Nine decades of electrocorticography: A comparison between epidural and subdural recordings

In recent years, electrocorticography (ECoG) has arisen as a neural signal recording tool in the development of clinically viable neural interfaces. ECoG electrodes are generally placed below the dura mater (subdural) but can also be placed on top of the dura (epidural). In deciding which of these modalities best suits long‐term implants, complications and signal quality are important considerations. Conceptually, epidural placement may present a lower risk of complications as the dura is left intact but also a lower signal quality due to the dura acting as a signal attenuator. The extent to which complications and signal quality are affected by the dura, however, has been a matter of debate. To improve our understanding of the effects of the dura on complications and signal quality, we conducted a literature review. We inventorized the effect of the dura on signal quality, decodability and longevity of acute and chronic ECoG recordings in humans and non‐human primates. Also, we compared the incidence and nature of serious complications in studies that employed epidural and subdural ECoG. Overall, we found that, even though epidural recordings exhibit attenuated signal amplitude over subdural recordings, particularly for high‐density grids, the decodability of epidural recorded signals does not seem to be markedly affected. Additionally, we found that the nature of serious complications was comparable between epidural and subdural recordings. These results indicate that both epidural and subdural ECoG may be suited for long‐term neural signal recordings, at least for current generations of clinical and high‐density ECoG grids.

monkeys by British physician Richard Caton in 1875 (Caton, 1875) and in humans by German psychiatrist Hans Berger in 1924 (Jung & Berger, 1979).Electrocorticography (ECoG), as it became known, was later popularized as a diagnostic tool to identify epilepsy seizure onset zones and for the functional mapping of eloquent cortical areas in patients with epilepsy by applying electrical stimulation currents between pairs of electrodes (Lesser et al., 1984;Reif et al., 2016).
ECoG grids consist of circular electrically conductive disks (electrodes) embedded in silicon sheets that are placed under the skull onto the surface of the brain.Compared with non-invasive neural signal recording methods such as scalp electroencephalography, ECoG recordings are highly specific to the tissue immediately underneath electrodes (high spatial resolution) (Crone et al., 1998;Freeman et al., 2000;Lesser et al., 2010;Leuthardt et al., 2004;Miller et al., 2009), and the signal amplitude is up to five times higher than that of scalp electrode recordings (Blume & Holloway, 2011).ECoG electrodes can be placed either on top of (epidural) or below (subdural) the dura mater (Figure 1a) and exist in many configurations, typically grids of N Â M electrodes, with N, M > 1, or strips of 1 Â N electrodes (Figure 1b).Standard clinical grids and strips used for cortical mapping and epilepsy monitoring have an inter-electrode distance (IED; or pitch) of 10 mm (Diehl & Lüders, 2000;Lesser et al., 2010;Penfield & Boldrey, 1937;Salles et al., 1994) and an exposed electrode surface diameter of 2-5 mm.In the last decade, electrode grids with smaller disks and inter-electrode space have also been introduced (with ≤4-mm IED and diameter < 1.3 mm).Hereafter, these are referred to as high-density grids (Figure 1b) (Bouchard et al., 2013;Bouchard & Chang, 2014;Flinker et al., 2011;Kellis et al., 2016;Slutzky et al., 2010;Torres Valderrama et al., 2010;Wang et al., 2016).
In the early 1930s, epidural placement of ECoG electrodes was common in clinical practice (Almeida et al., 2005;Foerster & Altenburger, 1935).In the decades thereafter, both subdural and epidural recordings were employed and arguments favouring epidural or subdural ECoG seemed to be driven by the technical advancements of individual clinical centres (e.g., Goldring &Gregorie, 1984, for epidural ECoG andWyler et al., 1984, for subdural ECoG).In the course of time, subdural recording became increasingly favoured, primarily because subdural cortical stimulation does not stimulate meningeal pain fibres (in contrast to epidural recordings), and anatomical landmarks are easier to identify when the brain is not covered by the dura during surgical electrode placement (Benbadis, 2000;Wirth & Van Buren, 1971;Zumsteg & Wieser, 2000).These factors have allowed subdural ECoG to be not only suitable for diagnostic purposes for patients with epilepsy (Jacobs et al., 2010;Lesser et al., 2010) but also for the fundamental study of perceptual, cognitive and motor brain functions (Crone et al., 2006;Jacobs & Kahana, 2010;Jerbi et al., 2009).
Today, ECoG is increasingly used beyond its diagnostic and fundamental research purposes.In particular, ECoG has arisen as a valuable chronic neural signal recording modality for neural interfaces, such as braincomputer interfaces (BCIs) (Benabid et al., 2019;Matsushita et al., 2018;Miller et al., 2020;Silversmith et al., 2021;Vansteensel et al., 2016;Wang et al., 2013).Most current ECoG-based chronic recording research is based on subdural recordings.However, it has often been suggested that epidural placement of ECoG may be associated with fewer infections and complications (e.g., Bundy et al., 2014;Fischer et al., 2019;Martens et al., 2014;Moran, 2010), which is conceptually attractive for the novel chronic applications of ECoG, including BCI.Yet, it is currently unclear if that is really the case and whether epidural recordings offer similar signal quality and precision as subdural ECoG electrodes.For future clinical implementation of ECoG-based neural interfaces, it is pertinent to assess if epidural recordings pose a viable alternative to subdural ones.
In this review, we assessed the influence of the dura mater by comparing studies that employ epidural ECoG to those using subdural ECoG recordings.More specifically, we report on several aspects, such as (1) signal quality (spatial resolution and signal amplitude), (2) signal decodability (offline and online classification and regression), ( 3) longevity (tissue response and impedance) and (4) type of complications (in acute and chronic settings).For that, we included studies on both acute and chronic ECoG recordings in human and non-human primates (NHPs) and discuss implications of our findings for the design of chronic ECoG-based neural interfaces.

