Interictal High Frequency Oscillations Detected with Simultaneous Magnetoencephalography and Electroencephalography as Biomarker of Pediatric Epilepsy

Ethics Statement: Experimental procedures here have been approved by the Institutional Review Board (IRB) of Massachusetts General Hospital, Boston, MA, USA. The following section will describe the experimental protocol for the noninvasive detection and source localization of HFOs using scalp EEG and MEG. Patient preparation is minimal, and the examination is generally well tolerated. The entire session lasts about 2 – 3 h with the actual recordings lasting around 60 min.

1. Patient's Preparation

1. Ensure that the child is comfortable with the environment.
   NOTE: Allow young children to explore the testing environment and to see the testing equipment. Screen the patient regarding safety and consent by using a screening form. Ask the patient (or his/her parents) whether he/she had a clinical seizure within the last 2 h.
2. Remove all metallic/magnetic materials and provide hospital-issued clothing to the patient. Remove the shoes since they are frequently magnetic. Check if the subject is free of magnetic artifacts by measuring MEG signals for a few minutes. Use a degausser for reducing artifacts from implanted materials, such as dental works.
   NOTE: The demagnetizer should not be applied if there is any possibility that ferromagnetic objects are within the body.
3. Measure the maximum head circumference to select the appropriate EEG cap size for the child. Use a measuring tape and hold it to the nasion. Then measure around the head around the maximum circumference (~1 cm above the inion).
4. Place the EEG cap on the head according to the International 10 – 20 system. Cleanse the skin where each electrode is located and apply paste/gel for each electrode.
   NOTE: More details regarding scalp EEG recordings in children are provided elsewhere⁴¹.
   1. Place the ground and reference electrodes on the head. Connect additional electrodes for measuring horizontal and vertical electrooculography (EOG), electrocardiography (ECG), electromyography (EMG) and additional EEG electrodes at locations covering the temporal regions (T1/T2).
      NOTE: The EOG, ECG, and EMG aid the identification of eye movements, magnetocardiographic contamination, muscle activity, and also to monitor the patient's state.
5. Ensure that the EEG channels have good contact with the scalp by positioning the sensors individually. Gently twist each sensor from side to side to move hair out of the way. Measure all electrode impedances with an EEG ohmmeter in order to be below 10 KOhm.
6. Place four HPI coils on the head: two behind the earlobes and two on the forehead at approximate symmetrical locations.
   NOTE: The HPI coils help to localize the relative position of the patient's head with respect to the location of the MEG sensors in 3D space. The number of HPI coils may differ depending on the vendor of the MEG system.
7. Obtain the locations of the HPI coils and the EEG electrodes using a digitizer.
   NOTE: The digitizer records the coordinates of a sensor in the 3D space. The sensor is placed on the tip of a stylus. The locations of the HPI coils must be known with respect to the head anatomy and the locations of the MEG sensors.
8. Obtain the locations of fiducial landmarks including left/right preauricular points and nasion using the digitizer. Digitize additional points (approximately 300 points) to obtain precise head shape.
9. Transfer the patient into the magnetic shielded room (MSR)⁴², where the MEG system is located.
   NOTE: The MSR is a shielded environment that minimizes the interference of MEG recordings from external electromagnetic sources (i.e., power lines, radiofrequency signals from portable devices, electrical devices and computers, magnetic fields from moving magnetized objects such as cars, elevators, and trains). It consists of three nested layers. Each layer is made of a pure aluminum layer plus a high permeability ferromagnetic layer (i.e., mu-metal, an alloy consisting mostly of nickel and iron).
10. Lay down the patient on the bed, put his/her head into the MEG helmet, and apply appropriate pads/sponges under the patient's head for comfort.
11. Connect the HPIs, the EEG leads, the EOG, the ECG, the EMG, and the additional electrodes to the recording machine. Adjust the patient's head position in the scanner ensuring that it is located as deep as possible in the helmet.

