IDENTIFYING THE EPILEPTOGENIC ZONE IN INTERICTAL RESTING-STATE MEG SOURCE-SPACE NETWORKS

Transcription

1 IDENTIFYING THE EPILEPTOGENIC ZONE IN INTERICTAL RESTING-STATE MEG SOURCE-SPACE NETWORKS Ida A. Nissen 1, Cornelis J. Stam 1, Jaap C. Reijneveld 2, Ilse E.C.W. van Straaten 1, Eef J. Hendriks 3, Johannes C. Baayen, Philip C. De Witt Hamer, Sander Idema, Arjan Hillebrand 1 1 Department of Clinical Neurophysiology and MEG Center, VU University Medical Center, Amsterdam, the Netherlands 2 Brain Tumor Center Amsterdam & Department of Neurology, VU University Medical Center, Amsterdam, the Netherlands 3 Department of Radiology & Nuclear Medicine, VU University Medical Center, Amsterdam, the Netherlands Neurosurgical Center Amsterdam, VU University Medical Center, Amsterdam, the Netherlands Epilepsia 2017; 58:137-18

2 Chapter ABSTRACT OBJECTIVE In one-third of patients, seizures remain after epilepsy surgery, meaning that improved preoperative evaluation methods are needed to identify the epileptogenic zone. A potential framework for such a method is network theory, as it can be applied to non-invasive recordings, even in the absence of epileptiform activity. Our aim was to identify the epileptogenic zone on the basis of hub status of local brain areas in interictal MEG networks. METHODS Preoperative eyes-closed resting-state MEG recordings were retrospectively analysed in 22 patients with refractory epilepsy, of whom 1 were seizure-free one year after surgery. Beamformer-based time series were reconstructed for 90 cortical and subcortical automated anatomical labeling (AAL) regions of interest (ROIs). Broadband functional connectivity was estimated using the phase lag index in artifact-free epochs without interictal epileptiform abnormalities. A minimum spanning tree was generated to represent the network and the hub status of each ROI was calculated using betweenness centrality, which indicates the centrality of a node in a network. The correspondence of resection cavity to hub values was evaluated on four levels: resection cavity, lobar, hemisphere and temporal versus extra-temporal areas. RESULTS Hubs were localized within the resection cavity in 8/1 seizure-free patients and in 0/8 patients who were not seizure-free (57% sensitivity, 100% specificity, 73% accuracy). Hubs were localized in the lobe of resection in 9/1 seizure-free patients and in 0/8 patients who were not seizure-free (6% sensitivity, 100% specificity, 77% accuracy). For the other two levels, the true negatives are unknown, hence only sensitivity could be determined: hubs coincided with both the resection hemisphere and the resection location (temporal versus extra-temporal) in 11/1 seizure-free patients (79% sensitivity). SIGNIFICANCE Identifying hubs non-invasively before surgery is a valuable approach with the potential of indicating the epileptogenic zone in patients without interictal abnormalities. 5

