JLE

Epileptic Disorders

MENU

Spike-associated networks and intracranial electrographic findings Volume 22, issue 3, June 2020

Figures


  • Figure 1

Tables

For the 30% of individuals with focal-onset epilepsy refractory to medical therapy, at least a quarter are estimated to be curable with epilepsy surgery (Engel, 2001). The benefits of epilepsy surgery with respect to quality of life, reduced mortality, and economics have been repeatedly demonstrated (Malmgren and Edelvik, 2017). The likelihood of seizure freedom after surgery, however, is highly variable and dependent on a number of clinical factors. For the most straightforward epilepsy surgery candidates, those with seizures arising from the mesial temporal lobe and mesial temporal sclerosis on MRI, up to 80% will achieve seizure freedom (Malmgren and Edelvik, 2017). In other focal epilepsy syndromes such as MRI-negative frontal lobe epilepsy, seizure freedom rates are as low as 30% (Malmgren and Edelvik, 2017). Functional connectivity and its associated network analysis methods are emerging as powerful tools with the potential to improve our understanding of the variability in epilepsy comorbidities and surgical outcomes (Bernhardt et al., 2015; Jin et al., 2015).

Functional connectivity is the inferred relationship between two anatomic regions based on their covarying neuronal activity. Functional connectivity can be measured from several different neuroimaging and neurophysiology modalities including functional MRI, both scalp and invasive electroencephalography (EEG), and magnetoencephalography (MEG). MEG is a useful neuroimaging tool in presurgical evaluations in epilepsy surgery. MEG's clinical applications, however, have been largely confined to localizing the source of interictal epileptiform discharges (“spikes”). In the past several years, new analytical methods have demonstrated MEG's potential for the study of epileptic networks (Englot et al., 2015). In this paper, we report on promising early findings of a new method to identify the functional epileptic network in surgical epilepsy patients using the high temporal resolution of MEG.

Methods

Patient enrollment

All individuals at the University of Colorado and Children's Hospital Colorado who underwent MEG recording for a presurgical epilepsy work-up and who subsequently had stereotactic electroencephalography (SEEG) monitoring from July 1st, 2017 to Dec 31st, 2018, were considered for inclusion. Individuals were excluded if visual review of the MEG revealed inadequate quality of the recording or fewer than five spikes per American Clinical Magnetoencephalography Society guidelines (Bagić et al., 2011). Medical records review and neurophysiology data analyses were performed with the approval of the Colorado Multiple Institutional Review Board. Source localization from the MEG data using single equivalent current dipole modeling was employed in the presurgical evaluation of all patients. The subsequent network analyses described in the following sections were performed after SEEG monitoring was complete as part of this retrospective study.

Data acquisition and preprocessing

MEG data were obtained using a Magnes 3600 WH whole-head MEG device (4D Neuroimaging, San Diego, CA, USA), comprised of 248 first-order axial-gradiometer sensors (5-cm baseline) in a magnetically shielded room (ETS-Lindgren, Cedar Park, TX, USA). Five head position indicator coils attached to the subject's scalp were used to determine the head position with respect to the sensor array. The locations of the coils with respect to three anatomical landmarks (nasion and preauricular points, with the intersection of the tragus and daith of the ear defining the preauricular points) and two extra non-fiducial points as well as the scalp surface (approximately 500 points) were determined with a 3D digitizer (Polhemus, Colchester, VT, USA). The MEG signals were acquired continuously in a 0.1-100-Hz bandwidth and sampled at 290.64 Hz and 24-bit quantization. Anatomic T1-weighted MRI scans were obtained for clinical purposes on MRI scanners at two separate locations (Children's Hospital Colorado and UCHealth at the University of Colorado) and used for co-registration and source localization. SEEG electrodes were stereotactically implanted using an image-guided system. The locations and total number of SEEG electrodes were tailored to each patient according to his or her presurgical evaluation and multidisciplinary epilepsy surgery conference discussions. Intracranial data were recorded at a minimum sampling rate of 512 Hz using 8-to-16-contact electrodes. Visual analyses of SEEG data, including identification of the irritative zone (IZ; spike-generating brain tissue) and the seizure onset zone (SOZ), were performed during routine clinical practice at each site.

Source reconstruction and estimation of the spike-associated network

MEG and MRI data were imported into MNE-Python v0.18.1 (Gramfort et al., 2013) for offline processing. A trained clinical magnetoencephalographer (JJB.) visually reviewed all MEG recordings. Noisy channels were visually identified and excluded from further analysis. Spikes were manually marked for each subject starting approximately 0-25 ms prior to the spike upslope. Spikes were included only if they were preceded by a minimum three-second spike-free interval. For each subject, a group of one-second baseline epochs starting 1.05 seconds prior to the spike marks (PRE) and a group of one-second spike-containing epochs starting at the spike marks (SPIKE) were extracted from the continuous time series data. The epochs were bandpass filtered from 12-30 Hz corresponding to the beta band. This frequency band preserves the predominant spike activity, which is by definition ≤70 ms in duration and has a favorable signal-to-noise ratio compared to higher frequency bands. Additionally, a minimum of 12 oscillations within the one-second time windows balances reliability of the measure with a high temporal sensitivity, as has been shown for coherence-based connectivity metrics (Sun et al., 2012).

