High-density Electric Source Imaging of interictal epileptic discharges: How many electrodes and which time point?

OBJECTIVE
To assess the value of caudal EEG electrodes over cheeks and neck for high-density electric source imaging (ESI) in presurgical epilepsy evaluation, and to identify the best time point during averaged interictal epileptic discharges (IEDs) for optimal ESI accuracy.


METHODS
We retrospectively examined presurgical 257-channel EEG recordings of 45 patients with pharmacoresistant focal epilepsy. By stepwise removal of cheek and neck electrodes, averaged IEDs were downsampled to 219, 204, and 156 EEG channels. Additionally, ESI at the IED's half-rise was compared to other time points. The respective sources of maximum activity were compared to the resected brain area and postsurgical outcome.


RESULTS
Caudal channels had disproportionately more artefacts. In 30 patients with favourable outcome, the 204-channel array yielded the most accurate results with ESI maxima < 10 mm from the resection in 67% and inside affected sublobes in 83%. Neither in temporal nor in extratemporal cases did the full 257-channel setup improve ESI accuracy. ESI was most accurate at 50% of the IED's rising phase.


CONCLUSION
Information from cheeks and neck electrodes did not improve high-density ESI accuracy, probably due to higher artefact load and suboptimal biophysical modelling.


SIGNIFICANCE
Very caudal EEG electrodes should be used for ESI with caution.


Introduction
In patients with pharmacoresistant focal epilepsy, epilepsy surgery aims at eliminating the epileptogenic zone, i.e. the specific area of cerebral cortex which is indispensable for the generation of seizures. A valuable, non-invasive neurophysiological tool to approach the epileptogenic zone is electric source imaging (ESI). Based on the patient's scalp electroencephalogram (EEG), ESI plots the sources of epileptic activity within a 3D model of the patient's brain (Zijlmans et al., 2019;Foged et al., 2020). ESI can be applied Abbreviations: ESI, electric source imaging; IED, interictal epileptic discharge; IQR, interquartile range; ILAE, international league against epilepsy; LSMAC, locally spherical model with anatomical constraints. on both ictal (Beniczky et al., 2016;Nemtsas et al., 2017) and interictal epileptic discharges (IEDs). IEDs occur more frequently than ictal EEG patterns and are less often corrupted by movement artefacts. They display relatively simple spatio-temporal dynamics and can easily be averaged to achieve a high signal-to-noise ratio (Pittau et al., 2014). In comparison to spherical head models and standard EEG with up to 32 electrodes, accuracy of ESI was found to increase with the use of individual realistic head models and high-density EEG with up to 257 electrodes (Lantz et al., 2003a;Brodbeck et al., 2011;Foged et al., 2020).
While the standard EEG array recommended by the International Federation of Clinical Neurophysiology (IFCN) covers the neurocranium above the hat brim line plus 6 electrodes in the inferior temporal chain, modern high-density EEG setups offer a much more caudal head coverage, including face and neck . Simulated data indicate a substantial increase in ESI accuracy if regions below the hat brim line are covered (Song et al., 2015). However, the additional value of inferior electrodes for ESI has never been assessed empirically in patients with epilepsy. On the one hand, due to their spatial proximity to the skull base, caudal EEG channels may help in localizing epileptic activity in basal structures such as the temporal lobes. On the other hand, located rather far from the convexity, they might be of less value for ESI in most extratemporal structures. In addition, caudal electrodes placed on face and neck record signals after complex propagation through several types of tissue and air (sinus cavities) and are relatively prone to muscle artefacts. These effects lower the signal quality recorded at these sites and may, in turn, reduce these electrodes' value for ESI.
At the EEG and Epilepsy Unit of the University Hospitals Geneva, ESI for clinical and scientific use is currently based on 257-channel EEG down-sampled to 204 channels and has been validated with this approach (Birot et al., 2014;Staljanssens et al., 2017). To eliminate the contribution of electrodes most affected by artefacts, the 204-channel setup does not include the most caudal row of electrodes covering the neck, nor does it cover the cheeks. In order to provide a reference for future clinical work, we aimed at evaluating this ''default" methodology by comparing its results to those achieved with 257, 219 or 156 EEG channels. We hypothesised that ESI based on the full 257-electrode setup would be superior to 204 channels in temporal lobe epilepsy but not in extratemporal epilepsy. We further expected an ''intermediate" 219-channel setup that includes inferior temporal electrodes to yield balanced results, and a very reduced 156-channel array above the hat brim line to perform rather poorly ( Fig. 1 A, Supplementary Fig. 1).
ESI is commonly performed at single time points, based on the respective EEG scalp voltage map at those specific moments in time. Thanks to the high temporal resolution of EEG in the order of milliseconds, different time points across an IED can be evaluated. The IED peak offers the highest signal-to-noise ratio; however, its source estimation may already be subject to propagation (Merlet et al., 1997;Lantz et al., 2003b;Bast et al., 2006;Plummer et al., 2010). Therefore, interictal ESI is most often performed at 50% of the IED's rising phase (Brodbeck et al., 2011;Megevand et al., 2014;Feng et al., 2016;Centeno et al., 2017). Still, some authors report highest ESI accuracy at very early time points of the IED (Plummer et al., 2019) or, on the contrary, at the IED peak (van Mierlo et al., 2017). Thus, we additionally compared ESI accuracy at five different time points along the IED's rising slope.