3 Interictal MEG networks INTRODUCTION The most common treatment for refractory epilepsy is epilepsy surgery, which aims to remove the epileptogenic zone, i.e. the brain region that needs to be removed or disconnected to result in seizure freedom (Lüders et al., 2006). At present, one out of three patients continues to experience seizures after surgery (Jobst and Cascino, 2015). To improve outcome in these patients, new approaches are needed. Epilepsy is increasingly seen as a disorder of brain networks (Stam, 201; van Diessen et al., 2013a). Considering the brain as a complex network of interconnected regions may provide opportunities for novel analysis strategies. For example, Ponten and colleagues showed that networks in patients with epilepsy deviate from the optimal network configuration seen in healthy controls (Ponten et al., 2007), and Jin and colleagues found that patients with left mesial temporal lobe epilepsy have stronger network hubs (nodes that are central in the network) in the affected lobe compared to healthy controls (Jin et al., 2015). Most studies report hyperconnectivity in the epileptogenic zone (Centeno and Carmichael, 201), but few studies (mainly fmri) find the opposite pattern (Bettus et al., 2009; Ortega et al., 2011; Pittau et al., 2012). However, Bettus and colleagues described increased connectivity with SEEG while connectivity was decreased in the same regions with fmri (Bettus et al., 2011). Networks can be constructed on the basis of structural or functional connectivity. Both structural and functional networks are affected in patients with epilepsy, but functional networks capture abnormalities in brain activity (i.e. epileptiform activity) even in the absence of structural abnormalities (Koepp and Woermann, 2005). Functional networks can be based on both invasive and non-invasive recordings. Non-invasive methods (e.g. Electro- and Magneto-Encephalography (EEG/MEG)) bear no risks of complications in contrast to invasive methods (e.g. stereo EEG (SEEG) and Electrocorticography (ECoG)). Interictal scalp EEG and MEG are complementary techniques (Barkley, 200; Baumgartner, 200), where MEG is less perturbed by the skull and other tissue in the head, is reference-free, and has a slightly higher spatial resolution than clinical scalp EEG (Ebersole and Ebersole, 2010). This allows for more accurate estimation of functional interactions between brain regions, and therefore more accurate reconstruction of the functional networks. Hubs are regions that play, by definition, a central role within the network; they are highly connected and/or are crucial for efficient transmission of signals across the network (van den Heuvel and Sporns, 2013). Hubs may also facilitate the spread of epileptiform activity to the rest of the brain and thus could play an important role in seizures. A hub that propagates epileptiform activity is referred to as a pathological hub. Several studies have revealed abnormal hubs in or near regions classified as the seizure onset zone (area 55

4 Chapter that initiates seizures (Lüders et al., 2006)) or the epileptogenic zone (Bernhardt et al., 2011; Jin et al., 2015; Liao et al., 2010; van Diessen et al., 2013b; Zhang et al., 2011b). Two invasive studies found that the seizure onset zone and epileptogenic lesions showed increased hub status (Varotto et al., 2012; Wilke et al., 2011), indicating that the epileptogenic zone can potentially be identified with invasive recordings using a network measure for hub status. These studies used betweenness centrality, which quantifies the centrality of a brain region in the network. It is defined as the fraction of shortest paths that pass through a given region (Freeman, 1977) and high betweenness values thus indicate regions that are intermediate to many connections between regions within the network. Results from non-invasive studies point in the same direction (Bernhardt et al., 2011; Jin et al., 2015; Liao et al., 2010; Zhang et al., 2011b). Most previous studies rely on ictal or interictal epileptiform activity to localize epileptogenic tissue. However, not all recordings capture ictal activity, and some recordings even fail to display interictal epileptiform activity (12% of 382 patients in a recent study (Nissen et al., 2016a)). Therefore, new methods are needed that do not depend on the presence of ictal or interictal epileptiform activity to identify the area that needs to be resected (Iannotti et al., 2016; Krishnan et al., 2015). Network theory offers a framework for such a method. The aim of our study was to identify the epileptogenic zone non-invasively in interictal data without epileptiform activity. Based on previous studies, we hypothesized that pathological hubs are located in or near the epileptogenic zone and therefore that the epileptogenic zone has a high betweenness centrality. Identification of the pathological hubs could therefore help to localize the epileptogenic zone, and would provide a target for removal or disconnection. MATERIAL AND METHODS Pa ents Twenty-two patients with refractory epilepsy received a clinical MEG recording as part of their preoperative evaluation at the MEG center of the VU University Medical Center, Amsterdam, the Netherlands. All patients underwent epilepsy surgery. Surgical outcome was scored according to the Engel classification more than a year after surgery. Patients were not subjected to procedures and were not required to follow rules of behavior other than routine clinical care, therefore approval for this study by the institutional review board and informed consent were not required according to the Dutch health law of February 26, 1998 (amended March 1, 2006), i.e. Wet Medisch-Wetenschappelijk Onderzoek met mensen (WMO; Medical Research Involving Human Subjects Act), division 1, section