MRI and MEG data were co-registered in the patient's anatomic space in MNE-Python. For each spike-containing epoch, source activity was estimated using standardized low-resolution brain electromagnetic tomography (sLORETA) constrained to cortical surfaces (Pascual-Marqui, 2002). The source-reconstructed data were then morphed to FreeSurfer's fsaverage brain (Reuter et al., 2012) and parcellated using the Human Connectome Project's Multimodal Parcellation (HCPMMP) (Glasser et al., 2016) consisting of 362 regions of interest (ROIs). The time course data were extracted for each parcel, and these time courses were used to generate whole-brain connectivity matrices for each PRE and SPIKE epoch. Connectivity strength was calculated using the absolute value of the imaginary component of coherence. The imaginary component of coherence is a common connectivity measure in MEG research because it removes zero-time lag correlations, which usually represent environmental artifacts and spatial leakage of inferred sources (Colclough et al., 2016).

Statistical analysis

Connectivity matrices from all PRE and SPIKE epochs were imported into the Network-Based Statistic (Zalesky et al., 2010) (NBS; v1.2) toolbox in MATLAB (R2018B). NBS compares network structure between two groups using permutation testing to determine the likelihood of finding a network of a given size after random re-organization. Many methods of comparing cortical networks operate at the level of individual node or edge features rather than at the whole network. By examining connectivity patterns at the network level, NBS is able to identify significant network differences even when some connections within the network would be too weak to be identified using other methods. NBS relies on a user-defined statistical threshold of the network graphs of interest. NBS tests were performed across a range of thresholds starting at t=2 and increasing in increments of 0.5 until no connections survived thresholding (maximum t-values ranged from 3.5 to 5 across subjects). A network difference between the PRE and SPIKE conditions was considered significant if NBS-based permutation testing indicated that the likelihood of identifying a network of a given size at a given threshold by chance alone was less than 5% (p < 0.05). For visualization purposes and preliminary anatomic comparisons, graph density thresholds for each individual patient were selected to yield graphs encompassing approximately 15% of the brain, or about 50 connections between the 362 ROIs in the HCPMMP. Networks of this size were large enough to include the IZ and adjacent tissue expected to be involved in propagation on visual review, and approximated the spatial extent and distribution captured during standard clinical SEEG at the first author's institution. In order to compare the results with findings from routine clinical practice, the anatomic distribution of the MEG-based spike-associated network was compared to the visually identified IZ and SOZ on SEEG.

Results

Thirteen subjects, eight adults and five children, met the initial inclusion criteria. Four of the adults were excluded due to an inadequate number of spikes. Nine individuals underwent subsequent network analysis (table 1). In six individuals (three adults and three children), a statistically significant spike-associated network (SAN) was identified (table 2). For these subjects, the SAN was significant in all cases at all t-statistic thresholds (supplementary table 1). The three individuals in whom a significant SAN could not be identified had <30 spikes. The only two individuals in the cohort with tuberous sclerosis were included in this group, though the clinical significance of this is uncertain. One third of the cohort was female. Gender did not appear to influence the detection of a SAN, although this study was not powered to detect such a difference.

In the six cases with a significant SAN, only one of whom had <30 spikes, nodes of the SAN overlapped with most or all of the anatomic regions in the SEEG-identified IZ and SOZ (table 2). Specifically, of 29 anatomic regions identified as part of the IZ among the six individuals, 21 were included in the SANs, yielding a sensitivity of 78%. Similarly, of the 20 anatomic regions identified as part of the SOZ among the six individuals, 16 were included in the SANs for a sensitivity of 83%. An illustrative case is shown in figure 1A-E. Determining an accurate specificity of the SAN method is more challenging. While the SAN consistently included areas that were not identified as part of the IZ or the SOZ by SEEG, most of the areas were not monitored during the intracranial recordings. If only those false positive nodes on the SAN, that were monitored by SEEG, were included, the overall specificity of the method in this cohort was 84%. On the other hand, if one assumes that unmonitored regions were truly not part of the IZ and/or SOZ, and therefore falsely positive when identified as part of the SAN, the specificity dropped to 67%.

In considering the potential effect of spike propagation from deeper sources, a separation of 50 ms between the PRE and SPIKE conditions was thought to be adequate given evidence that propagation appears to occur within 30 ms (Zumsteg et al., 2006). To confirm that the increased connectivity observed during the SPIKE condition did not precede the marked spikes, however, we performed a network comparison between the PRE condition and a 5sPRE condition (5 seconds before the IED). There were no statistically significant network increases between these time periods (supplementary table 2).

