Ictal-onset localization through connectivity analysis of intracranial EEG signals in patients with refractory epilepsy
Pieter van Mierlo,
Medical Image and Signal Processing Group, Department of Electronics and Information Systems, Ghent University – iMinds, Ghent, Belgium
Address correspondence to Pieter van Mierlo, Medical Image and Signal Processing Group, Department of Electronics and Information Systems, Ghent University – iMinds, De Pintelaan 185, Block B, 5th floor, B-9000 Ghent, Belgium. E-mail: Pieter.vanMierlo@UGent.be
Fifteen percent to 25% of patients with refractory epilepsy require invasive video–electroencephalography (EEG) monitoring (IVEM) to precisely delineate the ictal-onset zone. This delineation based on the recorded intracranial EEG (iEEG) signals occurs visually by the epileptologist and is therefore prone to human mistakes. The purpose of this study is to investigate whether effective connectivity analysis of intracranially recorded EEG during seizures provides an objective method to localize the ictal-onset zone.
In this study data were analyzed from eight patients who underwent IVEM at Ghent University Hospital in Belgium. All patients had a focal ictal onset and were seizure-free following resective surgery. The effective connectivity pattern was calculated during the first 20 s of ictal rhythmic iEEG activity. The out-degree, which is reflective of the number of outgoing connections, was calculated for each electrode contact for every single seizure during these 20 s. The seizure specific out-degrees were summed per patient to obtain the total out-degree. The electrode contact with the highest total out-degree was considered indicative of localization of the ictal-onset zone. This result was compared to the conclusion of the visual analysis of the epileptologist and the resected brain region segmented from postoperative magnetic resonance imaging (MRI).
In all eight patients the electrode contact with the highest total out-degree was among the contacts identified by the epileptologist as the ictal onset. This contact, that we named “the driver,” always laid within the resected brain region. Furthermore, the patient-specific connectivity patterns were consistent over the majority of seizures.
In this study we demonstrated the feasibility of correctly localizing the ictal-onset zone from iEEG recordings by using effective connectivity analysis during the first 20 s of ictal rhythmic iEEG activity.
In approximately 15–25% of patients with refractory epilepsy included in the presurgical evaluation protocol (Engel et al., 1992; Boon et al., 1999), invasive video–electroencephalography (EEG) monitoring (IVEM) is required to identify the ictal-onset zone (Carrette et al., 2010). However, the visual analysis and spatial interpretation of many simultaneously recorded intracranial EEG (iEEG) signals by the epileptologist has its limitations. Precise identification of initial, sometimes discrete, iEEG changes to determine the ictal-onset zone is time-consuming and requires extensive and specific expertise. The purpose of this study is to investigate whether effective connectivity analysis of intracranially recorded seizures provides an objective method to localize the ictal-onset zone. This may aid and support decision making on visual ictal-onset localization by the epileptologist in clinical practice.
In this study, the localizing value of a mathematical analysis technique (Van Mierlo et al., 2011) is investigated to overcome the limitations of a visual identification strategy. The proposed technique is more objective due to its quantitative nature. It investigates the causal relations in the spectral domain between the iEEG signals, making use of the concept of effective brain connectivity.
Effective connectivity is defined as the influence that one neural system exerts over another, either directly or indirectly (Friston et al., 1993a). In contrast to functional connectivity, that is, the temporal correlations between remote neurophysiologic events (Friston et al., 1993b), it describes the directionality of interactions between brain regions. These interactions can be modeled a priori and analyzed afterward with structural equation modeling (Tomarken & Waller, 2005) or dynamic causal modeling (Friston et al., 2003), or they can be derived from a data-driven technique such as Granger causality measures (Granger, 1969).
Many studies have been performed using functional connectivity to assess the epileptogenic network. Already in 1972, coherence analysis of iEEG was used to characterize spreading of epileptic discharges within the brain of patients with epilepsy (Brazier, 1972). Gotman successfully designed a method based on the coherence and phase spectra to make inferences on the possible routes of propagation of seizure activity (Gotman, 1983). This method was also used to study synchronization mechanisms, particularly at the onset of seizures (Duckrow & Spencer, 1992). The coherence was also used to successfully classify subtypes of temporal lobe epilepsy (Bartolomei et al., 1999). It has been shown that during seizures the neuronal network moves in the direction of a more ordered configuration compared to the more randomly organized interictal network (Ponten et al., 2007). For a detailed overview of the use of functional connectivity measures in epilepsy specifically during ictal iEEG epochs we refer the reader to Wendling et al. (2010).