5 Interictal MEG networks MEG acquisi on MEG recordings were performed with a whole-head system (Elekta Neuromag Oy, Helsinki, Finland), containing 306 channels (102 magnetometers and 20 gradiometers). Patients were placed in supine position in the scanner inside a magnetically shielded room (Vacuumschmelze GmbH, Hanau, Germany). Eyes-closed resting-state recordings were acquired for the identification and localization of interictal epileptiform activity, typically yielding three datasets of 15 minutes each. Additional paradigms were recorded but not analyzed here, namely paradigms for the localization of eloquent cortex, such as voluntary movements and somatosensory stimulation (see e.g. (Hillebrand et al., 2013)). The sampling frequency was 1250 Hz and the data were filtered online with a 10 Hz antialiasing filter and a 0.1 Hz high-pass filter. The head position was recorded continuously relative to the MEG sensors with or 5 head-localization coils. The head-localization coil positions and scalp outline (roughly 500 points) were digitized using a 3D digitizer (Fastrak, Polhemus, Colchester, VT, USA). The points on the scalp surface were coregistered with the anatomical MRI of the patient using surface-matching. Artifacts were removed by spatially filtering the raw data offline using the temporal extension of Signal Space Separation (tsss) (Taulu and Simola, 2006) as implemented in MaxFilter software (Elekta Neuromag Oy; version 2.1). Details and parameter settings have been described previously (Hillebrand et al., 2013). Atlas-based beamforming Neuronal activity was reconstructed using an atlas-based beamforming approach, modified from (Hillebrand et al., 2012). Here, we reconstructed time series of neuronal activation for the centroids(hillebrand et al., 2016b) of 90 ROIs in the AAL (automated anatomical labelling) atlas (Tzourio-Mazoyer et al., 2002), namely all 78 cortical ROIs (Gong et al., 2009) and 12 subcortical structures (excluding the cerebellar ROIs). These centroids were inversely transformed to the co-registered MRI of the patient. Consecutively, the time series for each centroid were reconstructed using a scalar beamformer (Elekta Neuromag Oy; beamformer; version ). The beamformer weights form a spatial filter and were calculated for each centroid separately in order to maximally let signals pass that originate from the centroid of interest and to attenuate all other signals. The weights are based on the lead fields (using a spherical head model based on the scalp surface obtained from the anatomical MRI of the patient), the data covariance and noise covariance. Data covariance was based on the entire recording. On average, 870 seconds of data (range: ) were used, and the time-series were filtered in the broadband (0.5-8 Hz). A unity matrix was used as noise covariance when estimating the optimum source orientation(sekihara et al., 200) for the beamformer weights. The broadband data were projected through the 57

6 Chapter Table 1: Baseline patient characteristics and findings from different modalities during preoperative evaluation. 58

7 Interictal MEG networks 59

8 Chapter normalised beamformer weights (Cheyne et al., 2007) in order to obtain time series (virtual electrodes) for each centroid (Hillebrand et al., 2016b). Func onal networks For each patient, 20 epochs without epileptiform activity and artefacts were selected containing 096 samples (3.28s) for further analysis in Brainwave (version available from in the broadband (0.5-8 Hz), theta band (-8 Hz) and lower alpha band (8-10 Hz), as previous studies have associated these bands with network measures in epilepsy patients (Douw et al., 2010b; van Dellen et al., 201). Customarily, 20 epochs (65.6s in total) are sufficient for stable connectivity values (Engels et al., 2015). However, to test the repeatability of the main results, we analysed all epochs in the entire recording that did not contain epileptiform activity or artefacts, which resulted in 16 epochs per patient (based on the patient with the lowest number of available epochs). A functional network was constructed based on the 90 virtual electrode time series. First, functional connectivity was estimated using the phase lag index (PLI) (Stam et al., 2007), which measures the asymmetry in the distribution of instantaneous phase differences between two time series. It ranges between zero (flat distribution due to noise or no connectivity, or a phase difference of zero modulus pi due to volume conduction) and one (full synchronization), and is robust against the effects of volume conduction and field spread (Porz et al., 201). Using the PLI, the minimum spanning tree (MST) was constructed, which forms the backbone of the original network, where the ROIs served as nodes and the inverted PLI values (1/PLI) as edge weights. The MST is formed by starting with an empty network, and by subsequently adding the edges with the smallest weights (i.e. with the highest PLI), until all nodes in the network are connected, while discarding edges that form loops (Kruskal, 1956). The MST always contains the same number of edges, when the number of nodes is kept constant, which enables direct comparison across groups or studies (Stam et al., 201; Tewarie et al., 2015) without the need for arbitrary thresholds (van Wijk et al., 2010). Betweenness centrality (Stam et al., 201) was calculated per node to identify hubs, i.e. nodes that play an important role in the network (Sporns et al., 2007). Rela on between resec on cavi es and hubs The resection cavity was segmented manually on the three month post-operative MRI scan using iplan 3.0 software (BrainLAB AG, Feldkirchen, Germany). The post-operative scan was linearly registered with the pre-operative MRI (used for MEG co-registration), after which the resection cavity was exported and co-registered using the same transformation 60