Discussion

This study demonstrates the anatomic correlation between a novel MEG analysis method of non-invasively estimating the epileptic network in epilepsy surgery candidates and subsequent intracranial EEG findings. Even after surgical resection, 20-70% of individuals, depending on their epilepsy type, will continue to have seizures (Malmgren and Edelvik, 2017). Additionally, 20-30% of potential surgical candidates never undergo resection at all despite extensive pre-surgical investigations and extended intracranial EEG recordings (Taussig et al., 2014; González-Martínez et al., 2016). In approximately 10% of children undergoing SEEG, the decision not to proceed is due to an inability to identify the epileptogenic zone (McGovern et al., 2019). Both invasive and non-invasive measures of functional connectivity are increasingly used to identify the epileptogenic zone, predict post-operative outcomes, explain surgical failures, and offer new insights into the widespread cognitive effects of focal epilepsies (Murakami et al., 2016; Bear et al., 2019). These methods have not yet, however, reached routine clinical practice due, in part, to a combination of complicated and labor-intensive analyses and limited clinical data.

The SAN method presented here offers some advantages over other approaches of applying functional connectivity to identify epileptic networks. Some of the most promising recent studies use intracranial electrodes to measure functional connectivity and identify the epileptogenic tissue (Lagarde et al., 2018). These studies are limited, however, to the brain areas covered by the intracranial electrode placement. In contrast, MEG is a non-invasive tool capable of studying whole-brain connectivity and is therefore not limited by electrode placement decisions. When considering only brain regions that were included in the SEEG recordings, nodes identified in the SAN were 84% specific for the intracranially-determined epileptic network. Based on these results, one might consider including these nodes during the SEEG planning stages. Additionally, the analytical pipeline can be largely automated, thereby alleviating some of the labor-intensiveness of other approaches. The only steps requiring direct human involvement are the visual identification of spikes for inclusion and co-registration of the MRI and MEG data. Notably, automated approaches to these latter steps are also increasingly feasible (Joshi et al., 2018).

These advantages are similar to another MEG-based method of exploring connectivity patterns related to spikes (Malinowska et al., 2014). In this study, the authors used ICA of MEG recordings to identify spike-containing components, which were then projected into source space. The identified networks showed excellent overlap with the results of their SEEG investigations. In contrast with our method, however, the authors found that the ICA-based networks were generally more anatomically limited than those identified on SEEG. One might expect that the ICA-based approach provides greater specificity and lower sensitivity compared with our method. As such, the preferred method will depend on the goal of the investigation.

In addition to surgical applications, the SAN method could provide novel insights into cognitive deficits associated with focal epilepsies. There is increasing recognition that the cognitive effects of focal epilepsies extend far beyond their foci, suggesting a larger epileptic network than might be expected based solely on the IZ and SOZ (Jehi, 2018; Bear et al., 2019). The broad spike-associated networks described in the present study could provide novel insights into the process underlying the widespread cognitive changes, and further study is warranted to test this possibility.

Some limitations also deserve consideration. The imaginary part of coherence reduces the effects of external noise and spectral leakage, but physiological zero-lag connections, i.e. co-varying physiological activity occurring at precisely the same time in multiple areas, are also inherently removed. These zero-lag connections can contain meaningful information (Vicente et al., 2008), and their loss should be balanced with the importance of eliminating spectral leakage and other sources of connectivity artifact. Future application of this technique will require further study and optimization of the network measurement parameters including the most informative frequency bands and the duration of the connectivity estimate time windows. Additionally, the significant SANs consistently overlapped the SEEG-identified IZ and SOZ but also extended well beyond these regions, particularly at lower network thresholds, thereby limiting direct clinical application. A large surgical epilepsy cohort with post-operative outcome data could address these issues by enabling comparison of network features at different computational settings with the clinical outcomes. The workflow could be further enhanced through the use of threshold-free approaches to the NBS, which eliminate the arbitration threshold decisions (Baggio et al., 2018).

Should the technique prove clinically useful, its implementation into the clinical workflow also faces a couple of hurdles. First, significant SANs were most consistently identifiable in individuals with ≥30 spikes, and not all patients will have an adequate spike burden. Second, access to MEG is restricted to a relatively small number of epilepsy centers. High-density EEG, while more computationally complicated, warrants investigation due to its lower cost and potential for longer-duration recordings that could enable capturing more spikes.

Network-based techniques are already helping us understand the network effects of epilepsy. In the future, such techniques, including the SAN presented here, might assist in the presurgical planning stages or even allow smaller resections if we can identify critical hubs in the epileptic network. Additionally, the extensive connectivity changes seen across the brain during spikes could provide new insights into the widespread cognitive comorbidities associated with focal epilepsy. The SAN results presented in this brief cohort report hint at the underlying potential of this technique to improve our understanding of epilepsy.

Supplementary data

Summary didactic slides and supplementary tables are available on the www.epilepticdisorders.com website.

Acknowledgements and disclosures

This study was supported, in part, by a grant from the National Institutes of Health, National Institute of Neurological Disorders and Stroke (NIH K12 NS089417).

Dr. Kirsch serves as a consultant for Ricoh. The remaining authors report no conflicts of interest.


* Preliminary findings from the results presented in this manuscript were presented at the 2019 Annual Meeting of the American Clinical Neurophysiology Society.