A commonly used effective connectivity measure in epilepsy to study ictal networks is the nonlinear correlation coefficient coupled to the direction index (Bartolomei et al., 2001; Wendling & Bartolomei, 2001; Bartolomei et al., 2005). It shows both the strength and the direction of the connections between two signals. These studies have demonstrated the existence of both generic and organized networks involved during temporal lobe seizures. However, it is a bivariate measure, meaning that each pairwise combination of channels requires separate investigation. This can be cumbersome when many signals are considered simultaneously.
Multivariate effective connectivity measures were designed to model all signals in the same system. The directed transfer function (DTF; Kaminski & Blinowska, 1991) is a measure commonly used in epilepsy. The DTF has been used to localize the epileptogenic focus during stationary seizure epochs out of the iEEG (Franaszczuk et al., 1994; Franaszczuk & Bergey, 1998; Wilke et al., 2010; Jung et al., 2011; Wilke et al., 2011). Another multivariate effective connectivity measure is the partial directed coherence (PDC; Baccala & Sameshima, 2001). It has been used to localize the epileptogenic focus in epilepsy secondary to type II focal cortical dysplasia (FCD; Varotto et al., 2012). However, the DTF and PCD require that the investigated signals are stationary during the considered epoch, whereas seizure onsets are intrinsically nonstationary.
The time-variant version of the DTF named the adaptive directed transfer function (ADTF) was designed to cope with nonstationary signals (Astolfi et al., 2008; Wilke et al., 2008). The localizing value of the ADTF has been investigated based on interictal iEEG data (Wilke et al., 2010, 2011). We developed a modified version of the ADTF, namely the full-frequency adaptive directed transfer function (ffADTF; Van Mierlo et al., 2011), to estimate the effective connectivity pattern during seizure activity. Initial testing of the ffADTF during simulated seizures resulted in a sample-based sensitivity and specificity of 98%. The analysis technique was successfully applied to the iEEG of one single patient in a proof-of-concept study (Van Mierlo et al., 2011).
In this study we clinically validate our previously developed methodology. We localize the ictal-onset zone in a series of eight patients based on iEEG data. We use the effective connectivity patterns of the seizure onset to accurately localize the ictal-onset zone. The results are compared to conclusions of the visual analysis of the IVEM performed by the epileptologist, the surgically removed brain region on magnetic resonance imaging (MRI), and seizure outcome.
The general concepts of the analysis technique have been described previously in Van Mierlo et al. (2011). The alterations made to the previously developed method are described below.
Time-variant effective connectivity
A time-variant multivariate autoregressive (TVAR) model is built from the iEEG signals by using the Kalman filtering algorithm (Arnold et al., 1998; Schlögl, 2000). The time-variant connectivity measure, the ffADTF, is calculated from the coefficients of the TVAR model as follows:
where Hij(f, t) is the time-variant transfer matrix of the system describing the information flow from signal j to signal i at frequency f at time t. While applying the ffADTF to the seizures of all the patients we noticed that the term Hij(f, t) can be high even when there is no power in the spectrum of signal j at that frequency and time. These new insights allowed us to alter/improve the ffADTF's formula. Each term Hij(f, t) should be weighted by the autospectrum of the “sending (in this case j)” signal. The modified effective connectivity measure based on Granger causality is called the spectrum-weighted adaptive directed transfer function (swADTF):
The swADTF allows us to investigate the causal relation between all the signals at a predefined frequency band over time. The measure weighs all outgoing information flow present in the terms Hij(f, t) by the power spectrum of the “sending” signal j, namely . The swADTF is normalized so that the sum of incoming information flow into a channel at each time point is equal to 1:
Each signal can be represented as a node in the graph. Each swADTF value corresponds to the directed time-variant strength of the information flow between two nodes. We assume that there is a directed edge between two nodes if the corresponding swADTF value is higher than a predefined threshold:
where thij is the predefined threshold for the connection from signal j to signal i. The matrix Cij(t) is the connectivity matrix and represents the directed edges between all signals over time. All self-edges are set to 0 at each time point t:
The out-degree of each node, that is, the number of outgoing edges from that node, is calculated from the connectivity matrix. The out-degree of node J (corresponding to signal xj) at time t is noted as φJ (t):
The sum of the out-degree of node J over the time points of a time interval [t1, t2] depicts the number of outgoing connections of node J during that time interval:
The iEEG dataset was obtained from a series of eight patients included in the presurgical evaluation protocol at the Ghent University Hospital Reference Center for Refractory Epilepsy in Belgium. The study was approved by an ethics committee, and all patients signed informed consent.
