Localization of the ictal onset zone with MEG using minimum norm estimate of a narrow band at seizure onset versus standard single current dipole modeling

1. Introduction

Literature on the yield of ictal magnetoencephalography (MEG) is limited to case reports and a few case series (Assaf et al., 2003; Eliashiv et al., 2002; Mohamed et al., 2007; Tang et al., 2003; Tilz et al., 2002; Yagyu et al., 2010). Most of these studies were done using single Equivalent Current Dipole (sECD) model in order to localize seizure onset zone. There are some conceptual challenges when it comes to implementing this model in localization of ictal rhythms, especially in cases of paroxysmal fast activity. Successful sECD-fitting rate appears to decrease as a function of frequency as studied by de Jongh et al. (2003).

In this report, we describe a new method of analysis of ictal rhythms and implement it in an illustrative case. Although both approaches yielded concordant results at the lobar level, the rhythm-based approach provided more accurate sublobar localization. Successful surgical resection, guided by intracranial EEG, contained the area of activation delineated by the rhythm-based method leaving behind the sECD area.

2. Case presentation

19-year-old right-handed woman presented with medically intractable focal epilepsy since infancy. She underwent two resections in the right peri-rolandic area at an outside hospital at 5 and 8 years of age. Seizures persisted immediately after surgeries. Her spells started with a sensation of “head spinning” followed by intense sensation of left arm shaking evolving to frank left arm clonic movements with infrequent generalization. Prior to her presentation at our center, the patient was having two seizures per day on average. Neurological examination showed mild left facial fold effacement and strength 4/5 in the left upper extremity representing residual dysfunction from her prior surgeries. Video-EEG showed ictal paroxysmal fast rhythms in the vertex region with earlier involvement of the right parasagittal chains (Fig. 1A). MRI of the brain showed two prior resection cavities over the right side rostral and caudal to the primary motor area (Fig. 1B). FDG–PET showed hypometabolism in the anterior margins of the anterior resection and ictal SPECT was showed a concordant area of hyperperfusion in the same region (Fig. 1C). One typical ictal event was recorded during routine outpatient MEG recording.

Fig. 1.

Summary of pre-surgical findings and localization of seizure onset zone by intracranial depth electrodes. MRI, PET, SPECT, and MEG images follow the neurological orientation (‘right is right’). (A) Several clinical seizures were recorded and had the same electro-clinical behavior. EEG onset (left – red vertical line) was marked by general attenuation of background and appearance of paroxysmal fast activity at the vertex with subsequent involvement of the right parasagittal chains (middle). The activity became more widespread shortly before its offset (right – blue vertical line). (B) Axial FLAIR and sagittal T1-weighted images showed two prior resection cavities caudal and rostral to the right primary motor cortex, which was preserved in prior resections. (C) Axial FDG–PET image (left) showed decreased uptake (arrow) in the right frontal region and subtraction ictal SPECT co-registered to MRI scan (right) showed increased blood flow anterior and medial to the prior anterior resection. (D) Seizure onset zone was delineated by stereotactically implanted depth EEG electrodes and was localized to the anterior edge of the previous anterior resection cavity. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

3. MEG analysis

60-minute resting-state recording of spontaneous MEG and EEG signals were performed using a whole-head 306-channel Neuromag system (Elekta, Helsinki, Finland) at a sampling rate of 1000 Hz and acquisition-band-pass filtered between 0.5 and 70 Hz. EEG was recorded simultaneously using the international 10–20 system electrodes with additional anterior temporal electrodes. Evaluation of the sources producing the interictal or ictal activity of interest was carried out in Neuromag software. Waveform analysis was performed on data segments that contained ictal onset. The location, orientation, and strength of the dipole sources that best fit the measured magnetic fields were calculated with standard software using least squares algorithms to fit the magnetic fields predicted by a single equivalent current dipole model. Dipoles were accepted when certain modeling parameters were met: goodness of fit >90%, confidence volume <1000 mm³ at 95% confidence limit and amplitude of dipole moment >100 nA (nano-Amperes). The dipoles were fitted to a multiple-sphere head model at the sensor level and co-registered to the patient's MRI per standard procedure using routine fiducial points. Head sensors were used to identify the position of the patient's head in the helmet to compensate for head movements during the procedure.