9 Interictal MEG networks matrix as for the MEG analysis (see above). The overlap of the resection cavity with the AAL ROIs was determined using in-house developed software (AH). We applied two methods. First, we determined whether the ROIs that contained the ten highest betweenness centrality values would fully or partly overlap with the resection cavity. We then calculated the sensitivity, specificity, and diagnostic accuracy using different thresholds, i.e. using an increasing number of ROIs from the (ranked) ten highest MST betweenness centrality values. The optimal threshold was determined by a receiver operating characteristic (ROC) curve using the Youden Index (Youden, 1950). After deriving the optimal threshold, we compared the correct identification of the resection cavity in seizure-free versus not seizure-free patients with Fisher s exact test (because of the small sample size), using IBM SPSS Statistics 20.0 (SPSS Inc., Chicago, IL, USA). Besides the comparison on a ROI level, we analyzed three other levels with different spatial resolution: lobar level, hemispheric level, and temporal versus extra-temporal level (i.e. whether ROIs with high betweenness centrality values were located within the lobe of resection, in the hemisphere of the resection, or in the temporal or extra-temporal areas without regarding lateralization, respectively). For the hemisphere and temporal/extra-temporal level analysis, 1,3,5,7,9, and 11 (odd to avoid ties) highest values were used as thresholds, of which the indication of the majority was used to determine the affected hemisphere or temporal/extra-temporal area. These analyses were only applied to data from seizure-free patients, since for patients who were not seizure-free it remains unknown whether the location of the epileptogenic zone is within the hemisphere of resection or contralateral hemisphere (and similarly unknown for the temporal versus extra-temporal level). Second, we averaged the betweenness centrality values of the ROIs that overlapped with the resection cavity in order to discriminate between seizure-free and not seizure-free patients. The number of ROIs that overlapped with the resection cavity varied per patient (median: 9, range: 2-18), as each patient had a different resection cavity. Different thresholds were calculated per patient using % of the maximum betweenness centrality value (out of all 90 ROIs). An average above the threshold in seizure-free patients was counted as a true positive; an average below the threshold in not seizure-free patients was considered a true negative. The Youden Index determined the optimal threshold. RESULTS This study included 22 patients with refractory epilepsy, of whom 1 were seizure-free at least one year after surgery (Table 1). Twelve patients (55%) were female and the median age was 33 years. Six patients had MTS and eight patients a tumor. The majority of the 61

10 Chapter 62

11 Interictal MEG networks Figure 1: Rela on between the lobe of resec on and betweenness centrality for the (A) 1 seizurefree pa ents and (B) 8 not seizure-free pa ents. Cortical ROIs overlapping with the resection cavity are shown (red; left panels) together with the betweenness centrality for each cortical ROI displayed as a colourcoded map on the parcellated template brain (right panels). patients (1/22, 6%) underwent mesial temporal lobe resection. The broadband betweenness centrality distribution and the ROIs that overlapped with the resection cavity for all patients are displayed in Figure 1. The overlap between the ten highest broadband betweenness centrality ROIs and the resection cavity is given in Table 2, with the corresponding ROC curve in Figure 2. The Youden index indicated that the optimal threshold was at the th highest betweenness centrality value, meaning that the four ROIs with highest betweenness centrality should be used to indicate the area that overlaps with the epileptogenic zone. In this case, correspondence to the resection cavity was observed in 8/1 seizure-free patients and in 0/8 not seizure-free patients (57% 63

12 Chapter Table 2: Correspondence of resection cavity ROIs to the regions of interest with the ten highest betweenness centrality values in the broadband for seizure-free patients (white cells) and not seizure-free patients (grey cells). 6