The patients included in the study were selected based on the following criteria: focal ictal onset on IVEM and subsequent resective surgery rendering the patient seizure-free (Engel class I outcome [Engel, 1993]) during a minimum follow-up of 2 years. The clinical patient characteristics are described in Table 1.
Table 1. Clinical patient characteristics
The columns represent: the patient's number, sex, age at the time of IVEM, type of epilepsy, ictal onset defined based on scalp EEG recording, notable MRI findings, type of surgery, surgical outcome, and duration of follow-up since the surgery (in years).
The details of the IVEM are found in Table 2. The iEEG was recorded using a monopolar reference electrode located on the right mastoid and with a ground electrode located on the left mastoid.
Table 2. Details of the IVEM data
No. electrode contacts
Strip and grid electrodes
Omitted electrode contacts
Sampling frequency (Hz)
The columns represent the following: the patient number; the number of analyzed seizures; the number of electrode contacts; the implanted depth electrodes; the implanted subdural strips and grids; the electrode contacts omitted due to visually identified artifacts; sampling frequency; available imaging of implanted electrodes; and available postoperative imaging.
LH1-10: left hippocampus
LG1-40: left gyrus angularis
CT and MRI
RH1-8: right hippocampus
RATB1-4: right temporoanterior
RPTB1-4: right temporoposterior
RTL1-6: right temporolateral
LTL1-6: left temporolateral
LATB1-4: left temporo anterior
LPTB1-4: left temporo posterior
RATB1 and LATB4
CT and MRI
LH1-8: left hippocampus
LFFA1-6: left basal posterior temporal
LTL1-6: left lateral temporal
LTA1-4: left basal anterior temporal
RTA1-4: right basal anterior temporal
RFFA1-6: right basal posterior temporal
CT and MRI
LA0-3: left amygdala
LH0-3: left hippocampus
RA0-3: right amygdala
RH0-3: right hippocampus
LG1-32: left temporo posterior
CT and MRI
LA0-3: left amygdala
LH0-3: left hippocampus
RA0-3: right amygdala
RH0-3: right hippocampus
LATS1-6: left temporo anterior
RATS1-6: right temporo anterior
LPTS1-4: left temporo posterior
RPTS1-4: right temporo posterior
RD1-12: insular lesion
TG1-32: right temporal lobe
SSG1-16: right suprasylvian
CT and MRI
LH1-12: left hippocampus
RH1-12: right hippocampus
LTA1-4: left temporo basal
LTM1-4: left temporo medial
LTP1-6: left temporo lateral
RTA1-4: right temporo anterior
RTM1-4: right temporo medial
RTP1-6: right temporo lateral
RH5-6 and RH11-12
CT and MRI
LD1-8: left hippocampus (LD1-4) and trough heterotopy (LD7-8)
LG1-56: temporal and parietal (LG1-8 lateral temporal and LG49-56 parietal)
Localization of the ictal-onset zone
From iEEG to out-degree
The iEEG of all the habitual seizures in all eight patients was selected. The starting point for the effective connectivity analysis was the moment where epileptiform rhythmic iEEG activity, with a frequency between 3 and 20 Hz, was observed at the onset of the seizure.
For each patient we selected the ictal iEEG from 5 s before until 20 s after the marked starting point. The epochs were filtered (low-pass filter [0.5, 40 Hz], with a notch at 50 Hz) and normalized (mean set to 0 and standard deviation to 1). The normalization is required to treat every signal equally and in this way avoid biasing the results based on the amplitude of the signals. Afterward the seizures were resampled from 200 to 100 Hz and from 512 and 256 to 128 Hz with a specific resampling procedure for time-variant autoregressive models (Schlögl, 2000). The resampling allows the use of a smaller model order while maintaining the time window used to estimate the coefficient matrices. The resampling also allows for coping with computational memory issues and resulted in a significant decrease of analysis time. For each resampled seizure the TVAR coefficients were estimated with an update coefficient (UC) equal to 10−3 and the model order p equal to 8 and 10 for 100 and 128 Hz, respectively. A detailed discussion on the parameters can be found in Van Mierlo et al., 2011. We discarded the coefficient matrices from the first 5 s before the marked starting point. This 5-s time frame was used only to allow the Kalman filter to adapt to the data.