| LITERATURE REVIEW SELECTION AND OUTPUT
We performed a systematic search (last search date 10th January 2021) on ECoG studies using four inclusion criteria.First, the studies should be based on ECoG recordings.Second, the subjects of the studies should be either humans or NHPs (DeCasien et al., 2017;Dunbar & Shultz, 2017;Jerison, 1973).Third, one or more signal properties should be described in the studies.This means that we searched for studies where specific mention was made of signal metrics, such as amplitude, coherence, power spectrum or the decoding of brain states.Lastly, only articles written in English were included.Of note, as the term 'ECOG' is also used in studies related to oncology, a filter was applied to leave oncology-related studies out of the database.Keywords and synonyms used as search terms can be found in Table S1.We applied seven exclusion criteria, namely, (1) no original contribution is made to the literature (this criterium includes records such as comments, errata, notes and conference reviews); (2) the topic of the paper does not relate to neuroscience (e.g., oncology); (3) the study discusses recordings in a study population other than humans or NHPs; (4) a recording method other than ECoG is used; (5) the study does not discuss neural recordings; or (6) the full text of the study is unavailable.Although we focus on neural signal recording using ECoG, articles that also discussed stimulation paradigms were not excluded, as they may contain relevant information for this paper.Five databases were used to screen title, abstract and keywords: Embase, PsycInfo, PubMed, Scopus and Web of Science.Additionally, we performed a search in Google Scholar and inspected the reference list of eligible papers.As Google Scholar performs full-text searches, we selected the top 300 results (based on Google's proprietary algorithm, which determines the article's relevance based on the full text, author list, journal and how often and how recently the article was cited) for title and abstract screening (Figure 2).
The extracted data were uploaded to Rayyan (www.rayyan.ai),a web-based tool for literature reviews (Ouzzani et al., 2016).The search resulted in a total 3904 unique articles.Based on title and abstract screening, 2515 of these were included: 55 discussed epidural ECoG, 2452 subdural ECoG and 8 both epidural and subdural ECoG, indicating that to date, the application of epidural ECoG to clinical and research practice has been scarce in comparison with its exponentially increasing subdural counterpart (Table 1 and Figure 3).Of note, most of the identified epidural studies originated from one of four research labs.
Here, we used the 63 papers discussing epidural recordings to investigate signal quality obtained with epidural recordings and to structure this review article.Epidural signal quality characteristics were compared with those obtained with similar subdural studies, which were selected from the 2460 subdural reports obtained in the systematic search.For that, the most recent state-of-theart subdural studies that discussed similar topics as the ones found in the epidural studies are described.In addition, to assess any difference in the number and nature of complications that arise with epidural and subdural recordings, we made a selection of ECoG studies that reported complications related to acute (shorter than 30 days) and chronic (longer than 30 days) recordings from the 2515 included articles.Moreover, we noted that some clinical studies did not discuss signal properties and therefore were not included in our initial search.Because these articles were potentially relevant for the complications assessment, we conducted an extra search specifically on complications in clinical subdural and epidural studies.