Patient recruitment
Until July 31st, 2019, the database of the EEG and Epilepsy Unit at the University Hospitals Geneva was retrospectively screened for patients meeting the following criteria: (a) a first resective brain surgery to treat pharmacoresistant focal epilepsy; (b) age older than 6 years at evaluation; (c) presurgical 257-channel EEG recording with a minimum of 3 focal IEDs; (d) presurgical highresolution magnetic resonance imaging (MRI); (e) known 12month postsurgical outcome. According to the International League Against Epilepsy (ILAE) criteria (Wieser et al., 2001), patients were classified as having favourable 12-month postsurgical outcome (ILAE 1 + 2: seizure-free or ''auras" only since surgery, equivalent to Engel IA and IB) or rather unfavourable outcome (ILAE 3 or higher: not seizure-free). The study was approved by the local ethics committee at the University Hospitals Geneva.

EEG acquisition and pre-processing
High-density EEG recordings (257 electrodes; unfiltered acquisition, sampling rate 500-1000 Hz) were acquired during presurgical epilepsy evaluation at the University Hospitals Geneva (EGI Philips / Magstim EGI, Eugene, US-OR). For clinical purposes, an EEG experienced neurologist (M. Seeck, S. Vulliemoz, and others) visually identified interictal spikes (<70 ms) and sharp waves (70-200 ms) using conventional display (e.g., longitudinal bipolar or average reference montages of the 10-20 electrode array plus 6 electrodes in the inferior temporal chain, corresponding to the 2017 IFCN recommendations ). In order to avoid any possible propagation/interference effects, IEDs occurring earlier than 1 second after a previous IED were not considered. IEDs of the same configuration and localisation were considered homologous. In the current study, 8 individuals had more than one type of homologous IEDs, and we used that IED type which occurred most often or, if the number of single IEDs was unknown, which fit best to the clinical focus hypothesis. One-second EEG epochs centred on the peak of the marked IED were filtered in the interval [1-70] Hz with a 4th-order Butterworth filter to avoid phase distortion, plus notch filtering at 50 Hz. EEG epochs that contained any artefacts surrounding the IED were discarded by visual inspection. IED epochs were averaged using the Cartool software (version 3.80) (Michel and Brunet, 2019) and spatially down-sampled to 219, 204, or 156 channels using MATLAB R2016a (The MathWorks, Inc., Natick, US-MA). The 257-and 219-channel setups included all 6 inferior temporal electrodes of the IFCN 25-channel array while the 204-channel setup included P9/P10 only, and all six were removed for the 156-channel setup ( Fig. 1 A, Supplementary  Fig. 1). Using Cartool, channels corrupted by continuous artefacts were identified via visual inspection of both EEG waveforms and surface voltage maps and corrected via interpolation from nearby channels using 3D splines. Finally, IED epochs were temporally down-sampled to 250 Hz.

MRI acquisition and pre-processing
Each patient's individual structural T1 or MPRAGE MRI image, acquired during pre-surgical evaluation, was re-sampled to 1 mm 3 isotropic resolution using cubic interpolation and preprocessed with Freesurfer (version 6.0.1) (Reuter et al., 2012) and the Connectome Mapper 3 open-source pre-processing software (Tourbier et al., 2019). A grey matter mask was generated which excluded brainstem and cerebellum. Based on the Desikan-Killiany anatomical atlas (Desikan et al., 2006;Destrieux et al., 2010), the grey matter was then parcelled into 82 regions of interest (41 per hemisphere) which were again grouped into 38 ''sublobar" areas (19 per hemisphere; Supplementary Table 1).

