Patient Selection
In this retrospective study we included patients who (i) had a previous positive investigation of scalp EEG FOs (Andrade-Valença et al.
2011; Melani et al.
2013), and (ii) underwent a separate simultaneous EEG-MEG 1-h recording session. Seventeen patients fulfilled the inclusion criteria. The assessment of scalp EEG FOs in nine of them (Patients 1–5 and 10–13) was in the context of a project in which the inclusion criterion was focal epilepsy with at least one IED per minute during a preliminary scalp EEG study (Andrade-Valença et al.
2011). The remaining eight patients belonged to a project in which the only inclusion criterion was focal epilepsy (Melani et al.
2013). Demographics and clinical information are given in Supplementary Table 1. This study was approved by the Montreal Neurological Institute Research Ethics Board and all patients signed a written informed consent prior to the study.
Epileptogenic Region
To define a gold standard for the source localization results, two specialists defined the epileptogenic region based on the available clinical information for each patient. This information is included in Supplementary Table 1, and consisted of (in order of priority; not all factors were available for every patient): resected region, ictal and interictal iEEG findings, visible lesion in the MRI, ictal and interictal scalp EEG findings. The epileptogenic region for each patient was independently marked by both reviewers, and a consensus was reached by the experts after discussing the cases in which there were differences (all differences were restricted to the extent of the epileptogenic region, with no patients showing discordant regions). The reviewers were blind to the source localization results at the moment of marking the epileptogenic zone. The marking was done over the inflated brain derived from each patient’s MRI, visualized in 6 orientations.
MEG Data Acquisition and Pre-processing
Simultaneous EEG/MEG recordings were performed at the Montreal Neurological Institute, McGill University and at the Department of Psychology, Université de Montreal using a 275 sensor CTF-MEG-system (MISL, Vancouver, Canada) and a 56 sensor EEG-cap with ceramic electrodes (Easy-cap, Herrsching, Germany). The sampling rate was 1200 Hz in one patient and 2400 Hz in the remaining patients. The duration of the recordings was constrained by each patient’s ability to stay still in the dewar. The data was acquired in blocks of 6 min duration, with a median of 9 blocks per patient (range 6–11), i.e. median total recording time of 54 min (range 36–66 min). Since this project aimed to characterize interictal FOs, we excluded the blocks containing ictal activity from two patients who had seizures during the MEG acquisition. All the recordings were downsampled to 640 Hz. The technical details of the data processing can be found in
Appendix.
Automatic Detection
The MEG automatic FOs detector was based on an algorithm developed for scalp EEG (von Ellenrieder et al.
2012). It searched for an increase in the root mean square (RMS) amplitude of the signal in narrow frequency bands, and compared it to the background in a 5 s sliding window. A detection occurred when the RMS amplitude in any narrow frequency band was at least 2.5 times larger than the background during an interval longer than 4 cycles of the central frequency of the band plus the effective duration of the narrowband filters impulse response. There were ten narrow bands in the 40–160 Hz range (40–46–53–61–70–80–93–107–123–142–160 Hz), five in the high-gamma band and five in the ripple band. The filter details are given in
Appendix. The automatic detector was set to work with high sensitivity in order to avoid missing any true FO. Since the high sensitivity implies low specificity, a high proportion of false positives was expected at this step. We will refer to the candidate FOs found in this first step as
pre
-
detections. To exclude muscle activity and movement artefacts, all pre-detections that fell in one of the following three categories were discarded: (i) Pre-detections that occurred simultaneously in more than 200 narrow frequency bands and channels, or in more than 100 channels. This criterion relies on the fact that muscle activity and movement artefacts usually involve many channels and have broad frequency content (Jeong et al.
2013). (ii) If the channel with highest rate of pre-detections was located at the edge of the helmet all pre-detections in this channel were excluded, as well as all pre-detections occurring at the same time in other channels. This criterion is based on empirical observations showing that the effect of movement artefacts is more pronounced in channels at the border of the helmet (Muthukumaraswamy
2013). (iii) Pre-detections in which the power increase with respect to the background in the high-gamma or ripple band was lower than in channels with no pre-detections. This led to the exclusion of potentially true positive FO events occurring simultaneously with other high frequency physiologic activity, instrumental noise, or artefacts that would have affected the source localization results. These events were excluded only to facilitate the source localization procedure.
The next step was to identify the channel with maximum rate of pre-detections for each band separately. The final output of the automatic detector consisted only of the pre-detections that occurred in any channel at the same time as a pre-detection in the highest rate channel. This approach simplifies the rest of the analysis since the visual review that follows is done only at the time of pre-detections in the highest rate channel. This is an easily replicable methodology that precludes a biased selection of a few pre-detections that might coincide with clinical information. At the same time, in epilepsy patients looking only at the channel with highest rate minimizes the detection of physiologic FOs, since the rate of pathologic FOs is higher (Matsumoto et al.
2013; Wang et al.
2013).