13 Interictal MEG networks Figure 2: ROC curve for the overlap of the highest betweenness centrality ROIs and the resection cavity with thresholds that include an increasing number of considered ROIs, from one to ten ranked highest broadband betweenness centrality values. The Youden index was maximal (0.57) using a cut-off at the th highest value. sensitivity, 100% specificity, 73% accuracy). A high betweenness centrality corresponded significantly more often to the resection cavity in seizure-free patients compared to not seizure-free patients (p =.018). Out of the 1 seizure-free patients, six did not have high betweenness centrality in the resection cavity, but five of those patients had the maximum hub located in adjacent ROIs or in the contralateral lobe. The repeated analysis using the entire recording (16 epochs) is shown in Table S1 with the corresponding ROC curve in Figure S1. The Youden index indicated an optimal threshold of the 8th highest betweenness centrality values. For all epochs, high betweenness centrality values corresponded to the resection cavity in 12/1 seizure-free patients and in 3/8 not seizure-free patients (86% sensitivity, 63% specificity, 77% accuracy). High betweenness centrality in the theta and lower alpha band overlapped to a lesser extent with the resection cavity than for the broadband (Tables S2 and S3, with the corresponding ROC curves in Figures S2 and S3). The lowest optimal threshold for the 65

14 Chapter theta and lower alpha band was the 9th and th highest value, which yielded a sensitivity of 71% and 50%, specificity of 38% and 88%, and accuracy of 59% and 6%, respectively. The analyses for the other three spatial resolution levels were only performed for the broadband, since identification of the lobe of resection achieved highest accuracy in that band. On a lobar level, the optimal threshold was at the 2nd highest betweenness centrality value, which resulted in a correspondence to the lobe of resection in 9/1 seizure-free patients and in 0/8 not seizure-free patients (6% sensitivity, 100% specificity, 77% accuracy) (Figure S and Table S). A high betweenness centrality corresponded significantly more often to the lobe of resection in seizure-free patients than in not seizurefree patients (p =.006). Note that ROI level correspondence was possible without lobar level correspondence, as ROIs of neighboring lobes might partly fall into the resection cavity. On hemisphere level, the maximum betweenness centrality indicated the resection hemisphere in 10/1 seizure-free patients, whereas using the 7 and 9 ROIs with the highest betweenness centrality and choosing the majority of the lateralization of those ROIs (right versus left), corresponded to the resection hemisphere in 11/1 seizure-free patients (Table S5A). For temporal versus extra-temporal, the best performance was obtained for the maximum betweenness centrality, which corresponded to the resection location in 11/1 seizure-free patients (Table S5B). For the second analysis, a threshold of 0% was optimal to discriminate between seizure-free and not seizure-free patients with an accuracy of 73% (Figure 3). DISCUSSION The purpose of this study was to investigate the usefulness of network characteristics (in particular the betweenness centrality), determined through non-invasive interictal MEG recordings, in delineating the epileptogenic zone in patients with refractory epilepsy. We hypothesized that pathological hubs would be in or near the resection cavity that contains the epileptogenic zone and therefore that the resection cavity would have a high betweenness centrality in seizure-free patients. Our results support this hypothesis: high broadband betweenness centrality corresponded to the resection cavity and lobe of resection in seizure-free patients and not in patients who were not seizure-free after surgery. In the majority of seizure-free patients, we were able to identify the resection hemisphere and the resection area (temporal versus extra-temporal). The resection cavity displayed high betweenness centrality values in 57% (and in 86% when using all epochs) of the seizure-free patients. Previous studies that have used a similar approach found corroborating results, but mostly required invasive methods. For example, Wilke and colleagues noted increased betweenness centrality in the seizure onset zone based upon invasive grid recordings (Wilke et al., 2011). Varotto and co-workers 66

