MATLAB Brain Connectivity Toolbox was used to analyse networks [30].Network density was calculated as the number of edges present in the network as a proportion of all possible edges.Node degree was calculated as the number of edges connected to that node, and mean network degree as the mean of each node's degree.Node strength was calculated as the sum of all the edge weights of edges connected to that node, and mean network strength was calculated as the mean of all node strengths in the network.The mean network edge weight was calculated as the mean of all edge weights.The average shortest path length across the network was calculated using the mean of the shortest path length between each pair of nodes.For weighted networks, the edge weights were first converted to lengths by inverting the edge weights.Network global efficiency was calculated as the mean of the inverse shortest path length (L) between each node.Path lengths (L) were normalised by comparison to randomised reference networks with the same number of nodes, edges, and node degrees as the original networks by computing 1,000 random networks from each patient's own individual network.The shortest path length was calculated for each of the randomly rewired networks, and each patient's mean shortest path length was divided by the mean of the mean shortest path lengths from the randomly rewired networks to create k, the normalised average path length [30].

Statistical analysis
Statistical analysis was carried out using R version 4.2.1.Structural network measures were compared between those with cognitive impairment and those with intact cognition.In the subgroup with IQ scores available, network measures were also correlated with IQ.Between-group comparisons were performed using t tests for IQ scores and connectome measures (normally distributed) and Mann-Whitney U tests for other numerical data.Categorical data were compared using X 2 tests.The correlation of IQ with network measures was performed with Pearson's correlation coefficient.Multiple comparisons were accounted for by controlling the false discovery rate (FDR) at 5% [31].
For multivariable analyses, logistic regression was performed for associations with cognitive impairment and linear regression for associations with IQ.Epilepsy characteristics selected for inclusion in the multivariable analysis were those hypothesised to be related to cognition: age at MRI, sex, age at time of first seizure, time from first seizure to MRI, seizure frequency, number of ASMs, and seizure onset type (focal or generalised).As age at MRI was correlated with age at onset of epilepsy and the duration of epilepsy, we could not include it in the model due to collinearity.For the multivariable models, we chose k as the measure of network efficiency as it is normalised for density and node degree, unlike average path length or global efficiency.We could not include density, degree, strength, or edge weight to control for these network features due to the high correlation between these variables and k.It is not possible to include L, k, and global efficiency because they are all highly correlated and global efficiency is the inverse of L.

Ethical approval
The Dutch Medical Ethics Committee (METC) of the UMCU considered the study and confirmed that the Medical Research Involving Human Subjects Act (WMO) did not apply and that official approval of the study was not required.This allowed retrospective analysis of routinely acquired clinical, neuropsychological, and imaging data without requiring patient consent.

Creation of the cohort
Of the 130 patients who underwent the epilepsy MRI protocol, we identified 62 children over the age of one year with epilepsy and documentation of cognitive function who did not have structural MRI lesions.Fig. 1 shows the study flow diagram with reasons for exclusion.In 28 per 62 (45%), a diagnosis of cognitive impairment was documented, and the remainder, 34 per 62 (55%), had documentation of normal cognitive function.Neuropsychological test results were available for 27 participants, 11 of those with normal cognitive function and 16 of those with cognitive impairment (see Fig. 1).Eight of the sixteen with cognitive impairment were too impaired to achieve an IQ or DQ score and were excluded from the analysis with IQ.For the 19 with a score, tests used for assessment were: Weschler Intelligence Scales for Children (WISC, n = 10, IQ); Wechsler Preschool and Primary Scales of Intelligence (WPPSI, n = 2, IQ); Snijders-Oomen Niet-verbale Intelligentietest (SON, n = 2, IQ); and Bayley Scales of Infant Development (BSID, n = 5, DQ).