Out of the TVAR coefficients, the swADTF was calculated in the frequency band (3, 40 Hz) based on eqn 2. This band is chosen to remove the background electrical activity up to 3 Hz. Due to the 1/f nature of the spectrum of the iEEG we tried to limit the influence of background activity by setting the lower bound of the frequency interval as high as possible. The upper bound of the frequency interval was set to 40 Hz based on the applied low-pass filter. All ictal rhythmic seizure activity was observed within the chosen frequency band.
The swADTF values were thresholded by a connection specific threshold based on connectivity analysis of interictal segments. For each patient we selected 10 interictal epochs of 25 s and processed them in the same manner as the ictal epochs. We derived the 99th percentile for each connection between each pair of electrode contacts and used this as the threshold for that specific connection. In this way we expect to see only 1% of connections during the analysis of an iEEG epoch.
The connectivity matrix of each seizure was calculated using these thresholds according to eqn 4. The time-variant out-degree of each node was derived from this connectivity matrix in each patient according to eqn 6. The seizure specific out-degree (SSO) is the sum of the out-degrees over the 20 s of each seizure epoch according to eqn 7. This depicts how many out-going connections arise from each contact during the considered seizure. The total out-degree is the sum of the SSO over all seizures per individual patient. The higher this total out-degree of one electrode contact, the more connections arose from this contact during all analyzed seizures.
The outcome parameters
We compared the electrode contact with the highest total out-degree with the electrode contacts showing ictal onset based on the visual identification of the iEEG performed by the epileptologist (AM, KV, PB). Furthermore, we also compared the location of the electrode contact with the highest total out-degree with respect to the resected brain area derived from the postoperative MRI. The SSOs are analyzed to evaluate whether the results are consistent over multiple seizures.
The total out-degree results of all patients are depicted in Table 3. In all eight patients the electrode contact with the highest total out-degree was among the electrode contacts that, visually defined by the epileptologist, showed the ictal onset on iEEG. The electrode contact with highest total out-degree also laid within the resected brain region in all patients in whom postoperative images were available (n = 7/8). We can conclude that the localization based on our method corresponded in all patients with the visual analysis.
Table 3. Seizure-specific results of all patients
Highest total out-degree
SSO seizure 1
SSO seizure 2
SSO seizure 3
SSO seizure 4
SSO seizure 5
SSO, seizure specific out-degree.
The three electrode contacts with the highest out-degree for each seizure individually are shown. The columns represent the following: the patient number, the result of the visual analysis of the iEEG, the electrode contacts lying in the resected area (if postoperative images are available), the highest total out-degree, and the seizure specific out-degree for the individual seizures.
The highest SSOs are shown in Table 3. The highest SSO was consistent over multiple seizures. In the majority of seizures from one individual, the highest SSO came from one specific or immediately neighboring contacts. In 62% of all seizures, the SSO was equal to the highest total out-degree. In 33%, the highest SSO was a neighboring contact (maximum distance of two contacts) of the one with the highest total out-degree. We can conclude that total out-degree corresponds well with the SSO. This means that most seizures of the same patient had the same driving brain region during the analyzed epochs.
Individual patient results
The SSO and the total out-degree of all patients are shown in supplementary Figure S1 and Figure S2. Below we discuss the results of patient 2 in more detail as an illustrative case.
In patient 2 the MRI showed a dysembryoplastic neuroepithelial tumor in the right inferior temporal gyrus, and the scalp EEG showed bilateral frontal temporal ictal onset. Because of these incongruent findings the patient underwent IVEM. A depth electrode was implanted in the right hippocampus and three strips covered the right and left temporal lobe. During the IVEM, three habitual seizures were recorded that started from contact RPTB4 (right temporal lobe strip) according to the visual analysis of the epileptologist. A right anterior temporal lobectomy and amygdalohippocampectomy rendered the patient seizure-free (follow-up, 3 years).