15 Interictal MEG networks Figure 3: ROC curve for the averaged betweenness centrality values of the ROIs that overlap with the resection cavity with thresholds using % of the maximum betweenness centrality value per patient. The Youden index was maximal (0.52) using a cut-off at 0% of the maximum betweenness centrality value, discriminating between seizure-free and not seizure-free patients with an accuracy of 73%. observed higher betweenness centrality in SEEG within epileptogenic lesions compared to outside epileptogenic lesions (Varotto et al., 2012). Furthermore, van Dellen and colleagues found that betweenness centrality decreased after surgery in areas close to the resection area in seizure-free patients, leading them to the conjecture that the removal of pathological hubs might result in seizure freedom (van Dellen et al., 201). Only one study found seemingly contradictory results: van Diessen and colleagues reported a decreased hub status within the seizure onset zone using SEEG in the hippocampus and amygdala (van Diessen et al., 2013b). However, SEEG has a higher spatial resolution than MEG virtual electrodes, but also a far more limited field of view as they cover a smaller area. It is therefore possible that they measured within the seizure onset zone, but not in the pathological hub area, leaving open the possibility that hubs are located near, but not 67

16 Chapter Figure : Schema c figure of a new concept about the presence of a pathological hub. A pathological hub is located in the proximity of the epileptogenic zone, with an established connection to the epileptogenic zone. Epileptiform activity might spread from the epileptogenic zone to the pathological hub and from there to the rest of the brain network. According to this concept, the pathological hub can be removed or disconnected to achieve seizure freedom in cases where resection of the epileptogenic zone is not possible. necessarily within, the epileptogenic zone. Indeed, in a recent MEG study we found the hub status of the irritative zone to be higher at the boundary than in the center (Nissen et al., 2016b). These studies, together with our results, indicate an association between network hubs and the epileptogenic zone. This supports a new concept about the presence of a pathological hub, located in the proximity of the epileptogenic zone, with an established connection to the epileptogenic zone (Figure ). Epileptiform activity, which originates in the epileptogenic zone, may spread to the rest of the epileptogenic network via this pathological hub. Resection of a pathological hub should lead to seizure freedom, provided that it can be reliably identified and is not part of eloquent cortex. Similarly, disconnecting the seizure onset zone and pathological hub may prevent spread of seizure activity. In fact, the current notion of epileptogenic zone (Lüders et al., 2006) is not confined to a zone or region, but includes a broader epileptogenic network. In this respect, the epileptogenic zone could be seen as consisting of 1) epileptogenic tissue (the focus ), 2) pathological 68

17 Interictal MEG networks connections, and 3) a pathological hub near, or in, the seizure onset zone and connecting the seizure onset zone to the rest of the brain. Consequently, the removal of any of those three components could be sufficient to achieve seizure freedom. If this concept holds true, it would have implications for patients in whom the epileptogenic tissue itself cannot be removed due to overlap with eloquent cortex or inaccessible location. These patients may still become seizure-free after resection of a pathological hub or pathological connections. Preoperative identification of the ROIs that fall within the epileptogenic zone and the lobe that contains the epileptogenic zone is a prerequisite for neurosurgery in refractory epilepsy. Information on a lower spatial resolution level, like determination of the correct hemisphere or of the temporal versus extra-temporal areas, is important for preoperative evaluation (Kharkar and Knowlton, 2015). We showed that MEG hubs may contribute to identification of the epileptogenic zone at a reasonable spatial resolution (i.e. at the resolution of ROIs in the AAL atlas). Additionally, we found that a high average betweenness centrality in the resection cavity discriminated seizure-free from not seizurefree patients. This may result in future preoperative resection simulations and calculation of the average betweenness centrality to optimise the surgical plan. A major advantage of our study lies in the non-invasiveness of MEG in combination with the use of virtual electrodes. Virtual electrodes allow an extensive coverage of the entire brain, including subcortical regions (Hillebrand et al., 2016a), with varying spatial resolution (from 1-20mm (Barnes et al., 200; Hillebrand and Barnes, 2005)). Hence, our approach did not require any surgical procedures while still allowing us to study the brain extensively. Another advantage is that the analysis was performed on interictal MEG recordings without epileptiform activity, which might provide an alternative method for localization as such MEG recordings are of no use with the current clinical techniques. High betweenness centrality was not found in the resection cavity in six (or in two when using all epochs) out of 1 seizure-free patients. A possible explanation for the negative findings in these patients is related to the concept described above. According to this concept, a surgical intervention that disconnects the epileptogenic tissue from a pathological hub would render patients seizure-free without removing the hub itself. In these patients, the pathological hub might be located outside the lobe of resection; however, the reason that the surgery was still successful could be that it eliminated the connections between the pathological hub and the epileptogenic tissue (or indeed the epileptogenic tissue itself). Interestingly, in five out of six patients without high betweenness centrality in the resection cavity, the maximum hub was located in adjacent ROIs or in homologous ROIs in the contralateral lobe, i.e. regions that are likely to be strongly connected to the resected 69