| SIGNAL QUALITY OF EPIDURAL RECORDINGS
A key property of a viable ECoG-based neural interface is the quality of the recorded signals.Although this quality can be described by different properties or features, in this review, we focus on two features that are frequently reported to be affected by the dura, in particular (1) the specificity of recorded signal to the tissue immediately underneath an electrode (denoted here as 'spatial resolution') and (2) the signal strength as defined by the amplitude of the recorded electrical potential (denoted here as 'amplitude').

| Spatial resolution
The possibility to discriminate healthy from epileptic neural activity and delineate epileptic foci using epidural ECoG at millimetre precision has been shown in the early 1970s (Harris, 1972).More recent modelling and NHP studies further addressed the minimum epidural IED between electrodes where no signal overlap occurs across electrodes and where electrodes show independent activity.Several investigators indicated the minimum epidural IED to be below 5 mm (Kim et al., 2007;Rouse, 2012;Rouse et al., 2016;Slutzky et al., 2010;Thongpang et al., 2011).An interesting observation was made by Slutzky and colleagues, who suggested in their human modelling work that the distance between the epidural recording electrode and the signal source was strongly driven by the cerebrospinal fluid (CSF) layer.The authors created a finite element model of the human F I G U R E 2 Review selection and output.Systematic review flow diagram for queries of databases and registers, adapted from Prisma 2020 (Page et al., 2021).Details of the search output are described in the text.
T A B L E 1 Summary of the epidural studies found in the systematic literature search.head and systematically decreased the thickness of the CSF layer in their model.They concluded that the thickness of the CSF layer affected the minimum IED more than the mere presence of the dura (Slutzky et al., 2010).In fact, they showed that as the CSF layer decreased from 3.1 to 0.2 mm, the minimum epidural IED sharply decreased from 9.0 to 3.2 mm, approaching their estimated minimum IED value of subdural electrodes (2.6 mm).Although the authors stated they considered the thinning of the CSF a realistic scenario given the pressure applied by the implanted grids, this observation has not been confirmed in humans or NHPs.Slutzky and colleagues also attempted to estimate the maximum IED required to avoid spatial aliasing (i.e., insufficient spatial sampling) based on a spatial spectral analysis.This metric attempts to characterize the notion of intrinsic resolution of neuronal activity patterns.They suggested that the maximum epidural IED to avoid spatial aliasing in humans would be approximately 1.4 mm, close to the 1.25-mm maximum value earlier computed for subdural electrodes (Freeman et al., 2000).Although the interpretation of the relation between the minimum and maximum IEDs may require additional investigation, this relation seems similar for epidural and subdural ECoG.Building on abovementioned findings, Fisher and colleagues showed that nine adjacent brief visual stimuli could be decoded from primary visual cortex of NHPs (Fischer et al., 2019) using epidural ECoG with an IED of 1.8 mm (Schander et al., 2019;Strokov et al., 2017).Furthermore, when looking specifically at the higher frequency band range of the ECoG signals in NHPs (75-105 Hz), Rouse and colleagues showed a decline in signal correlation between pairs of epidural electrodes as the IED increased, which was particularly steep between 3 and 6 mm (Rouse, 2012;Rouse et al., 2016).Additionally, they reported a decrease in performance of one-dimensional BCI cursor control of around 5% when using a bipolar electrode pair with an IED of 3 mm compared with electrode pairs with larger IEDs (9 and 15 mm), suggesting that the limit of discriminating signals from separate spatial locations with epidural high-frequency band ECoG lies around this value.Taken together, the studies described above suggest that ECoG signals recorded epidurally can be discerned at $3-mm precision, which is comparable with the precision of subdural recordings (2.6 mm).In these studies, recordings were made from the tips of cut flush microwires.In this review, we grouped these under high-density recordings.f Slutzky et al. (2010) do not describe human or non-human primate recordings and were therefore not part of the output of our literature search.