Demographic and clinical features
Demographic and clinical characteristics in the groups with and without cognitive impairment are shown in Table 1.Those included in this analysis (n = 62) were similar to the whole cohort of those undergoing MRI (n = 130; 41% female; median age 6 years; 30% experiencing daily seizures; 23% focal onset seizures).Eleven of the 62 included children (18%) had focal seizure onset, and the remainder had generalised seizure onset.Children with cognitive impairment were younger at the time of the first seizure (median 1 year) compared with those without cognitive impairment (median 3 years).The duration of epilepsy at the time of the MRI scan was longer in those with cognitive impairment (median 30 months) compared with those without cognitive impairment (median 14 months).The structural MRI scan was carried out within six months of the first seizure in 18 out of 58 (31%).The time of the first seizure was unknown to four participants.There was a higher proportion of children with daily or more frequent seizures (50%) in the group with cognitive impairment compared with those without (12%), and a lower proportion with focal onset seizures (11%) compared with those without cognitive impairment (24%; see Table 1).The proportion with daily seizures was the only group difference that was statistically significant (p = 0.002).
Participants were taking between zero and four ASMs at the time of their MRI scan.For some participants, the MRI scan was their first imaging for suspected epilepsy, and ASMs were started after the MRI.For six participants, ASM use at the time of the MRI scan was not documented.In the group with cognitive impairment, 4 out of 23 (17%) were not taking any ASMs, compared with 12 out of 33 (36%) of those without cognitive impairment.For those with cognitive impairment whose seizures were less frequent than daily, seizure frequency was weekly (n = 3), monthly (n = 1), and less than once a month (n = 7).In those without cognitive impairment having fewer than daily seizures, seizure frequency was weekly (n = 4), monthly (n = 5), and less than once a month (n = 20).

Network characteristics and cognitive impairment
In 48 out of 62 (77%) of the participants, MRI scans were carried out under general anaesthesia.Structural connectomes were created using tract-averaged FA and NoS successfully for all participants.Both weighted and binary networks were created for both FA and NoS networks.Group averaged weighted FA networks are shown in Fig. 2. Network average path length, L, was statistically significantly higher (p = 0.002), and as a consequence, global efficiency was significantly lower (p < 0.001) in the group with cognitive impairment in both binary and weighted networks after controlling for multiple comparisons (see Table 2).Network density and mean degree were lower in both the FA-weighted and NoS binary networks.In the weighted networks, mean edge weights (p < 0.001) and mean node strength (p < 0.001) were lower in the group with cognitive impairment in networks constructed from tract averaged FA, but the difference was not statistically significant in weighted networks constructed from NoS.After normalisation of the mean average path length (L) to random networks with the same number of nodes, edges, and node degrees, normalised path length (k) was also statistically significantly lower in the group with cognitive impairment in binary and weighted FA and NoS networks (p = 0.003).

Network characteristics, epilepsy, and cognitive impairment
To assess whether the relationship between average path length and cognitive impairment is influenced by differences in clinical features, we created a multivariable logistic regression model with cognitive impairment as the dependent variable and added clinical epilepsy characteristics to the association between k and cognitive impairment (see Table 3).We chose k as the measure of path length as it is normalised for density and node degree.The relationship between k and cognitive impairment remained in the multivariable model (p = 0.007, see Table 3).The relationship between daily or more frequent seizures and cognitive impairment also remained when controlling for other epilepsy-related factors and k (see Table 3).As lower mean node strength and network mean edge weight were associated with cognitive impairment (Table 2), the model was repeated in the binary networks (see Table 3) to ensure the relationship between k and cognitive impairment was not entirely driven by lower network degree or density.Similar results were found in the binary networks (see Table 3), suggesting that the association between k and cognitive impairment cannot be accounted for by differences in clinical or epilepsy characteristics or overall network density and degree.

Intelligence quotient and mean path length
To further confirm our finding of an association between cognitive impairment and lower network efficiency, we assessed the relationship between IQ score and path length in the 19 participants for whom an IQ score was available.The Pearson correlation for global efficiency and IQ score was 0.62 (95% CI: 0.32-0.81,p < 0.001) for FA-weighted networks.We reported global efficiency due to the positive direction of the association.There were expected negative correlations for the inverse measure of path length.This correlation between network global efficiency and IQ supports the finding that path length is related to cognitive function.We also created a linear regression model using IQ as the dependent variable to assess the associations with both clinical epilepsy features and path length.For the multivariable model, we used k because it is normalised for network density and degree.Table 4 shows that both lower k and a lower number of ASMs are independently associated with IQ in this cohort.