A separate analysis was done using a rhythm-based method. The peri-ictal and ictal data from all sensors were transformed into the time-frequency space using Morlet Wavelet transformation (MWT). Compared to short time Fourier transformation (STFT), MWT has superior time frequency resolution and is ideal for analyzing non-stationary signals like EEG, and MEG. The multiple scales of the wavelet transform permit the equivalent of large- or small-scale transform windows in the same time series. A complex Morlet Wavelet ψ(t) was used, which is defined as:

$$\mathrm{\Psi }(t)=\frac{1}{\sqrt{2\pi \cdot {\sigma }_t^2}}{e}^{i2\pi f_{c}t}{e}^{\frac{-t^2}{2\cdot {\sigma }_t^2}}.$$ (1)

The central frequency (f_c) of the mother wavelet was set to 1 Hz. The standard deviation of the Gaussian window was set to σ_t = 1. The family of wavelets can then be generated by dilation and translation of the mother wavelet ψ(t) as

$${\mathrm{\Psi }}_{a,b}(t)={|a|}^{\frac{1}{2}}\mathrm{\Psi }\left(\frac{t-b}{a}\right),$$ (2)

where (a) is the dilation (scaling) factor and (b) is the translation (time shift) factor. Through this process of translation and dilation, it becomes possible for us to study the frequency content of signal at different time scales using wavelet transforms without resorting to fixed-window-based analysis as conventional STFT based techniques do. We chose the scale to correspond with frequency steps of 1 Hz and a range 1–50 Hz. The wavelet coefficients were then estimated from the peri-ictal and ictal magnetometer recordings. Then a matrix P(f,t) was calculated from wavelet ψ(a,b) by transforming the scale (a) into pseudo-frequency f, and replacing (b) with time. The values of the matrix (wavelet coefficients) were then squared to render signal power. After applying the continuous wavelet transformation the squared coefficients were averaged over 100 ms windows to generate the final product for each sensor which is a matrix P(f,T) of the power at a time window T that is 100 ms wide, and frequency f. In order to identify the dominant frequency, data from the sensors involved at seizure onset which were visually identified as sensors which exhibited paroxysmal fast activity at the onset of seizures (18 sensors in this case) were then averaged in the time frequency space following the transformation above. Evolving patterns representing the seizures were then identified visually in the time–frequency space and correlated with the onset in the time series. Seizures in the time–frequency space appear as accelerating “chirps” (evolving from lower to higher frequencies – this case) or decelerating “chirps” (evolving from higher to lower frequencies – typical scenario). After marking the exact time of onset, the frequency with the highest power at the time of onset was identified visually and verified from the power matrix (15 Hz in this case) (Fig. 2).

Fig. 2.

MRI, PET, SPECT, and MEG images follow the neurological orientation (‘right is right’). (A) 10-second MEG data from the right frontal and vertex sensors at the time of onset. The blue frame highlights 400-millisecond window at onset, which is magnified in the box (arrow). The vertical blue line marks the time point at which sECD analysis was performed. The red frame highlights 100-millisecond window on which L2-MNE was subsequently performed. The black-rimmed box shows the distribution of the magnetic field, which remained stable for several seconds following ictal onset. (B) Morlet wavelet transformation of the same 10-second window shown in (A) after averaging power spectra of the involved electrodes. This step is important for visual identification of the dominant rhythm at onset in terms of signal power (in this case 15 Hz, as highlighted by the red arrow). (C) Mapping of current density of the narrow band identified in B on a 2D helmet using 200-millisecond intervals for 1 s to demonstrate signal relative stationarity (in space and time), focality (in space), and evolution (in time and power) prior to implementing L2-MNE solution. This step is important to confirm stability of the ictal rhythm and to rule out temporal overlap with other ‘contaminating’ rhythms such as physiologic beta rhythms (illustrated in the lower box). Such rhythms tend to show bilateral distribution across the midline. (D) sECD analysis (upper row) at the time point shown in A (vertical blue line) localized the ‘point’ of activation in the precentral gyrus. On the other hand, L2-MNE solution (lower row) of a narrow band of interest (14–16 Hz) localized the seizure onset in the anterior edge of the prior anterior resection. (E) Post-surgical T1-weighted images confirmed the resection of the area delineated by L2-MNE of the narrow band of interest. (F) L2-MNE is biased to signals with high power. This figure illustrates L2-MNE solution of the signal prior to filtering. The active areas shown do not overlap with the seizure onset zone. Rather they represent distant areas of slowing in the setting of an active seizure. This illustration highlights the significance of selection of a narrow band tailored to the ictal rhythm. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

The cortical surface of the brain was extracted using BrainVISA, which is a freely available package for structural morphometry of the cortical sulci (www.brainvisa.info). Refer to the website for details of methodology of extraction of the cortex. For faster processing and computation, the vertices were down-sampled to 15,000 vertices over each hemisphere. It is important to note that the same step can be accomplished by other freely available academic packages such as Freesurfer and Brainsuite. For the purpose of this analysis, BrainVISA has the advantage of faster extraction compared to other methods.