18 Chapter area. Even though we did not test the following hypothesis further, we speculate that the resection could have disconnected the pathological hub and the epileptogenic tissue in those cases. Similarly, in the one other patient the maximum hub was located posterior (occipital), which could be a physiological hub of the default mode network that was disconnected from the epileptogenic tissue during surgery. Limita ons The resection cavity often corresponded to a high betweenness centrality, but not necessarily with the maximum betweenness centrality. At the highest spatial resolution (ROI level), we obtained the highest specificity and accuracy using the four highest betweenness centrality values. The resection cavity did not include a hub in any of the eight patients who were not seizure-free, corresponding to a specificity of 100%. However, the high specificity was not replicated when using all epochs, which yielded a specificity of 63%. In order to apply the results to clinical practice, a method is needed that results in an unambiguous answer, e.g. where the maximum betweenness centrality indicates the area that needs to be resected. The maximum betweenness centrality correctly identified the resection hemisphere and temporal or extra-temporal areas in more seizure-free patients (10/1 and 11/1 patients, respectively) than it identified the lobe of resection or resection cavity ROIs (both 7/1 patients). The maximum betweenness centrality can therefore be used for identification of the epileptogenic zone at a reasonable spatial resolution (i.e. at the resolution of ROIs in the AAL atlas). The sensitivity is currently relatively low, hence the clinical relevance of our findings is still undetermined. Our method should therefore be refined to give a more consistent and reliable measure, for example by using other centrality measures, different characteristics of network topology (Rubinov and Sporns, 2010), or different connectivity measures (Engel et al., 2013a; Hillebrand et al., 2016b). Our results demonstrated that in seizure-free patients a pathological hub could be found in the ROIs that were at least partly overlapping with the resection cavity, however, there are several limitations: (1) in seizure-free patients, the resection cavity contained the epileptogenic zone, but it might be more extensive than the latter, (2) the epileptogenic zone does not follow the outlines of the AAL ROIs, and (3) the spatial resolution was too low to reveal whether the hub was located completely within, or just in proximity of, the resection cavity. Previous studies also lacked the spatial resolution (van Dellen et al., 201; Wilke et al., 2011) or spatial coverage (van Diessen et al., 2013b; Varotto et al., 2012) to make this distinction. A higher spatial resolution, but with whole brain coverage, could be achieved by using MEG virtual electrodes with a higher resolution atlas than the currently used AAL atlas, in combination with improved source reconstruction approaches (e.g. (Vrba et al., 2010)). However, localization at the ROI and lobar level already helps surgical 70

19 Interictal MEG networks planning, because current preoperative evaluation methods aim at localizing the epileptogenic zone to a lobe (PET, EEG, MEG, SPECT) before invasive methods with a higher spatial resolution (SEEG, ECoG) are used to refine the localization. Similarly, localization at the hemispheric and temporal/extra-temporal level helps to improve preoperative evaluation, particularly in case other modalities provide inconsistent findings. CONCLUSION The epileptogenic zone can be identified by estimating the betweenness centrality value in non-invasive interictal MEG recordings without epileptiform activity. Our results suggest that the resection cavity or lobe with the epileptogenic zone contains a hub, which has a pathological role in the epileptogenic network. This supports the concept of a pathological hub either within or in the vicinity of the epileptogenic zone, with an established connection to the epileptogenic zone during seizures. An implication of our results is that the removal of the pathological hub instead of the epileptogenic tissue or disconnecting these two regions might also lead to seizure freedom. This would be particularly beneficial in patients who were previously not eligible for surgery because of overlap of epileptogenic tissue with eloquent cortex or due to its inaccessibility. Further research is needed to increase consistency of the hub indications (i.e. maximum betweenness centrality instead of high betweenness centrality). In a prospective study containing patients for planned epilepsy surgery, betweenness centrality can be calculated for the planned resection cavity to predict seizure freedom, while the surgeons are blinded to the prediction. Our results are a first step towards a new method for a preoperative estimate of the location of the epileptogenic zone without relying on interictal epileptiform activity, thereby potentially improving surgical outcome with respect to attaining seizure freedom. ACKNOWLEDGEMENTS We would like to thank Hennie Evers for manually segmenting the resection cavities; Elvira Ruijter and Esther van Dam for preparatory work and testing of the software pipeline; Nico Akemann, Ndedi Sijsma, Karin Plugge, Marlous van den Hoek, and Peter- Jan Ris for MEG acquisitions; Matteo Demuru for help with Matlab; Marjolein Engels for verification of selected epochs; Giacomo Marseglia for help with the figures; Christiaan Bloeme and Nikki Thuijs for the Engel classification; Ingrid Moor for the preoperative evaluation information; and Dimitri Velis for the critical comments on the manuscript. I.A. Nissen is supported by the Dutch Epilepsy Foundation (project 1-16). The funding sources had no role in study design, data collection and analysis, interpretation of results, decision to publish, or preparation of the manuscript. 71