Species
However, as the results of the modelling study are highly relevant, the results are described in the paper.
A factor that plays an important role in the amplitude of both epidurally and subdurally recorded signals is the diameter of the exposed electrode surface.Indeed, Tolstosheeva and colleagues showed, using epidural ECoG grids in NHPs with electrodes of varying diameters (100, 300 and 500 μm in diameter), that larger electrode diameters were associated with larger signal amplitude in the high-frequency band range between 50 and 100 Hz (as reflected in higher power) (Tolstosheeva et al., 2015).These results are similar to those found for NHP subdural ECoG, for which it was suggested that the cortical tissue around an electrode that contributes to its activity (i.e., cortical spread) linearly increases with increasing electrode surface (Dubey & Ray, 2019).
Not surprisingly, the dura also affects the amplitude of recorded signals.Ryapolova-Webb et al. (2014), for example, reported a significantly lower signal amplitude for epidural recordings than for subdural recordings in an NHP implanted with standard clinical ECoG grids.Interestingly, however, the influence of epidural electrode placement on signal amplitude seems especially evident for smaller electrode diameters (Bundy et al., 2014).Indeed, Bundy and coworkers reported that, although for standard clinical grids (2.3-mm exposed diameter) the epidural signal is attenuated to some extent for frequencies above 60 Hz compared with the subdural signal, this attenuation is stronger and present for a broader range of frequencies (0-200 Hz) for epidural grids with small electrode diameter.This is likely due to the lower signal amplitudes of smaller electrodes, which approach the amplifier noise floor (i.e., level of background noise) at lower frequencies for epidural recordings when compared with subdural recordings.This result is in line with a previous study that estimated a massive 80% signal attenuation by the dura in an anaesthetized human implanted with a high-density ECoG grid (4-mm diameter and IED of 5 mm) for frequencies between 5 and 100 Hz (Torres Valderrama et al., 2010).Yet, significant signal attenuation does not necessarily mean loss of detectable responses.The abovementioned article from Tolstosheeva et al. (2015), for example, reported that epidural recordings were able to detect clear responses to standard visual stimuli in the visual cortex using each of the three electrode diameters.Taken together, the results suggest that the dura can attenuate the neural signal considerably, mostly affecting highfrequency bands in clinical grids and affecting the whole frequency spectrum in small electrode diameters.

| SIGNAL DECODABILITY FOR EPIDURAL RECORDINGS
In general, two decoding strategies are used in the BCI field: decoding discrete classes and regressing continuous variables.While the former directly converts discriminable classes into independent degrees of freedom, the latter allows to (continuously) predict a quantity that can be translated as continuous n-dimensional control of a computer cursor or end effector.Either one can be accomplished offline or online, that is, with or without providing real-time feedback to the subject controlling the BCI.Below, we summarize offline and online classification and regression studies.For an overview, see also Tables S2-S4.

| Offline classification studies
Several human and NHP epidural studies with clinical and high-density grids have accomplished levels of decoding of upper limb movements (Baker et al., 2009;Choi, Kim, et al., 2018;Choi, Lee, Park, Lee, et al., 2018;Larzabal et al., 2021;Spüler, Walter, Murguialday, et al., 2014;Thomas et al., 2019) that are comparable with those for subdural electrodes (Gruenwald et al., 2019;Jiang et al., 2017Jiang et al., , 2018;;Kub anek et al., 2009;Shiraishi et al., 2020;Yao & Shoaran, 2019).Of note, one study Histogram of subdural (blue) and epidural (orange) studies per year included in this review.For simplicity, studies discussing both types of recordings are included in both curves.Year points after 2019 are dashed because the number of outputs of those years may have been incomplete at the time of the search or decreased because of the COVID-19 pandemic.Of note, only studies where signal metrics are described or where the electrocorticography (ECoG) signal is applied for the classification of brain states were included in this search.
described both epidural and subdural classifications of finger, wrist and elbow flexion and extension in abledbodied participants using clinical and high-density grids (Thomas et al., 2019).This study showed an epidural accuracy (93%) in the same range as the subdural accuracies (71-100%).Other examples of epidural studies are Larzabal et al. (2021), who reported up to 60% classification accuracy for 12 imagined upper-limb movements, including finger movements, from long-term high-density epidural ECoG in an individual with tetraplegia, and Spüler, Walter, Murguialday, et al. (2014), who decoded seven classes (on average 61% accuracy) in severely paralysed individuals implanted with epidural clinical grids, at a level similar to that described for subdural counterparts (e.g., 80% accuracy for five classes; Kub anek et al., 2009).In Spüler, Walter, Murguialday, et al. (2014), the highfrequency band contributed most to the decoding accuracy, and, importantly, frequencies up to 150 Hz still contributed useful information.Additionally, Choi, Lee, Park, Lee, et al. (2018) achieved 84% average accuracy in classifying four classes of (bi)manual movement in NHPs with epidural recordings.
Other human studies aimed to identify cortical states over the course of the progression of amyotrophic lateral sclerosis, using low frequencies recorded with epidural grids.Among other results, they described three statedependent frequency domains (<4 Hz for sleep-like periods, $7 Hz for low arousal and $20 Hz for elevated arousal) (e.g., Martens et al., 2014), showing for the first time the usability of epidural ECoG to detect cortical states in individuals with complete paralysis.
Finally, using high-density epidural grids in NHPs, researchers have attempted to decode saccadic eye movement directions from the frontal eye fields, supplementary eye fields and superior parietal lobes (Baek et al., 2014;Lee et al., 2013Lee et al., , 2017)), to decode the direction of visual attention from visual cortex (Rotermund et al., 2013) or to estimate the visual receptive field (Fischer et al., 2019;Schander et al., 2019;Strokov et al., 2017).Although it was speculated in one study that the dura may have diminished the use of higher frequencies for decoding (Lee et al., 2017), others found nearperfect classification of visual stimuli (>95% classification accuracy for 25 stimuli and >99% for 13 stimuli) using high-frequency band modulations recorded from the visual cortex (Fischer et al., 2019) and 87% average accuracy in classification of four eye-movement directions (Baek et al., 2014).
Altogether, the above studies show that both highfrequency and lower frequency band features recorded with clinical and high-density epidural grids can be used for successful decoding.Moreover, the available evidence suggests that the dura has limited influence on decoding ability, which is in agreement with the findings on spatial resolution described in Section 3.