Then an L2-minimum norm estimation of sources (L2-MNE) (Hamalainen and Ilmoniemi, 1994; Ioannides et al., 1990; Uutela et al., 1999) was applied on a narrow band (3 Hz band width) with the central frequency being the frequency of interest identified at onset in the time frequency space (14–16 Hz, 15 Hz, respectively in this case). The choice of the bandwidth is proportional to the frequency of interest in that lower frequencies can be represented precisely with 1 Hz bandwidth, whereas the width should be increased at higher frequencies to compensate for relatively lower frequency resolution at that level. The sources were constrained to the cortical orientation. To avoid influences from weak insignificant estimates, we used only areas with amplitudes exceeding 55% of the maximum amplitude at a certain time point (Lin et al., 2006). The analysis of ictal rhythms was performed on the earliest time window showing relative stability and focality of the signal (narrow band) in the current density plot (Fig. 2A–C). The implementation of L2-MNE solution of the inverse problem was performed using Brainstorm (Free to download at http://neuroimage.usc.edu/brainstorm/), which is inspired by Matti Hamalainen's MNE software. A full description of the method can be found in the MNE software User's Guide, chapter 6, “The current estimates” at http://www.martinos.org/meg/manuals/MNE-manual-2.7.pdf.

In this case, ictal rhythm consisted of evolving paroxysmal fast activity better seen in the right fronto-parietal and vertex sensors. sECD analysis of this activity was reproducible at different time points within the first few seconds of onset and suggested that the seizure onset was located within the primary motor area, posterior to the anterior resection (Fig. 2D upper row).

The area of activation delineated by the new method was stable and reproducible in different time windows following the onset (Fig. 2C) and suggested that the seizure onset zone was located anterior to the area delineated by sECD (Fig. 2D lower row). The distance between sECD and the center of the area of activation was 21.8 mm (i.e. vertices with highest amplitude). Patient underwent invasive evaluation with stereotactically-implanted depth electrodes to pinpoint the area of ictal onset. Seizures were found to originate just anterior to the anterior resection cavity in a region that overlapped with the area identified by L2-MNE. The patient underwent a resection of the seizure onset zone delineated by stereo-EEG, preserving the primary motor cortex (thus, resecting the area delineated by L2-MNE, and leaving behind the area containing the ictal sECD). Surgical pathology revealed focal cortical dysplasia. The patient continues to be seizure free 12 months following surgery.

4. Discussion

In this paper we propose an alternative seizure-specific approach for analysis of ictal activities, which may be captured in some patients undergoing preoperative MEG investigations.

Modeling of certain ictal rhythms poses additional challenges when compared to interictal analysis of sharp waves using standard sECD model (Tilz et al., 2002). Using parametric dipole models (Brenner et al., 1978; Mosher et al., 1992; Uutela et al., 1999), the active brain areas can be located with a reasonable accuracy. That being said, it is not straightforward to select an appropriate model when several temporally overlapping sources are active. The sECD model produces accurate results whenever the magnetic field is generated by activity arising from a focal neuronal population (Eliashiv et al., 2002; Gallen et al., 1993). Extending sECD modeling to dynamic ictal patterns is conceptually problematic but may be applied in cases of relative magnetic fields stationarity at seizure onset. Instead of point-like sources, some inverse solutions search for the best estimate of a distributed current using minimum-norm principles. The solution with the smallest norm is selected from all those current distributions that could explain the measured magnetic field. L2-MNE describes more accurately distributed sources without making a priori assumptions, and therefore it is considered to be less prone to modeling errors. The L2-MNE has been applied to several types of magnetic inverse problems, such as source localization of event-related magnetic fields (Ahlfors et al., 1992; Babiloni et al., 2000; Haan et al., 2000; Rinne et al., 2000; Shibata and Ioannides, 2001) including epileptiform areas (Fuchs et al., 1999). However, to our knowledge there is no literature on it being applied for analysis of ictal rhythms in the way implemented in this paper.

Methodologically, our suggested approach takes advantage of the uncertainty inherent to the MNE and assumes that the seizure onset zone corresponds with area of higher activation of a narrow band of interest at the time of onset as delineated by MNE. It is important to beware that MNE solution of the inverse problem is biased to signals with high power. However, signals that correlate with the seizure onset are not always more powerful than background rhythms, especially with onsets consisting of paroxysmal fast activity. Hence, we applied L2- MNE solution only on a narrow band identified visually in the time–frequency space at the onset. This eliminates the bias created by non-ictal signals such as physiologic rhythms, and remote regions of slowing in the setting of active seizure. This step is important and produces more localizing images than those produced by the same analysis done on the entire width of the signal (Fig. 2D and F).

Intrasubject test–retest variability of sECD localization attributed to different reasons was 7 mm for measurements conducted with the patient in the same position (Gallen et al., 1993). Researchers in previous studies have demonstrated that the total registration-related spatial uncertainty involved in these processes is approximately 4 mm (Knowlton et al., 1997). In our case, the distance between sECD and the center of L2-MNE area was 21.8 mm making it difficult to attribute the differences simply to registration differences. The patient underwent a tailored resection based on subsequent intracranial recordings. The resection contained the area delineated by the rhythm-based method whereas the area pinpointed by sECD was not resected. The patient remains seizure free 12 months following resection.

5. Conclusion

We present a rhythm-based method of analysis of ictal events recorded during MEG. This method has conceptual advantages over point-like models, namely sECD and does not require priori assumptions. The feasibility of the analysis is illustrated in one case. Further studies are needed to confirm the reliability and accuracy of this method.