The purpose of this dissertation is to develop a method to noninvasively localize the seizure onset zone (SOZ) in patients with drug-resistant epilepsy based on electroencephalographic (EEG) recordings.
Epilepsy is a neurological disorder characterized by unprovoked and recurrent seizures. Seizures are caused by abnormal electrical activity in the brain, which can lead to a wide range of behaviors ranging from subtle absences to jerking of the whole body. For 60-70% of the patients, anti-epileptic drugs allow adequate seizure control. The
remainder of the patients have so-called drug-resistant epilepsy. For them, the most efficient treatment option is epilepsy surgery during which the brain region responsible for the seizures is removed (resective surgery) or disconnected (disconnective surgery). Epilepsy surgery can only be performed when the region responsible for the seizures can be
delineated and when there is no overlap with functional tissue in order to avoid functional deficits, such as speech or motor problems, after surgery. Unfortunately, no single modality or unique technique allows measuring or identifying this epileptogenic focus directly.
Therefore, these patients undergo a presurgical evaluation, during which a team of experts tries to form a hypothesis about this region, based on the integration of the results of different investigations. The cornerstone investigations are long-term video-EEG monitoring and Magnetic Resonance Imaging (MRI). EEG is a technique which measures
the electrical activity of the brain with sensors or electrodes placed on the scalp. Since seizures (ictal) are caused by abnormal electrical activity, they result in abnormal EEG patterns. Yet, also in between seizures (interictal), short, abnormal EEG patterns can be observed. Based
on these abnormal EEG patterns, epileptologists can get a rough idea about the SOZ, i.e. the region where the seizures originate from, or the irritative zone (IZ), i.e. the region where the interictal activity originates from. MRI allows imaging the brain to reveal structural abnormalities such as lesions which could be responsible for the epilepsy. Often these
investigations do not allow forming a solid hypothesis about the exact epileptogenic focus and extra investigations such as SPECT, PET, MEG and/or invasive EEG (iEEG, EEG recorded with electrodes placed inside the brain) are needed.
This whole presurgical process is time-consuming, labor-intensive, occasionally riskful for the patient (brain surgery is needed for invasive EEG) and possibly subjective since human interpretation is required.
Furthermore, sometimes unambiguous delineation is still impossible. Therefore, it would be of high clinical value to have a method that is able to localize the responsible brain region in a more accurate, faster or automated, safe, and/or more objective way. In this dissertation, we aim to develop such a method for SOZ localization. The modality of choice is EEG since it is simple, safe, relatively inexpensive and the
most important tool to diagnose epilepsy.
Developing such a method is not trivial because several challenges need to be addressed. The two main problems are the low spatial resolution of the EEG and the fact that epilepsy is a network disease. The low spatial resolution of EEG is due to the fact that the brain activity is propagated
through the brain, skull, scalp and other tissues before it reaches the (often limited amount of) electrodes. The different conductivities of these tissues, and specifically the low conductivity of the skull, attenuate and distort the brain signals. As a consequence, the signal at a given
electrode does not necessarily represent the activity of the directly underlying brain area, but is distorted and mixed with the activity of other brain areas. In this dissertation, we will tackle this problem in two ways. First, the number of electrodes can be increased with respect to standard EEG, resulting in so-called high-density EEG. Second, EEG
source imaging (ESI) will be used. ESI is a technique that estimates the brain activity that generated the measured EEG. It consists of a forward model and an inverse problem. The forward model calculates how a certain source of activity in the brain will translate to an EEG measurement. In this dissertation, this forward model will be based on
the MRI of the patient. The inverse solution tries to find the generating sources by minimizing a cost function based on the difference between the generated EEG by the forward model and the actually measured EEG.
By applying ESI on an EEG segment containing ictal activity, the brain regions active during the seizure can be identified. One way to estimate the SOZ is to select the most active brain region during the seizure.
However, and this is the second problem, epilepsy is a network disorder rather than a focal disease. This means that during an epileptic seizure, different brain regions become active as a part of an epileptic network.
There is no guarantee that the most active region during a seizure is the one causing the seizure, because a small group of neurons (associated with a less active region) could be triggering a larger group (associated with a more active region). To overcome this problem, we will analyze the epileptic network to find the driving region behind it. This can be
achieved by functional brain connectivity, the study of the interactions and information flows between brain regions. In this dissertation, the connections or information flows between the active brain regions found after ESI are calculated using a time-varying multivariate autoregressive (TVAR) model in which the brain signals associated with the regions are modeled as a linear combination of their past samples plus uncorrelated
white noise. This way, it can be investigated how the past samples of one signal influence the current samples of the other signals to assess the causality between these signals. The coefficients of the TVAR model are transformed to the frequency domain which results in a time-varying transfer matrix from which the spectrum weighted Adaptive Directed
Transfer Function (swADTF) can be calculated. The swADTF is a measure of the information flow between two brain regions and stresses connections with high spectral power in both the receiving and the sending signal. Finally, the region with the highest outgoing information flow or swADTF values was used as an estimation for the SOZ.