2. Data Acquisition

1. MEG and EEG
   NOTE: MEG/EEG data acquisition is performed based on the method described in a previous study⁴². More details about the clinical use of MEG in pediatric epilepsy can be found elsewhere^43,44.
   1. Record MEG signals with a whole-head MEG system.
      NOTE: The MEG system employs thin-film sensors of two types (planar gradiometers and magnetometers) integrated on 102 sensor elements. Each element contains a magnetometer that consists of a single coil, and two orthogonal planar gradiometers that consist of a "figure-of-eight"-type coil configuration. The magnetometer measures the magnetic flux perpendicular to its surface and the gradiometers measure the difference between the two loops of the "eight", or the spatial gradient. The MEG system has 204 planar gradiometers and 102 magnetometers (306 sensors in total). MEG systems from different vendors have different number and types of coils (i.e., axial gradiometers).
   2. Record simultaneously EEG signals using a nonmagnetic 70-channel electrode cap with Ag/AgCl sintered ring electrodes and additional electrodes in T1/T2⁴². Use a common reference montage.
   3. Close the door of the MSR. Communicate with the patient via an intercom system to check if he/she feels comfortable. Ask the parent to stay inside the MSR during the recording if the child feels uncomfortable to stay alone.
   4. Start the recordings by clicking the 'Go' button in the MEG acquisition software. Use a high sampling rate of 1 KHz (or more). Use a low-pass Infinite Impulse Response (IIR) filter of 6^th order at 400 Hz. Check online all the recorded signals. Fix bad MEG channels by using a sensor tuner.
      NOTE: Bad MEG channels are defined sensors (gradiometers or magnetometers) that have a relatively high level of white noise (above 2 to 5 fT/√Hz for magnetometers) or sensors that record spurious environmental electromagnetic noise. This is usually happening when the sensors are exposed to strong (relative to the signals measured) magnetic fields and specific parts of the coils "trap" the magnetic flux destroying the superconductivity. A sensor tuner then is used that heat the coil by applying an electric current through it. This procedure is called tuning and is used when the white noise level of the sensor is above a specific threshold (i.e., 2 - 5 fT/√Hz). Some MEG systems do not have sensor tuners.
   5. Measure the patient's head position by clicking the 'Measure' button in the MEG acquisition software. If the patient's head is not well-covered by the sensory array, ask the patient to move his/her head deeper into the helmet.
      NOTE: This maneuver activates the 4 HPI coils by applying transient oscillatory electrical signals through the coils that generate artificial magnetic fields. These fields are detected by the MEG sensors, thus the head position is determined. The procedure may differ between different MEG vendors. 
   6. Record MEG, EEG, and peripheral recordings by clicking the 'Record' button in the MEG acquisition software (i.e., ECG, EOG, and EMG) for ~60 min.
      NOTE: The data is stored as a .fif file in the Redundant Arrays of Independent Disks (RAID). The file type is different for other MEG vendors.
   7. When the recording ends, open the MSR, disconnect the cables, and take out the patient from the MSR room. Remove all the tapes, electrodes, HPI coils, and EEG cap gently. Provide head washing for the patient.
   8. After the acquisition is completed, record the magnetic signals of the empty MSR without the patient. Start the recordings by clicking the 'Go' button in the MEG acquisition software. Record MEG data for 2 min using the same parameters as in step 2.1.4.
      NOTE: This data is used for estimating the environmental electromagnetic noise.
2. MRI
   1. Acquire anatomical MRI data with magnetization-prepared rapid acquisition gradient-echo sequences (MPRAGE; TE = 1.74 ms, TR = 2,520 ms, voxel size = 1 × 1 × 1 mm) with a high-resolution 3T scanner. Details about the MRI scanning protocol can be found elsewhere⁴⁵.
      NOTE: We do not perform the MRI scan the same day as the MEG session in order to avoid artifacts in the MEG recordings due to magnetization of possible patient's metallic implants, such as from dental works.

3. Identification of Interictal Activity

1. Open the data using Brainstorm⁴⁶, which is documented and freely available for download online under the GNU general public license.
2. Select visually portions of the EEG data with interictal activity occurring at least 2 h apart from clinical seizures.
   NOTE: Figure 1 presents a portion of EEG and MEG data with frequent IEDs.
   1. Identify empirically well-defined IEDs in the EEG signals: this includes spikes (20 – 70 ms) and sharp waves (70 – 200 ms)^47-48.
      NOTE: The clinical significance of both types of IEDs in epileptic focus localization is equivalent.
   2. Try to identify (if possible) portions of the recordings with: (i) minimal motion artifacts, (ii) more than 3 – 4 IEDs per 10 s display, and (iii) slow-wave non-REM sleep that usually present a high number of HFOs⁴⁹.