Review by Experts
As mentioned above, a large number of false positive pre-detections was expected and required a second step carried out by human reviewers to visually identify and select true positive events. This task was performed on a number of pre-detections limited to 100 in the high-gamma and 100 in the ripple bands, i.e. for patients with more than 100 pre-detections in a given band, we only reviewed a random subset of 100, assuming that the sample was representative of all the detected events.
The visualization of the pre-detections for review was done using Brainstorm (Tadel et al.
2011), which is documented and freely available for download under the GNU general public license (
http://neuroimage.usc.edu/brainstorm). The reviewers were presented with with 1-s epochs centered around the pre-detections, showing the bandpass filtered signal in the high-gamma (40–80 Hz) or ripple (80–160 Hz) band from the highest rate channel and 17 neighboring channels. A topographic map of the time–frequency decomposition of the signal in the high-gamma or ripple band at all channels was also presented to help identify potential muscle artifacts.
The pre-detections were independently classified by two expert reviewers (EK and GP) following the definition of HFOs commonly used in EEG investigations, i.e. four oscillations clearly standing out of the background in the analyzed high-gamma (40–80 Hz), and ripple (80–160 Hz) frequency bands (Jacobs et al.
2012,
2012; Andrade-Valença et al.
2011; Melani et al.
2013). To maximize the reliability of the detections, we admitted as true positives only the pre-detections that both reviewers considered FOs based on these criteria. The signal from FOs labelled as true positive by both experts was then carefully reviewed by one of the experts in the clinical frequency range (0.3–70 Hz) to remove any remaining detections associated to artefacts. This last step also allowed us to determine the co-occurrence of FOs and IEDs. Concordance at channel level was defined for FOs when the channel with highest FO rate was in the affected lobe.
MSI of High-Gamma and Ripple FOs
The anatomical MRI of each patient used for MSI consisted of a T1 W MPRAGE 1 mm isotropic 3D acquisition (192 sagittal slices, 256 × 256 matrix, TE = 2.98 ms, TR = 2.3 s, Siemens Tim Trio 3T scanner). The MRI was segmented and the cortical surface obtained using BrainVISA v.4.4.0 (IFR49, France) (Mangin et al.
2004) and the MEG forward problem was solved with the boundary element method for a single layer model using OpenMEEG (Gramfort et al.
2010). Source localization of the FOs was performed with the wavelet version of the Maximum Entropy on the Mean (wMEM) method (Lina et al.
2014). We used implementations of OpenMEEG and wMEM available in the current version of Brainstorm.
The MEM principle (Amblard et al.
2004) is a non-linear inverse problem solver, coupled with a data-driven parcellation (DDP) to cluster the whole cortical surface into K non-overlapping parcels, as originally proposed by Lapalme et al. (
2006). DDP consists in using partial information from the available data in order to guide this spatial clustering. The key aspect of DDP lies in the pre-localization of the sources of brain activity using the multivariate source pre-localization (MSP) method (Mattout et al.
2005). MSP is a projection method that estimates a coefficient, which characterizes the possible contribution of each dipolar source to the data. DDP in K parcels is then obtained using a region-growing algorithm around the local maxima of the MSP map. In the MEM reference model, a hidden variable is associated to each parcel in order to model the probability of the parcel to be active (probability initialized using the MSP coefficients). Parcels that do not contribute to explain the measured data are automatically switched off by maximizing the entropy of a mixed probability distribution, allowing the method to recover accurately the source location together with their spatial extent along the cortex (Chowdhury et al.
2013). This property of the MEM framework is particularly important in the context of epilepsy where spatially extended generators are expected, and we have carefully demonstrated the ability of MEM to perform accurate source reconstruction of interictal spikes in EEG and MEG (Heers et al.
2014;
2015). The wMEM extension of the MEM framework (Lina et al.
2014) decomposes the signal in a discrete wavelet basis before performing MEM source localization on each time–frequency box. Thus, wMEM is particularly suited to localize oscillatory patterns, as evaluated with realistic simulations (Lina et al.
2014).
In this study, we localized only the time–frequency box with highest amplitude during the detected FO, in the high-gamma and ripple bands separately. The resampling of the signals to 640 Hz ensured that the third scale of the discrete wavelet transform corresponded to the high-gamma band (40–80 Hz), and the second scale to the ripple band (80–160 Hz). A diagonal model was adopted for the noise covariance matrix in the data space, estimated independently for each FO. The estimation was done based on the MEG background in the high-gamma or ripple band in a 0.5 s window immediately before the beginning of each FO.
Source localization was performed for each FO. The resulting map consisted of a cortical activation value associated to each vertex of the cortical tessellation. Each map was normalized in order to get a maximum activation value equal to one for all FOs, and then the average of the activation value at each vertex was computed across all FOs. In this way a single cortical source localization map per patient and per frequency band was obtained.
For displaying the resulting maps over the cortical surface, a threshold of 30 % of the maximum activation was applied, as in previous studies (Heers et al.
2014,
2015). The concordance with the epileptogenic region was evaluated at sublobar level as concordant, partially concordant, or discordant. The partially concordant category corresponds to cases in which local maxima of the source map were found both inside and outside the clinically defined epileptogenic zone.