Forward models
With the use of Cartool and based on fiducial points at the head surface, the average 257-channel array was interactively co-registered to the patient's individual 3D-MRI. Using MATLAB, this co-registered array was down-sampled to 219-, 204-and 156electrode setups, respectively. For the forward solution, a simplified realistic 3-shell head model was used with consideration of age-adjusted skull thickness (Locally Spherical Model with Anatomical Constraints; LSMAC) (Michel and Brunet, 2019). For ESI of IEDs, the LSMAC was found to yield as accurate results as more sophisticated head models, namely the Boundary Element Model and the Finite Element Model (Birot et al., 2014). A grid of around 5000 sources was distributed equally throughout the grey matter mask. This grid was kept unmodified for all electrode setups to allow for comparison of localisations. Both the lead-field matrix and the inverse matrix were computed independently for each electrode setup.

Source localisation
The inverse solution was calculated using LAURA (Local Autoregressive Average), one of the state-of-the-art techniques of distributed inverse solutions which attempts to impose reliable biophysical and physiological constraints on the minimum norm algorithm (Grave-de Peralta et al., 2004). The patient's postsurgical MRI or computed tomography (CT) scan was resliced using SPM8 (Wellcome Centre for Human Neuroimaging, University College Lon- Fig. 1. A, High-density EEG arrays with 257 channels, down-sampled to 219 channels, 204 channels, and 156 channels, in 3D view. For comparison, electrodes included in the 25-channel array recommended by the International Federation of Clinical Neurophysiology (IFCN)  are given in pink. Please note that the high-density electrode array does not perfectly match the IFCN 25-channel array. -B, Example of the maximum source localising inside the resection cavity (left) and, thus, in a sublobe overlapping with by the resection (right). don, UK) and co-registered to the grid of solution points. Since we did not use any covariance matrix to correct for noise, ESI could be robustly performed at every single time point. The IED's rising phase could be as short as 20 ms (Appendix 2), so we based the evaluation on predefined time points and not on changes in the scalp voltage map. The source of maximum amplitude at 50% of the IED's rising slope (by default; or alternatively at 10%, 25%, 75%, or the peak) was identified and visually compared to the resected brain area.

Assessment of concordance
Two separate measures of concordance were applied.
-Distance: The shortest spatial distance in one of the 3 orthogonal planes (axial, sagittal, coronal) between the maximum source and the edge of the resection was assessed. If the maximum source was located within the resection cavity, the distance was considered 0 mm; otherwise distance was classified as <10 mm, <20 mm or >20 mm from the border of resection. Distances of 0-10 mm were considered concordant to the resected brain area (Fig. 1 B). -Level of precision: According to the parcellation of the grey matter, the maximum source was compared to sublobes, lobe (s), and hemisphere affected by the surgery (Fig. 1 B). Here, sublobar precision was considered concordant.
To estimate the quality of the results, concordance in favourable outcome patients (ILAE 1 + 2) was considered true positive (TP), and discordance in unfavourable outcome patients (ILAE 3-5) was considered true negative (TN). Accordingly, concordance in unfavourable outcome was considered false positive (FP), and discordance in favourable outcome was considered false negative (FN). Sensitivity was calculated as TP/(TP + FN), and specificity as TN/(TN + FP). Overall accuracy was calculated as (TP + TN)/(TP + TN + FP + FN). The diagnostic odds ratio (OR) of favourable postsurgical outcome in case of concordant vs. discordant ESI results was calculated as (TP*TN)/(FP*FN) (Sharma et al., 2019). Please note that this approach has a limitation: Unfavourable seizure outcome does not reliably prove incorrect estimation of the epileptogenic zone, and neither does discordant ESI in such a case imply that the ESI maximum lies in the true epileptogenic zone (outside the resection). Thus, the definitions of TN and specificity are less reliable than the definitions of TP and sensitivity .