| Offline regression studies
With respect to the decoding of continuous variables from the epidural space, Shimoda et al. (2012) were the first to decode continuous three-dimensional hand trajectories using high-density epidural ECoG grids with 3.5-mm IED from NHP.The results, although encouraging, were lower than previously reported on the same task using subdural ECoG over similar cortical areas (Chao et al., 2010).The authors made clear that a considerable amount of chewing artifacts in the epidural dataset was an important cause for decreased decoding accuracy.The epidural dataset (publicly available; Nagasaka et al., 2011) gave origin to a series of follow-up studies that attempted to improve decoding performance (Eliseyev & Aksenova, 2014, 2016;Eliseyev et al., 2017;Engel et al., 2017;Farrokhi & Erfanian, 2018;Foodeh et al., 2020), reaching an average correlation coefficient of 0.65 between actual and predicted hand trajectory (Farrokhi & Erfanian, 2018).In comparison, the method applied by Farrokhi and Erfanian (2018) resulted in an average correlation coefficient of 0.76 in the dataset acquired during the same task with subdural ECoG.Remarkably, Farrokhi and Erfanian note that useful information was present in the 200-to 400-Hz frequency band, both in the epidural and subdural recordings.Other NHP and human epidural datasets have been used to decode two-dimensional position of one arm (Flint et al., 2012;Marathe & Taylor, 2013), the wrist (Spüler, Rosenstiel, & Bogdan, 2014;Spüler et al., 2016) or the three-dimensional trajectory of both arms (Choi, Kim, et al., 2018) using grids with an IED ranging from 1 to 10 mm.The results of these studies displayed similar, albeit slightly inferior (range of correlation coefficients for epidural r = $0.3-0.7 < subdural r = $0.5-0.9,where the reported r2 scores were translated to r; Table S3), results compared with NHP and human subdural studies that decoded trajectories of the arm (e.g., Nakanishi et al., 2017;Shin et al., 2018;Talakoub et al., 2017) and fingers (e.g., Flint et al., 2020;Xie et al., 2018).Of particular interest is a study from Flint et al. (2017) who attempted to directly compare performance between epidural and subdural recordings using both standard clinical and high-density grids (4-mm IED) to decode continuous grasp kinematics.They reported that highdensity grids outperform standard grids and that epidural performance is similar to subdural performance (Table S3).This indicates that even though signals recorded by the epidural high-density grids may be attenuated compared with those of the subdural grids (Bundy et al., 2014), decoding results are not necessarily greatly affected.In sum, the above studies support the premise that epidural recordings can yield similar, albeit often slightly lower, decoding performance compared with subdural recordings, especially for high-density grids.