Summary of findings
In this cohort of CWE, higher structural network average path length and lower network global efficiency were associated with cognitive impairment and lower IQ.This finding was apparent in network models created using different methods: binary and weighted; FA and NoS.This association is unlikely to be caused only by differences in epilepsy characteristics between the groups because it persists in multivariable analysis, including epilepsy characteristics.It is also unlikely to be due to differences in the number of edges and nodes because when path length is normalised to networks with the same number of nodes and edges, the normalised value remains associated with cognitive impairment and IQ.

Comparison to previous studies
Our findings are in keeping with previous studies in which a non-statistically significant higher path length was seen in those with frontal lobe epilepsy and cognitive impairment [19], and in those with new-onset epilepsy, in which those with lower IQ scores had higher average path lengths [20].However, in children undergoing epilepsy surgery, no correlations between presurgical global network properties and IQ were found [10].The association between structural network efficiency and IQ may change with development [16], and findings may be dependent on the groups studied.In particular, focal structural epilepsies-inherently included in surgical populations-may have different global network properties than children with non-lesional epilepsy included in our cohort.Previous studies have mostly investigated children who were able to complete an age-appropriate IQ test and had mean IQ scores near the population mean of 100 [10,19,20].This is in contrast with the high proportion of children with a significant intellectual disability in our cohort.The lack of association between IQ and structural network characteristics in healthy children could be due to the lack of variability in cognitive abilities studied or different population characteristics [15][16][17][32][33][34][35].
Fig. 2. Weighted-FA group averaged connectomes.Group averaged connectomes are shown for the group with cognitive impairment (top) and the group with intact cognition (bottom).Connections which were present in more than 60% of the group were kept.Weight of the edge is colour coded from zero (absent, blue) to 1 (high, yellow).Node order is listed in the Supplementary material (FA: fractional anisotropy).

Significance of findings
We controlled for important clinical epilepsy features in the relationship between longer network average path lengths and cognitive impairment to assess whether one or many of the features of more severe epilepsy could account for the network findings and cognitive impairment.Daily or more frequent seizures were associated with cognitive impairment in both univariable and multivariable models.Epilepsy has been associated with widespread micro-structural brain changes even when brains are macroscopically structurally normal [36,37], and more frequent seizures and an early age of onset are associated with lower cognitive functioning [11].Changes in white matter structure as a result of uncontrolled seizures may mediate the association between epilepsy, cognitive impairment, and impaired structural networks [11].Our findings of less efficient structural networks in CWE and cognitive impairment and the correlation of connectivity changes of functional and structural networks with cognitive impairment in other cohorts with epilepsy support the hypothesis that the disruption of structural and functional networks by sei-zures may mediate the association of epilepsy with cognitive impairment. 9 11-14 20However, the variables included in our multivariable models did not completely account for the network group differences.This suggests that the differences in network structure may also be driven by other factors not included in the models, such as genetic associations with the link between structural brain networks and cognition [16,33].Microstructural abnormalities, independently or as a result of epilepsy, genetic, and other factors may all play a multifactorial role in the association between cognitive function and structural network efficiency in epilepsy.Further work is needed to elucidate these relationships.
Network density, mean node degree and strength, and mean node edge weights were lower in those with cognitive impairment, in addition to the finding of higher average path lengths and lower global efficiency.Path lengths are generally lower in networks with a higher number of edges [38].It can therefore be difficult to know whether there is a true difference in path length and global efficiency between the two groups or whether this is led by the difference in the number of edges [38].As the associations remained after normalisation of the average path length to k, the findings   are likely due to differences in network topology and not only to a higher degree and strength.