Figure 1: IEDs in EEG and MEG Signals. Portion of simultaneously recorded scalp EEG and MEG signals with frequent IEDs. 1 s of highlighted section that contains one sharp wave is presented on the right panels in an extended time scale display. Red dots indicate the peak of the IEDs. Please click here to view a larger version of this figure.

3. Using Brainstorm, display the data with standard display settings (10 s/page). Go to the Filter tab and put the following filter display parameters: high-pass filter: 1 Hz, Low-pass filter: 80 Hz, Notch filter: 50 or 60 Hz (according to the frequency of the power line). Inspect the data and identify portions of data with IEDs.
   NOTE: Only portions of signal with IEDs will be scanned to look for HFOs (step 3.4). The selected filters are for visualization only; they have not been applied to the data. In order to apply these filters permanently to the data, use a band-pass Butterworth filter (4^th order) following the instructions in Brainstorm website (http://neuroimage.usc.edu/brainstorm/).
4. Mark the peak of each IED occurring in both EEG and MEG data (see red spots in Figure 1).
   NOTE: More details about marking IEDs using Brainstorm can be found elsewhere (http://neuroimage.usc.edu/brainstorm/Tutorials/Epilepsy).

4. Semi-automated Detection of HFOs in Simultaneous Scalp EEG and MEG Data

NOTE: Here we describe a semi-automated method to detect HFOs, which includes an automated detection (step 4.1; Figure 2), followed by a visual review of the automatically detected HFOs (step 4.3). In order to avoid the spurious oscillations of sharp transients as true ripples and to ensure that the HFOs are not due to a filtering phenomenon, we followed the latest suggestions in the relevant literature: we required the HFOs to have a minimum number of 4 oscillations since it has been observed that the impulse response of the filter has fewer oscillations than the chosen number of cycles⁵⁰, we used the Finite Impulse Response (FIR) filter to minimize ringing effect and the "Gibbs" phenomenon⁵⁰, we required the candidate HFO events to be inspected also visually by an expert to check whether the HFOs were also visible overlaid on the IEDs^50,51, and we required an isolated island to be observed in the time-frequency plain because a sharp event and an oscillation have different signatures: a real HFO is represented by an isolated peak in the time-frequency plot (restricted in frequency, as "island") located in the band of 80 – 500 Hz, while a transient event generates an elongated blob, extended in frequency^50,52,53.

Figure 2: Schematic Diagram of the Algorithm Steps. The algorithm works in two stages: the first one identifies candidate HFOs from each EEG signal in the time domain (left and middle columns); the second one classifies the previously detected candidate events in order to distinguish real HFOs from artifacts in the time-frequency domain (right column). Please click here to view a larger version of this figure.