| Online classification and regression studies
Following the first epidural demonstration of real-time one-dimensional cursor control in the 2000s in humans (Leuthardt et al., 2006), several groups have confirmed that epidural (clinical and high-density) ECoG can be used for decoding of several kinematic and kinetic aspects of reaching movements in real time (Karande, 2016;Marathe & Taylor, 2013;Williams, 2013), to study synchrony and plasticity during BCI control (Moran, 2010;Rouse, 2012;Rouse et al., 2013Rouse et al., , 2016;;Rouse & Moran, 2009) and to distinguish idle from active states (Gomez-Rodriguez et al., 2010;Williams, 2013;Williams et al., 2013).Also, work from Gharabaghi and colleagues has shown that the real-time control of robotic prosthetics in humans who had a stroke or an amputation of the upper limb is feasible with epidural recordings using motor imagery of opening and closing of the hand, with up to 95% decoding accuracy of movement initiation (Gharabaghi, Naros, Khademi, et al., 2014;Gharabaghi, Naros, Walter, Grimm, et al., 2014;Gharabaghi, Naros, Walter, Roth, et al., 2014).The group also closed the BCI loop by additionally stimulating the sensorimotor cortex for rehabilitation purposes (Gharabaghi, Naros, Walter, Roth, et al., 2014;Walter et al., 2013Walter et al., , 2012) ) and, perhaps surprisingly, reported no pain related to epidural stimulation.Similar to the offline studies described above, results of these epidural studies leveraging real-time control of robotic prosthetics were comparable ($80%) with those reported in subdural studies that used similar strategies (Bashford et al., 2018;Degenhart et al., 2018;Silversmith et al., 2021), such as a recent study that demonstrated >85% accuracy in real-time click control using subdural ECoG (Silversmith et al., 2021).Perhaps, the most outstanding epidural achievement so far has been the first successful long-term epidural high-density ECoG implant for control of an exoskeleton (Benabid et al., 2019).This system, known as WIMAGINE (Mestais et al., 2015), was first validated in NHPs (Eliseyev et al., 2014) and subsequently implemented by Clinatec ® in a paralysed human (Benabid et al., 2019).This study was the first ECoG-BCI to increase the number of degrees of freedom to 8 (bimanual control in the x-, y-and zdirections, plus pronation and supination of the arms) and accomplished an accuracy of 71% on average, which is considerably higher than homologous subdural reports (Degenhart et al., 2018;Wang et al., 2013).Overall, the above findings indicate, once more, that performance of epidural ECoG decoding is not fundamentally inferior compared with subdural ECoG decoding and therefore carries promise for control of complex prosthetics and exoskeletons.

| LONGEVITY OF EPIDURAL RECORDINGS
In the development of clinically viable BCI systems, maximizing the longevity of the signal is of critical importance to minimize the degradation of the system's performance over time and the resulting need for reimplantation surgery.

| Tissue response and signal quality
The most direct way to observe the long-term effect of electrode implantation on the brain is via histological analysis of the tissue around the electrodes after explantation.Although such procedures have not been performed on deceased human subjects with ECoG implants, several studies in NHPs have reported on the effects of long-term grid implantation on the brain tissue.One study, by Mestais et al. (2015), presented a histological analysis after a grid had been implanted in the epidural space of an NHP for 26 weeks.The authors show no tissue response nor damage due to the implantation of an epidural grid.Interestingly, although this study did not find evidence for a significant tissue response to electrodes placed epidurally, subdural placement in NHP has often been associated with encapsulation of the electrodes after at least 30 weeks (Degenhart et al., 2016;Romanelli et al., 2018;Ryapolova-Webb et al., 2014).Encapsulation refers to the formation of a new dural layer around the electrodes, often encapsulating these.Romanelli et al. (2018), however, argue that the neoformation of dura is specifically linked to NHPs, which show intense tissue regrowth at the site of dural opening.Coinciding with the development of this newformed tissue, some of these studies also reported a gradual decrease of signal amplitude and power on all frequencies, but well detectable movement-related power changes over 2 years after implantation (Degenhart et al., 2016;Ryapolova-Webb et al., 2014).As such, the exact relation between encapsulation and signal quality remains unclear.Taken together, although the limited available evidence suggests that epidural ECoG placement is less susceptible to encapsulation in NHP, tissue responses to both subdural and epidural human ECoG are still largely unknown.