In Figure 1 the results of the connectivity analysis of patient 2 are displayed. The upper panel shows the electrode positions and how they were located with respect to the resected area. Electrode contact RH1-3, RATB1-4, and RPTB1-4 laid within the subsequently resected area. The middle panel shows the connectivity analysis of the first seizure. The seizure is divided into four windows of 5 s (A, B, C, and D) for visualization purposes only. During the first 10 s most connections arose from contact RPTB4. During time window C from RPTB4 and RATB4 and during window D from RATB4.
The lower panel shows the SSO and the total out-degree. We can clearly observe that RPTB4 had the highest SSO during all three seizures. This resulted in the highest total out-degree for this contact, which corresponded with the visual analysis of the epileptologist and the subsequently resected brain area.
In this study we investigated a new method based on effective connectivity to localize the ictal-onset zone from epileptiform ictal rhythmical activity in the iEEG. The results corresponded to the ictal-onset zone as defined by the visual analysis of the iEEG by the epileptologist and to the resected area. The method investigated the causal relations between the signals without taken any form of prior knowledge about the electrode contact locations into account. This framework is user independent and can be applied readily in a clinical setting. In an initial stage this method can support epileptologists' ictal-onset localization and may significantly reduce the time spent to search for the earliest visually identifiable changes in the numerous simultaneously recorded channels observed one dimensionally but reflecting three-dimensional brain regions.
The method uses a mere 20 s of iEEG, starting from when ictal rhythmical activity, between 3 and 20 Hz, was seen at multiple electrode contacts at the beginning of a seizure. We did not look at the patterns corresponding with low amplitude rapid discharges above 25 Hz, during which a decorrelation between the brain regions was found (Wendling et al., 2003). In this article we showed that during the first 20 s of epileptiform ictal rhythmic activity, we are able to identify the electrode contact at which the first signs of ictal activity were observed. This means that the earliest spreading pattern contains information about the seizure initiator. The analysis of the 20 s of rhythmic activity identifies the same electrode contacts as the analysis of hours of iEEG by the epileptologist.
Other connectivity studies based on Granger causality (Franaszczuk et al., 1994; Franaszczuk & Bergey, 1998; Wilke et al., 2010; Jung et al., 2011; Wilke et al., 2011) localized the ictal-onset zone on the basis of stationary ictal iEEG epochs. The brain region “driving” the seizure during a specific stationary epoch is not necessarily the region where the seizure started. There is also no guarantee that the signals were stationary in the analyzed window. In our study we used an effective connectivity measure (the swADTF) capable of analyzing nonstationary spread of ictal rhythmic activity. Because the onset of the spread is nonstationary, our method is preferred to analyze these epochs.
Wilke et al. (2010, 2011) showed promising results in localizing the primary sources of interictal spikes from a patient with epilepsy based on a nonstationary multivariate connectivity measure. Although the localization of interictal epileptic events has proven to be valuable, some studies (Alarcon et al., 1997; Hufnagel et al., 2000; De Curtis & Avanzini, 2001) have concluded that the brain region showing interictal activity (the irritative zone) is larger than the area that is involved in seizure generation. In this study we considered only the ictal epileptic activity and did not take into account the interictal epileptic discharges, because accurate localization and determination of the ictal networks is the prerequisite for defining the subsequent therapeutic strategy, precisely aimed at suppressing seizures by the annihilation of epileptogenic networks (Wendling et al., 2009).
The seizure-specific analysis of individual patients led to comparable results over the majority of seizures. This is concordant with the finding of Gnatkovsky et al. (2011), who were able to identify reproducible ictal patterns in patients based on quantified frequency analysis of intracranial iEEG signals. In this study we can extend these findings to connectivity patterns, meaning that the connectivity patterns were consistent over most seizures in all patients. This could be an indication that most seizures of the considered patients started within the same brain area. We suspect this explains why all patients are seizure-free after resection of the ictal-onset zone. Successive studies in patients with unsuccessful surgery outcome are needed to determine whether there is more variability in the connectivity patterns that may explain why they were not rendered seizure-free.