Statistics
Statistical analyses were performed using SPSS 25 (IBM, Armonk, US-NY). Data are given as percent or as median and interquartile range (IQR). ESI accuracies obtained with a specific methodology (referring to electrode setup and time point) were compared intra-individually and pairwise to that achieved with the default methodology (204 channels, 50% of the IED's rising phase) using Wilcoxon test. Mann-Whitney U test was used for comparisons across patients. A value of p < 0.05 was considered statistically significant. All p-values given in this manuscript were corrected for multiple comparisons according to the Benjamini-Hochberg procedure, assuming a false discovery rate of 0.05 (Hemmerich, 2016). Graphs were created using SigmaPlot 14 (Systat Software Inc., San José, US-CA).

Average IEDs and artefacted channels
Depending on the duration of the high-density EEG recording and on the individual patient, the number of homologous IEDs ranged from 3-269 with a median of 27 (IQR 18-38; Appendix 2). For 16 patients, only the IED average was available in the database, and the number of averaged spikes could retrospectively not be ascertained. It was not possible to retrospectively determine electrode impedance values throughout the EEG recordings. For the default 204-channel array, a median of 1 channel (IQR 0-4) was visually considered artefact-ridden and required interpolation, representing 0.5% of all channels. By comparison, the median of corrupted channels was 6 for the original 257-channel array (IQR 1-13; 2.3%; p = 0.001), 3 for 219 channels (IQR 0-7; 1.4%; p = 0.018), and again 1 for 156 channels (IQR 0-4; 0.6%; p = 0.97; Appendix 2).

ESI with 204 channels at half-rise of IED
Using the ''default method", the maximum source was located inside the resection cavity in 19 patients (42%) and less than 10 mm remote from its edge in another 9 patients (20%), yielding an overall concordance of 62% in terms of distance (Fig. 2). Regarding the levels of precision, in 32 patients (71%) the maximum source was located inside a sublobe affected by the surgery (i.e., concordant), and outside the sublobes but within the affected lobe(s) in another 3 patients (7%).
Overall accuracies were similar in patients with unknown vs. known numbers of IEDs to be averaged (distance: 59% vs. 63%; level of precision: 72 vs. 75%; Supplementary Table 3 and Appendix 2). Descriptively, ESI was not more accurate in case of very large resections (e.g., patients #1 and #5) than in case of smaller resections (e.g., patients #10-27). In those 8 patients with more than one cluster of IEDs, ESI based on another IED type did not lead to more concordant results (data not shown). When levels of Tikho-nov regularization were changed manually (Michel and Brunet, 2019), ESI results did not change either (data not shown).

Different electrode setups
Individual ESI maxima gradually varied between the different electrode setups tested (Supplementary Fig. 3). In comparison to the default 204-channel setup, overall accuracies and diagnostic ORs decreased when other electrode setups were applied (Supplementary Table 2). When only patients with favourable postsurgical outcome were assessed (ILAE 1 + 2, n = 30), results obtained with 257 electrodes were significantly less precise than those based on 204 channels (sublobar precision: 57% vs. 83%, p = 0.042; Fig. 3). Otherwise, differences were statistically not significant, but the default 204-channel approach tended to perform best.

Different time points across the IED's rising phase
Time points different from 50% of the IED's rising slope did not improve ESI concordances. At 10% of the IED's rising phase, accuracy was even below 50% with 23% of ESI maxima in the hemisphere contralateral to the resection (p < 0.05; Fig. 4). Results at 25%, 75% and at the IED peak were not significantly different from the default approach at 50%.