| Impedance and signal stability
An indirect way to assess the stability of ECoG signals is by comparing the impedance between electrodes or the average amplitude of the electrophysiological signal over time.Thongpang et al. (2011) reported stable power of neural signals over a period of 8 weeks with an epidural high-density grid in NHP.This fits with other NHP epidural studies that show stable impedance over 4 months (Choi, Kim, et al., 2018) and a year (Mestais et al., 2015).However, another group has reported large deviations in epidural electrode impedance values in a macaque monkey (up to 10 MΩ) over several months after implant (Schander et al., 2019).In contrast to epidural recordings, a gradual decrease of signal (power) amplitude on all frequencies has been observed for NHP subdural recordings, despite stable impedance, stabilizing after around 300 days (Degenhart et al., 2016).
In studies on human chronic epidural (Benabid et al., 2019;Bensch et al., 2014;Martens et al., 2014;Murguialday et al., 2011;Soekadar et al., 2013) and subdural (Bergey et al., 2015;Morrell, 2011, p. 19;Nair et al., 2020;Pels et al., 2019;Swann et al., 2018;Vansteensel et al., 2016) ECoG implants for BCI or neuromodulation purposes, impedance seems to settle at stable values after an initial period of fluctuation.As an example, our group has used a subdural ECoG implant connected to a commercial amplifier (Activa ® PC+S) to restore communication in an individual with locked-in syndrome (Utrecht NeuroProsthesis; Vansteensel et al., 2016).In this first in man study, we showed that impedance stabilizes after approximately half a year (Pels et al., 2019), which is in line with the changes in subdural recordings reported by others also using the same electrode type (Sillay et al., 2013;Swann et al., 2018).Interestingly, in some studies, the changes in impedance were accompanied by changes of the neural signals.Indeed, the Utrecht NeuroProthesis showed a slow decrease in the neural signals over time over more than 12 months of implantation (Aarnoutse et al., 2019;Pels et al., 2019), which did not seem to affect performance of the system (Aarnoutse et al., 2019).In addition, both NeuroPace ® and NeuroVista ® studies (all subdural recordings) report an initial gradual decrease of spectral band power in the first half year after implantation followed by a stabilization (Sun et al., 2018) and small changes in effective signal bandwidth on a group level (Nurse et al., 2018).Intriguingly, however, the NeuroVista ® study further reported substantial signal variability in amplitude and power values across participants, which for some participants (2/15) remained unstable over the course of the 18-month study period (Nurse et al., 2018;Ung et al., 2017) suggesting that ECoG-based stability likely depends on individual factors, such as histological changes, surgical complications or electrode location, and perhaps on lead integrity and amplifier stability.
Another measure of signal stability is decoding performance.Several epidural studies reported a stable decoding performance over a duration of up to 6 months in NHPs (Choi, Kim, et al., 2018;Shimoda et al., 2012) but report no data after 6 months.On the other hand, subdural NHP studies reported that movement-related subdural signals remained well detectable over 2 years of implantation (Degenhart et al., 2016;Ryapolova-Webb et al., 2014), despite the initial signal amplitude decrease during the first 300 days (Degenhart et al., 2016).In humans, Martens et al. (2014) were able to identify three levels of arousal as an individual with locked-in syndrome transitioned into complete locked-in syndrome over the course of 6 months, whereas Benabid et al. (2019) presented stable exoskeleton control using epidural ECoG in a tetraplegic patient for 2 years.Similarly, click-based decoding performance in an individual with locked-in syndrome implanted with subdural ECoG remained stable over 3 years (Pels et al., 2019).Conjointly, the above studies suggest that both epidural and subdural ECoG can provide robust and stable long-term decoding performance.

| ECOG-RELATED COMPLICATIONS
The (nature of) complications of long-term implantation of ECoG grids are of critical importance in the development of clinically viable BCI and neuromodulation systems.The use of ECoG can be acute (up to 30 days) or chronic (longer than 30 days).Acute implantations typically relate to intraoperative recordings (e.g., during tumour brain surgery, for localization of epilepsy foci or brain function) or to extra-operative (or post-surgical) epilepsy recordings, whereas chronic implantations typically relate to BCI or closed-loop neuromodulation systems.To our knowledge, there are no studies directly comparing complications in epidural versus subdural placement of ECoG grids.In an attempt to fill this gap, we have collected and summarized complications associated with acute and chronic epidural and subdural ECoG studies (see Table 2 for a selection of studies).

| Complications associated with acute recordings
Several studies have reported on complications associated with acute ECoG recordings.Early epidural and subdural studies on epilepsy monitoring have already shown that ECoG is a relatively safe procedure, even with relatively large and stiff grids (Goldring, 1978;Goldring & Gregorie, 1984;Wyler et al., 1984).As far as we are aware, the respective epidural studies are the only ones reporting on complications with this approach.The authors reported merely two complications (wound infection and aseptic necrosis of the bone flap) over 100 patients (Goldring, 1978;Goldring & Gregorie, 1984).In the last decades, most acute recordings have been performed subdurally.A number of subdural studies report infections, intracranial haemorrhages and intracranial haematomas related to subdural invasive monitoring (Behrens et al., 1997;Burneo et al., 2006;Fountas & Smith, 2007;Hamer et al., 2002;Johnston et al., 2006;Nagahama et al., 2018;Panov et al., 2017;Tandon et al., 2019;Tanriverdi et al., 2009;Van Gompel et al., 2008;Wong et al., 2009), as well as a limited number of permanent neurologic deficits and deaths (Hamer et al., 2002;Van Gompel et al., 2008;Wong et al., 2009).Some studies have suggested that the risk of complications tends to increase with total area of the grid, number of electrodes and number of cables passed through the skull and skin (Hamer et al., 2002;Nagahama et al., 2018;Rahman et al., 2013;Wellmer et al., 2012;Wiggins et al., 1999).In two studies, some patients showed, upon electrode removal, a thin layer of blood underneath the implanted subdural grid, along with an indentation of the underlying cortex (Fountas & Smith, 2007;Hudgins et al., 2016), presumably caused by irritation of the cortex by the subdural grid.Unfortunately, minor complications, including, for example, headache or CSF leakage, are often not reported in these large studies.The low number of complications reported in the early acute epidural studies is striking but given the overall lack of documented complications associated with acute epidural recordings in the last decades, it is hard to assess how the complication rates of acute epidural and subdural recordings really compare.