In a first study, we verified this technique using simulations. We
generated an epileptic network consisting of three nodes in the brain,added background activity to obtain a signal-to-noise ratio of 5 dB and used a forward model to generate high-density (204 electrodes) EEG measurements of 3 s. ESI was applied on 1000 of these simulated EEG segments and the most active brain sources were selected. Connectivity
analysis as explained above was performed on the neuronal correlates of these selected brain sources and the source with the highest outgoing information flow was selected as estimation for the SOZ. The distance between the estimated SOZ and the true driver of the epileptic network, which we know from the simulation, was used to assess the quality of
the estimation. To make a benchmark comparison, we compared this distance to the distance between the most active region after ESI and the true driver. Our findings were threefold. First, we found that ESI followed by connectivity analysis (ESI + connectivity) is better than selecting the most active brain region after ESI (ESI power) to estimate the SOZ. Second, the median distance to the true driver was small, 12 mm. Third, when fewer electrodes were used (only a subset of
the high-density EEG was used for ESI and connectivity analysis), the distances increased, up to a median distance of 21 mm for 32 electrodes.
In the second phase of this study, we validated the method in
retrospective data of 5 patients. These patients ultimately underwent successful epilepsy surgery, so we know that the true SOZ was inside the tissue that was resected. In each of these patients, one seizure was preoperatively recorded using high-density EEG with 204 electrodes. The first three seconds of the seizures were analyzed and connectivity analysis
was done in the wide 1 to 30 Hz band, in order to be certain that the seizure frequency band was included. The estimated SOZ was compared with the resected zone. We confirmed that ESI + connectivity performs better than ESI power. ESI + connectivity was able to localize the SOZ within 10 mm of the border of the RZ in all patients, whereas ESI power could do this in only 2/5 patients. When lowering the number
of electrodes used for analysis, we found again that the performance decreased. ESI + connectivity was able to localize the SOZ within 10 mm in only 1/5 patients, using 32 electrodes. We concluded that the proposed method can correctly localize the SOZ, given that the seizure is recorded with high-density EEG. Although more validation is
needed, we suggested that it could become a useful tool in the presurgical evaluation of epilepsy.
Unfortunately, most clinics still lack the equipment to perform highdensity EEG and it is usually not part of the default presurgical evaluation. Furthermore, long-term high-density EEG recordings are rare, since they are uncomfortable for the patient. Therefore, the chance that a patient has a seizure during a high-density EEG recording, is rather low.
For these reasons, we adapted the proposed method in a second study. Instead of strictly selecting the first 3 s of the seizure, we selected, together with an expert epileptologist, a (quasi) artifact-free epoch during the beginning of the seizure, lasting 1-5 s. Furthermore, connectivity analysis was limited to the frequency band of the rhythmic activity in the EEG generated by the seizure. We validated this adapted
approach retrospectively in 111 seizures of 27 patients (24 temporal lobe epilepsy, 3 extratemporal lobe epilepsy) who were rendered seizure-free after surgery. Again, we found that ESI + connectivity outperformed ESI power, this time statistically significant. ESI + connectivity resulted
in a SOZ estimation within 10 mm of the border of the resected zone in 93.7% of the seizures. We confirmed our conclusion of the previous study that this method could serve as a useful tool in the presurgical evaluation, but now the standard long-term EEG monitoring can be used and seizures recorded with high-density EEG are not necessary.
However, user-dependent input for the initial epoch and frequency selection is required. Larger studies are needed, notably with more extratemporal epilepsies and localization correlation with a range of different surgery outcomes.
In a final study, we investigated the importance of epoch and frequency band selection. In terms of epoch selection, best results were obtained when an artifact-free epoch was selected during the electrographic onset phase of the seizure, before the ictal patterns have evolved. Furthermore, we found that it is impossible to obtain a trustworthy SOZ estimation with this method during the preictal or postictal period. In terms of the
frequency band, best results are obtained when the analysis is limited to the seizure frequency band corresponding to the analyzed ictal epoch.
Although performance was lower when using the 1 to 30 Hz band, the difference did not reach significance and this band can possibly be used for analysis whenever the seizure frequency band is unclear.
In conclusion, we developed a method that combines ESI and functional connectivity analysis to noninvasively localize the SOZ based on ictal EEG recordings. The method is safe and proved to be accurate.
Given more research, it could be made completely user-independent and automated. Yet, more validation, certainly in more heterogeneous population groups is necessary. Altogether, the method could serve as
a useful and accurate tool in the presurgical evaluation of epilepsy