Discussion
In our study, ESI did not perform better if information from caudal EEG channels over cheeks and necks was included. In both extratemporal and temporal lobe epilepsy, the reduced 204channel setup tended to yield the most accurate ESI results for patients who underwent successful epilepsy surgery. To our knowledge, this is the first study to show that increasing the numbers of EEG electrodes beyond 204 channels and below the recommended ''9/10" lines of the 10/10 international system  does not necessarily improve the precision of ESI.
Using our default approach of 204 EEG channels and 50% of the IED's rising slope, ESI achieved accuracies of 60-73% and diagnostic ORs of 1.8-5.7. This is in the range of previous studies, although it Fig. 2. A, individual electric source imaging (ESI) results obtained with the ''default method" of 204 EEG channels, measured at 50% of the averaged interictal epileptic discharge's (IED's) rising phase. X-axis: Individual patients, sorted by 12-month outcome according to the International League Against Epilepsy (ILAE) classification (Wieser et al., 2001). Y-axis: ESI maximum (symbol) compared to brain areas affected by resection (grey bar), given on a sublobar level, sorted by brain regions. SC, subcortical grey matter; OL, occipital lobe; TL, temporal lobe; PL, parietal lobe; FL, frontal lobe. Symbols indicate distances between ESI maximum and resected brain area: blue circle, 0 mm; green rhombus, <10 mm; yellow upward triangle, <20 mm; red downward triangle, >20 mm. B, overview on distances between maximum source and resected brain area (left, coloured bars) and levels of precision (right, greyscale bars), each given for all 45 patients, those 30 with favourable 12-month outcome only, and those 15 with rather unfavourable outcome (left to right). *, p < 0.05. Electric source imaging (ESI) concordances, sensitivities, specificities, overall accuracies, and diagnostic odds ratios for all patients and subgroups, respectively, estimated in distance between ESI maximum and resection, and in sublobar concordance. ILAE, International League Against Epilepsy surgical outcome classification (Wieser et al., 2001). TP, true positives. FP, false positives. OR, odds ratio.
B.J. Vorderwülbecke, M. Carboni, S. Tourbier et al. Clinical Neurophysiology 131 (2020) 2795-2803 is difficult to formally compare these quality measures to other studies on interictal ESI, because of different methodological approaches, e.g., different sublobar parcellations (Centeno et al., 2017), different durations of post-operative follow-up (Lascano et al., 2016), consideration of Engel class II (i.e., persistence of rare disabling seizures) as favourable surgical outcome (Brodbeck et al., 2011;Feng et al., 2016), or validation by intracranial EEG instead of post-operative outcome Koessler et al., 2015). Nevertheless, a recent meta-analysis across 19 interictal ESI studies published from 2003 to 2018, all validated by postsurgical outcome of ILAE 1 + 2, found overall accuracies of 74% (95%-CI, 70-78%) and diagnostic ORs of 4.0 (95%-CI, 2.3-7.0). These measures are comparable to ours. As in the current study, sensitivities of interictal ESI were higher than specificities (81% vs. 45%) (Sharma et al., 2019). In a newer, prospective study on simultaneous electric and magnetic source localisation that followed the rigorous STARD criteria (standards for reporting diagnostic accuracy), ESI based on up to 80 EEG electrodes, obtained with two different source models and two different software packages, yielded overall accuracies of 44-50% and diagnostic ORs of 0.5-1.3 (Duez et al., 2019). A recent study on low-density ESI reported an overall accuracy of 57-62% (Sharma et al., 2018) while another low-density ESI study found 61% for an automated ESI approach and 78% for a semi-automated method (Baroumand et al., 2018).  As expected, in the current study, caudal EEG channels were prone to artefacts. The percentage of corrupted channels was highest in the full 257-channel setup (median: 2.3%; maximum: 11%). Information from these artefacted channels had to be interpolated from their neighbours, as long as the neighbouring channels were not subject to interpolation themselves. Manual identification of artefacted channels costs working time that could be saved through down-sampling of the electrode setups. Apart from EEG artefacts, cheek and neck electrodes are located on muscles and conjunctive tissue but remote from the brain. Eyeballs, air cavities of sinuses, nose and mouth as well as the holes of the skull base are challenges for the head model (Montes-Restrepo et al., 2014). These circumstances make it difficult for a three-shell forward model like LSMAC to properly reproduce the volume conduction to caudal electrodes, what is likely to be another reason for lower ESI accuracy in the 257-channel setup.
Interestingly, in the meta-analysis cited above, high-density ESI (i.e., 64-257 EEG channels) had virtually the same accuracy as lowdensity ESI (19-32 channels) although patients populations were not detailed in both groups, particularly with respect to proportions of temporal lobe epilepsy and extratemporal epilepsy (Sharma et al., 2019). In the light of our current findings, it seems possible that for ESI based on the full 128-and 257-channel setups, the advantages of high electrode density were counterbalanced by inaccurate signal recording and processing from very caudal face and neck electrodes. On the other hand, inferior temporal and occipital regions do need to be covered for ESI, since our 156channel array located above the hat brim line also led to less accurate results than the 204-channel setting. Thus, a relocation of high-density EEG electrodes from cheeks and neck towards the scalp, leading to smaller interelectrode distances within the borders of the 204-or 219-channel setup, could improve ESI accuracy.