Strengths and limitations
The methodological strengths of this study are that two different methods of determining network edge weights from DTI were investigated (NoS and FA), and the findings were consistent across modelling methods.Findings of a sensitivity analysis within the subset of participants with IQ available support the previously reported association between path length or network global efficiency and cognition [5].
This study is limited by its retrospective nature, which meant some data were unavailable, and we were unable to include the whole cohort in the analysis.However, those included had similar characteristics to those excluded.Accurate parcellation of the cortex and subcortical regions and determination of streamlines are necessitated, excluding those with structural abnormalities.Therefore, our findings cannot be applied to those with conditions such as tuberous sclerosis or tumours where pathological lesions interfere with the creation of structural connectomes.Examination of single (contralateral) hemisphere connectomes or exclusion of pathological anatomical regions are alternative methods for dealing with structural abnormalities, but these methods do not allow examination of whole brain connectomes.
Our study benefits from the pragmatic diagnosis of cognitive impairment, which facilitates the inclusion of those who are too impaired for assessment with standard neuropsychological tools.This makes our findings relevant to those who are most disabled and impacted by the cognitive impairment associated with epilepsy.This group is poorly studied with imaging, but the retrospective analysis of clinical data allowed us to use the MRI scans carried out under general anaesthesia for clinical purposes.All of those included had macroscopically normal MRI scans.Although using a dichotomized outcome measure limits the applicability across the range of cognitive disability, the findings in the less impaired subgroup with IQ scores available support the findings of an association between network efficiency and cognition.Using our dichotomized outcome facilitated inclusion of both those too impaired to complete IQ tests and those in whom IQ tests were not deemed clinically necessary due to good intellectual functioning.Including only those in whom formal neuropsychological assessment was undertaken may have biased the sample to those suspected of impairment.Due to the age range included, we combined IQ scores from age-appropriate instruments, as no single test could be used for all participants.The small study size and the heterogeneity of the cohort in terms of aetiology, age, and severity limit our conclusions to being preliminary findings that should be validated in larger cohorts.

Conclusions
In conclusion, in a cohort of CWE and non-lesional epilepsy, cognitive impairment was associated with higher structural DTI network average path lengths and lower global efficiency.This finding was consistent across networks built using different modelling networks, when path lengths were normalised for the density and degree of the networks, and when including markers of epilepsy severity in multivariable models.

Funding statement
This work was supported by The Wellcome Trust via the Edinburgh Clinical Academic Track (ECAT) Clinical Lectureship Scheme (106364/Z/14/Z).The funder had no role in study design, collection, analysis, and interpretation of data, the writing of the report, or the decision to submit the article for publication.For the purpose of open access, the author has applied a Creative Commons Attribution (CC BY) licence to any Author Accepted Manuscript version arising from this submission.
Epilepsy & Behavior 148 (2023) 109462 Contents lists available at ScienceDirect Epilepsy & Behavior j o u r n a l h o m e p a g e : w w w .e l s e v i e r .c o m / l o c a t e / y e b e h

Fig. 1 .
Fig. 1.Cohort inclusions and exclusions.Flow diagram showing number of participants eligible and reasons for exclusion.

Table 1
Clinical characteristics by cognitive impairment.
Comparison of the clinical and demographic features of those with and without cognitive impairment.There were four participants without data about the age of onset and duration of epilepsy, and six participants without data about the number of AEDs.IQ scores were available for 8 out of 28 (29%) of those with cognitive impairment and 11 out of 34 (32%) of those with intact cognition.IQ is mean with SD, ages and time are median with IQR, and categorical data are count with percentage.(MRI: magnetic resonance imaging; ASM: antiseizure medication; IQ: intelligence quotient; SD: standard deviation, IQR: interquartile range).

Table 2
Global network characteristics by cognitive impairment.Comparison of global network characteristics between those with cognitive impairment and those with intact cognition.Data are mean (SD) and statistical comparison was carried out using t-tests.All p values are adjusted after FDR correction at 5% for 20 comparisons (2 network types and 10 network measures).FA values fall between zero and one, but NoS networks were raw counts of streamlines and can therefore take any values.The mean network strength and edge weights given are those for the raw number of streamlines matrices.(FA: fractional anisotropy; NoS: number of streamlines; Eglob: global efficiency; L: average path length; k: normalised average path length; S: mean network strength; EW: mean network edge weight; SD: standard deviation; FDR: false discovery rate).

Table 3
Multivariable association with cognitive impairment.
A multivariable logistic regression model with the predictors listed was created.Cognitive impairment was the outcome variable.Both binary and weighted-FA networks were examined.(OR: odds ratio; 95%CI: 95% confidence interval; ASM: antiseizure medication; k: normalised average path).

Table 4
Linear regression of IQ scores.For the linear regression, 19 patients were included.Variables included are in the table.(ASM: antiseizure medication; k: normalised average path length).