20 Chapter DISCLOSURE E.C.W. van Straaten is an independent contractor at Nutricia Research Utrecht, the Netherlands. A. Hillebrand received financial support from Elekta Neuromag Oy to attend a symposium. The remaining authors have no conflicts of interest. We confirm that we have read the Journal s position on issues involved in ethical publication and affirm that this report is consistent with those guidelines. 72

21 Interictal MEG networks: Supplementary data SUPPLEMENTARY MATERIAL Figure S1: ROC curve using all epochs for the overlap of the highest betweenness centrality ROIs and the resection cavity with thresholds that include an increasing number of considered ROIs, from one to ten ranked highest broadband betweenness centrality values. The Youden index was maximal (0.8) using a cutoff at the 8th highest value. Figure S2: ROC curve for the overlap of the highest theta band betweenness centrality ROIs and the lobe of resection with thresholds that include an increasing number of considered ROIs, from one to ten ranked highest betweenness centrality values in the theta band. The Youden index was maximal (0.11) using a cutoff at the 2nd, 3rd, and 10th highest value. 73

22 Chapter Figure S3: ROC curve for the overlap of the highest lower alpha band betweenness centrality ROIs and the lobe of resection with thresholds that include an increasing number of considered ROIs, from one to ten ranked highest betweenness centrality values in the lower alpha band. The Youden index was maximal (0.36) using a cut-off at the 2nd highest value. Figure S: Lobar level analysis: ROC curve for the overlap of the highest betweenness centrality ROIs and the lobe of resection with thresholds that include an increasing number of considered ROIs, from one to ten ranked highest broadband betweenness centrality values. The Youden index was maximal (0.6) using a cut-off at the 2nd highest betweenness centrality value. 7

23 Interictal MEG networks: Supplementary data Table S1: Correspondence of resection cavity ROIs to the regions of interest with the ten highest broadband betweenness centrality values using all epochs for seizurefree patients (white cells) and not seizure-free patients (grey cells). Correspondence: X, no correspondence: -. L: left, R: right. 75

24 Chapter Table S2: Correspondence of resection cavity ROIs to the regions of interest with the ten highest betweenness centrality values in the theta band for seizure-free patients (white cells) and not seizure-free patients (grey cells). Correspondence: X, no correspondence: -. L: left, R: right. 76

25 Interictal MEG networks: Supplementary data Table S3: Correspondence of resection cavity ROIs to the regions of interest with the ten highest betweenness centrality values in the lower alpha band for seizure-free patients (white cells) and not seizure-free patients (grey cells). Correspondence: X, no correspondence: -. L: left, R: right. 77

26 Chapter Table S: Correspondence of lobe of resection to the regions of interest with the ten highest betweenness centrality values in the broadband for seizure-free patients (white cells) and not seizure-free patients (grey cells). Correspondence: X, no correspondence: -. L: left, R: right. 78

27 Interictal MEG networks: Supplementary data Table S5: Correspondence of (A) resection hemisphere and (B) resection area (temporal versus extratemporal) with the indication given by the majority of 1,3,5,7,9, or 11 highest betweenness centrality values in the broadband for seizure-free patients. Correspondence: X, no correspondence: -. L: left, R: right. 79