As a limitation of our study, there were no cases with fronto-or occipito-basal epilepsy and favourable surgery outcome. Thus, to compare epileptic activity in basal structures vs. the convexity, we relied on the comparison of temporal to extratemporal cases. ESI in temporal lobe epilepsy had higher sensitivities (79-100% vs. 46-55%), overall accuracies (67-82% vs. 50-61%) and ORs for favourable outcome (2.3-1 vs. 1.1-3.0) than in extratemporal epilepsy, but lower specificities (38% vs. 57-71%; Table 1, Fig. 3 BC). This was already the case in some previous studies (Brodbeck et al., 2011;Coutin-Churchman et al., 2012;van Mierlo et al., 2017;Toscano et al., 2020) but not in others Abdallah et al., 2017). In an earlier work on interictal lowdensity ESI in children, accuracy was markedly higher in extratemporal vs. temporal lobe epilepsy (Sperli et al., 2006). The authors argued that in temporal lobe epilepsy, a lack of inferior temporal EEG electrodes led to a shift of ESI maxima from basal temporal to extratemporal sources. In our current study, this shift seems reversed: Inclusion of very caudal electrodes as in the 257-and the 219-channel setup led to a shift of ESI maxima towards the temporal lobes even in extratemporal epilepsy ( Supplementary  Fig. 3). However, the accuracy of ESI in extratemporal epilepsies has only been assessed in small stand-alone studies or as a small subset of patients in larger clinical cohorts that mostly included temporal cases. Thus, the yield of ESI in extratemporal epilepsies needs to be better assessed in a larger study.
In a recent ESI study based on automatically detected IEDs, most accurate results were found at the IED peak (van Mierlo et al., 2017). This was most likely attributable to an optimal signal-to-noise ratio at that time point, since IEDs were centred around the peak for averaging. Although we did the same, ESI accuracy in our study decreased at time points later than 50% of the IED's rising phase. The most likely explanation for this is propagation: During the time course of an IED's upswing, the source of maximum activity can move away from the IED origin (Merlet et al., 1997;Lantz et al., 2003b;Bast et al., 2006;Plummer et al., 2010).
Consistent with this, another recent article on electric and magnetic source localisation reported best results at very early time points of IEDs (Plummer et al., 2019). In our study again, earlier time points than 50% yielded less accuracy. In a quarter of successfully operated patients, source maxima at 10% of the IED's rising phase were even contralateral to the resection (Fig. 4, Supplementary Fig. 4). This points to a relatively high instability of the source at IED onset where the signal-to-noise ratio is well-known to be low (Bast et al., 2006;Ray et al., 2007;Wennberg and Cheyne, 2014). The discrepancy between the two articles may be due to a higher overall signal-to-noise ratio in the other study: There was a 3-times higher number of IEDs per average (median: 76) than in ours (median: 27). In addition, IEDs were described as ''relatively complex with polyphasic components preceding a dominant negative-peak". This suggests a high signal-to-noise ratio already at early time points of the IED, compared to rather simple, biphasic spikes or sharp waves in our data. In our case, 50% of the IEDs rising phase seemed an optimal compromise between little source propagation (early stages) and high signal-to-noise ratio (towards the peak). In general, selection of the time point may have to be adapted to the number of IEDs per average and to the IED morphology which both affect the signal-to-noise ratio.
At least for low-density ESI, a minimum of 8-25 single IEDs have been recommended for IED averaging to yield acceptable signal-to-noise ratios and both reliable and valid localisation results (Bast et al., 2006;Wennberg and Cheyne, 2014). Unfortunately, six patients of our study cohort had less than 10 IEDs to be averaged, and the number of IEDs could retrospectively not be ascertained in another 16 patients. Still, ESI in these specific cases was as accurate as in the other patients with more than 10 IEDs ( Supplementary Fig. 5).
For our study, we used a methodology that is easy to handle and, thus, most feasible in clinical routine: We took averaged IEDs and a simplified realistic 3-shell head model, and we considered the spatial ESI maximum within a distributed inverse solution at one specific time point. We welcome future studies to replicate our findings using different methodological approaches, e.g. more sophisticated head models, single dipole inverse solutions, and/or ictal EEG. For our approach, we can conclude that adding information from caudal EEG electrodes over cheeks and neck does not necessarily improve accuracy of high-density ESI, most likely because of both higher artefact load and suboptimal biophysical modelling for the lower skull structures.