The framework presented herein was designed with the sole purpose of localizing the ictal-onset zone based on the origin of the epileptiform rhythmic activity during a seizure. How well other connectivity measures such as the nonlinear correlation coupled with the direction index (Wendling & Bartolomei, 2001) or nonlinear granger causality measures (Marinazzo et al., 2011) would perform in localizing the ictal onset still needs to be assessed. However, these measures are bivariate, meaning that all pairwise combinations need to be investigated and a new framework that defines the ictal onset-zone based on these connections needs to be constructed. This will require more advanced graph analysis measures, such as the betweenness centrality (Sporns et al., 2007).
Our method is obviously limited by the spatial resolution and sampling of the implanted electrodes. Due to the difficulties in obtaining broad cortical coverage by means of implantable electrodes, it is not unlikely that invasive electrodes do not cover the entire ictal-onset zone (Rosenow & Luders, 2001).
In addition to seizure freedom, there is an increasing interest for more global outcome parameters such as cognitive and psychiatric outcomes and health-related quality of life (Spencer & Huh, 2008). A more precise and accurate localization of epileptogenic tissue could allow for resection of even smaller brain regions, thereby preserving more function. This has already been shown in studies comparing lobectomy to selective amygdalohippocampectomy, where most authors found similar seizure outcome and evidence for better neuropsychological outcome (Schramm, 2008).
How ictal-onset zone delineation based on connectivity could affect the cognitive and psychiatric outcomes as well as the health-related quality of life of the patient after resective surgery still needs to be assessed. The connectivity measures have the potential to more precisely localize the ictal onset with respect to the visual analysis of epileptologist and could therefore affect the outcome parameters. However, there will always be a tradeoff between resecting a broad area that certainly encompasses all epileptogenic tissue and a smaller resection that preserves more functionality with the higher chance to leave residual epileptogenic tissue (Schramm, 2008). Can a more limited resection yield seizure-freedom rates similar to those afforded by wider/more aggressive resection? If not, is the quality of life better or worse than that with a wider resection that increases seizure freedom rate but yields a higher complication rate? One thing we do know is that as imaging and mathematical techniques advance, limited resection will become the more attractive option (Okonma et al., 2011).
How connectivity measures can play a role in planning the extent of resection needs to be assessed. The potential of the connectivity analysis is promising, but studies using resection strategies based on effective connectivity measures need to be conducted. The potential benefits include the ability of connectivity analysis to pinpoint to a part of a structure (e.g., anterior hippocampal region), allowing a smaller resection or even ablation around the “driving” electrode contact. Herein we need to keep in mind the poor spatial sampling of the implanted electrodes. Another potential benefit is the mapping of the spreading pattern of seizure activity that could guide subpial transections. Increased spatial resolution identification of the ictal-onset zone based on effective connectivity analysis could also be used in the more rational design of deep brain simulation protocols to treat epilepsy. Even if surgery is not an option, a deep brain stimulation system could be developed where stimulation pulses are delivered to the electrode contact depicted as driver by the connectivity analysis.
However, many clinical and preclinical studies are needed to understand the added value of the connectivity measures over the visual analysis of the epileptologist. There is a growing need for studies showing how connectivity analysis can be used in clinical practice to ultimately ameliorate the quality of life of patients with refractory epilepsy.
In this article, the applicability of effective connectivity analysis to localize the ictal-onset zone from iEEG recordings during early ictal rhythmic activity was investigated in a series of eight patients. We showed that the ictal-onset zone defined by our method corresponded with the results of the standard visual analysis performed by the epileptologist in all eight patients. Moreover, our results also corresponded with the resected brain tissue that rendered the patients seizure-free. The framework does not require any type of a priori knowledge and can be used readily in clinical practice.
Research funded by a PhD grant of the Institute for the Promotion of Innovation through Science and Technology in Flanders (IWT-Vlaanderen). Prof. Dr. K. Vonck is supported by a BOF-ZAP grant from Ghent University Hospital. Prof. Dr. P. Boon is supported by grants from FWO-Flanders; grants from BOF; and by the Clinical Epilepsy Grant from Ghent University Hospital. Robrecht Raedt is supported by the special research funds of Ghent University. The authors would like to thank Prof. Karel Deblaere for providing the medical images and Prof. Daniele Marinazzo for the fruitful discussions.
None of the authors have any conflict of interest to disclose. 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.