1. Automatic HFO detection
   NOTE: Figure 2 describes the flow chart of the automatic detection of HFOs on each EEG signal. The goal of the developed method is to reduce the burden for the EEG expert of marking HFO events on each EEG channel using a 2 s/page display that is recommended for the visual inspection of HFOs. An HFO was defined as an event within the ripple frequency band (80 – 250 Hz), which has at least 4 oscillations of sinusoidal like morphology standing out from the surrounding background⁵⁴ (step 4.1.1), and which appears as a short-lived event with an isolated spectral peak at a distinct high frequency¹⁴ (step 4.1.2).
   1. Detection of candidate HFOs in the time domain
      1. Band-pass (BP) filter the EEG signals between 80 and 250 Hz, restricting the frequency content of the signals to the ripple band of interest.
         NOTE: It is recommended to use a FIR filter to minimize ringing effect and the "Gibbs" phenomenon⁵⁵, and zero-phase digital filtering to avoid phase distortion.
      2. Calculate the envelope of the BP signal using the Hilbert transformation. Calculate the mean and the standard deviation (SD) of the envelope over 10 s sliding windows centered on each point of the time-series. Estimate the overall mean and SD using the median value over all the windows (in order to obtain values that are robust to the possible presence of portions of the signal with many HFOs and high SD).
      3. Calculate the z-score of the envelope and mark a candidate HFO every time the z-score is higher than the minimum threshold, set equal to 3⁵⁶.
      4. Define the starting and ending points of the detected event as the upward and downward crossings of half the threshold. Consider the HFOs with an inter-event interval of less than 30 ms as one single HFO. Calculate the number of peaks in the BP signal between the HFO starting and ending points, and discard events with less than 4 peaks. Also, discard events with a z-score higher than 12.
         NOTE: Modify your maximum z-score threshold according to the amplitude of the artifacts that may occur in your recordings. Events with a low number of oscillations can be caused by filtering effects^57,58, whereas events with extremely high amplitude can be due to muscle or electrode artifacts.
   2. Reject possible artifacts in the time-frequency domain.
      NOTE: This step is necessary to distinguish real HFOs from events that might be elicited by other EEG activity and filtering artifacts, whose frequency content is not restricted to the frequency band of interest. It is based on the assumption that a real HFO appears as a short-lived event with an isolated spectral peak at a distinct frequency above 80 Hz, in contrast with a transient event that generates an elongated blob extended in frequency⁵⁹. Figure 3 illustrates an example of a detected HFO showing the BP filtered EEG signal (upper panel), its envelope (middle panel), and the corresponding time-frequency plane (lower panel), during the period of [-0.5, +0.5] s around the HFO peak. The display of the time-frequency plane is restricted from 80 – 150 Hz because no prominent activity was observed for frequencies above 150 Hz.
      1. Transform all candidate HFOs events into the time-frequency space using the Morlet transformation in the frequency range from 1 Hz to the highest frequency of interest, i.e., 250 Hz (central frequency = 1 Hz, Full-Width-At-Half-Maximum = 3 s).
      2. Analyze the instantaneous power spectra of the time-frequency representation over each time point of the event duration. For each power spectrum, follow the automatic criteria described by Burnos et al.⁵⁶ to detect the peak in the high-frequency band and to verify whether it is clearly distinct from the closest peak in the lower frequency range. Discard HFOs that do not show a power spectrum with an isolated high-frequency peak in at least 90% of the time points.
2. Sort all the detected HFOs events by their temporal occurrence across channels. Group together all the consecutive HFOs whose duration overlaps. Keep only groups of HFOs involving at least two EEG channels for further analysis.
   NOTE: The algorithm requests the HFOs to occur in at least 2 channels in order to avoid capturing spurious random artifacts, which may resemble real HFOs and occur in single EEG leads. Two consecutive HFOs are considered as overlapping when the starting time of the second HFO precedes the starting time of the first one.

Figure 3: HFO Event Detected by the Algorithm. Upper panel: the BP filtered (80 – 150 Hz) EEG signal (in µV) from one channel (F8 – T8) from patient 1. Middle panel: the envelope of the BP signal (z-score). The peak of the envelope (red asterisk) indicates the timing of the HFO peak (red vertical dotted line). The blue asterisks mark the upward and downward crossings of half the threshold (blue dotted line), which indicate the start and end time-points of the HFO (blue vertical dotted lines). Lower panel: the time-frequency analysis plane. Note the isolated peak in the ripple frequency band (~100 Hz) around the peak of the HFO. Please click here to view a larger version of this figure.

3. Visual review of HFO events
   NOTE: Part of the visual review is based on the guidelines proposed by Andrade-Valenca et al.²⁸ and Zelmann et al.⁶⁰.
   1. Vertically align 2 computer screens; one for the inspection of EEG and one for the inspection of MEG signals. Display the detected events on both the expanded (2 s/page) and typical scale (10 s/page) showing, respectively, the 80 – 250 Hz and the 1 – 40 Hz BP filtered signals.
   2. Ignore events cooccurring with muscle or electrode artifacts in the unfiltered EEG and MEG, as well as events with large frequency variability, irregular morphology, or large amplitude variations.
   3. Observe the EOG and EMG signals during the detection of the HFOs and discard any event that is thought to correspond to EOG or muscular activity. Consider only the HFOs that overlap with EEG/MEG IEDs (detected in step 3.3) as they are more likely to be true HFOs^15,28,56.
      NOTE: This approach offers high specificity at the cost of low sensitivity; thus, it provides confidence that the identified HFOs are of cortical origin.
   4. Keep only HFO events that occur in both EEG and MEG signals at the same time.

5. Source Localization of IEDs and HFOs

1. Localize the generators at the peak of the MEG IEDs, marked in step 3.3, using the Equivalent Current Dipoles (ECD). Use the Minimum Norm Estimates software that is freely available (http://martinos.org/mne/stable/index.html). Consider only spikes with goodness-of-fit (GOF) > 80% and dipole moment Q<500 nA – m. Overlay the ECD location on the MRI of each patient.
   NOTE: The Maximum Entropy on the Mean (MEM) is an attractive alternative method that determines the location and extent of the sources⁶¹.
2. HFO source localization at both EEG and MEG using the wavelet Maximum Entropy on the Mean (wMEM) method (as proposed by von Ellenrieder et al³⁸).
   NOTE: The MEM is an efficient technique that has been successfully used to determine the location and extent of sources of epileptic activity^62-64. The wMEM is an extension of MEM that has been developed for localizing oscillatory activity as evaluated with realistic simulations⁶⁵. It decomposes the signal in a discrete wavelet basis before performing MEM source localization on each time-frequency box. Thus, wMEM is particularly well suited to localize HFOs.
   1. Segment the MRI and obtain the cortical surface using Freesurfer^66-67.
   2. Solve the EEG/MEG forward problem with the boundary element method (BEM) for a 3-layer model using OpenMEEG⁶⁸.
   3. Resample the signals to 640 Hz in order to ensure that the second scale of the discrete wavelet transform corresponds to the frequency band of interest.
   4. Estimate the noise covariance matrix in the data space independently for each HFO, based on the background in the ripple band in a 150 ms window immediately before each HFO. Perform the source localization for each HFO in the ripple band and average along the HFO duration. Note: The resulting map consists of a cortical activation value associated to each vertex of the cortical tessellation.
   5. Normalize each map in order to have a maximum activation value equal to 1 for each HFO.
   6. Compute the average of the activation values across all HFOs at each vertex. Apply a threshold of 60% of the maximum activation in order to display the final maps over the cortical surface.

6. Validation

1. Intracranial EEG (iEEG):
   1. Acquire extra-operative iEEG by employing subdural grids and/or stereotactically guided depth electrodes. Guide the placement of the electrodes based on the results of previous presurgical evaluation tests and the clinical hypotheses to be addressed, which is specific for each patient.
      NOTE: Intracranial EEG is recorded with a digital EEG system using a 2 KHz sampling rate, as part of the presurgical evaluation. Subdural grids are typically the best choice if the topography of the cortical seizure area is important to resolve, and if mapping of the eloquent cortex is critical (e.g., with an epileptogenic area near primary motor cortex, where the exact boundaries can be estimated by stimulation mapping and ictal intracranial EEG analysis).
2. Definition of the SOZ:
   1. Identify the SOZ as defined by an expert epileptologist based on the clinical information available for each patient.
      NOTE: The SOZ is defined as the area showing the earliest and rapid spread of iEEG change (ictal discharge) from baseline prior to or concomitant with the clinical onset. All channels involved at the beginning of the ictal electrographic discharge, usually the first 5 s, were considered as the SOZ. In patients where the seizures are originating from more than one area independently of each other, all contacts within the different SOZs are regarded as SOZ contacts. The expert is blind to the HFO source localization results at the time of defining the SOZ. The clinical information also includes: ictal and interictal iEEG findings, visible lesion on the MRI, ictal and interictal scalp EEG findings.
3. Identification of the HFO channels:
   1. Detect the HFOs in each intracranial electrode as described in 4.1. Use Kittler's method to determine the threshold based on the histogram of the number of HFOs from all channels, then apply bootstrapping and calculate the mean as the final threshold⁶⁹. Finally, identify the channels with an HFO rate above the threshold.
4. Compare the HFO localization with the SOZ identified in 6.2 as the gold standard for the source localization results.