| Complications following chronic implants
If discussed at all, the majority of chronic epidural and subdural NHP studies reports no instance of inflammation nor tissue reaction due to the ECoG grids (Degenhart et al., 2016;Marathe & Taylor, 2013;Mestais et al., 2015;Romanelli et al., 2018;Ryapolova-Webb et al., 2014).However, three studies have reported complications: detachment of the array and subsequent need for epidural reimplantation to reattach the array (Marathe & Taylor, 2013); the formation of granulation tissue (tissue regrowth to close an open wound) and effusion (outpouring of fluid) occurring around the epidural electrodes over 4 months of implantation (Shimoda et al., 2012); and minor inflammatory processes underneath the subdural grids (Romanelli et al., 2018).
With respect to human epidural chronic studies, no complications arose during the studies by the groups of Benabid (one patient) (Benabid et al., 2019) and Gharabhaghi (one patient) (Martens et al., 2014).A risk factor related to epidural recordings seems to relate to the application of cortical stimulation (Tsubokawa et al., 1993).Indeed, non-habitual seizures were the most occurring complication reported in this type of studies.These are reported 10 times (Levy et al., 2008(Levy et al., , 2016;;Rasche et al., 2006) in the 207 epidural cortical stimulation patients included in Table 2 (Katayama et al., 1998;Levy et al., 2008Levy et al., , 2016;;Nguyen et al., 1999Nguyen et al., , 2008;;Rasche et al., 2006;Tsubokawa et al., 1993).In detail, Rasche et al. (2006) report the occurrence of intraoperative epileptic seizures evoked by direct epidural cortical stimulation in seven patients.Levy et al. (2008Levy et al. ( , 2016) ) report in total three seizures in the investigational groups, versus zero in the control groups.
Regarding chronic human subdural studies, a large patient population is described in the NeuroPace ® studies investigating a neuromodulation system to reduce refractory epileptic seizures.According to the authors, a considerable number of the adverse events could be attributed to epilepsy (a complete overview of the complications in the NeuroPace ® studies can be found online; "RNS ® System LTT Study -Study Results -ClinicalTrials. gov,", n.d.).Of the device-related malfunctions of chronic ECoG implants, implant site infection was the most predominant adverse event (9.4%, n = 24 out of 256 patients) (Bergey et al., 2015).Nevertheless, all of these were superficial soft-tissue infections, and no infection in the brain or subdural space occurred and no sepsis or longlasting neurological deficits were reported.In total, 4.7% (n = 12) patients suffered from intracranial haemorrhages, of which four occurred in the first days after the implant and five due to seizure-related head trauma.The remaining three patients suffered haemorrhages several years (2.5-3 years) after the implant, which led to device explant for one of them.In a follow-up report after 9 years of implantation, the only serious complications related to the device that occurred in more than 5% of the patient population were 'implantation site infection' and 'elective explantation of the neuro-stimulator, leads or both' a These studies are not part of our database, as they do not discuss the ECoG signal.They are included here as high-powered studies that illustrate the complications most often occurring with ECoG.
b Two notes on Bergey et al. ( 2015): (1) The authors only report serious complications that occur in ≥2.5% of the participant population.
(2) For more information on serious adverse events in this patient population, please see Wei et al. (2016), for a case report on implant site infection and bone flap osteomyelitis, Weber et al. ( 2017), for an investigation in infection and erosion rates, and Nair et al. (2020), for a comprehensive follow-up study on the efficacy and safety of the NeuroPace ® device. c Minor complications in the NeuroPace ® studies are described online for only those partaking in the long-term treatment clinical trial (n = 230).
d Goldring & Gregorie (1984) discusses both new results and those already described in Goldring (1978).For clarity, we keep these groups separate in this table.
e Paediatric patient population.
f Ninety-eight per cent of patients were implanted with both subdural grids and depth electrodes. g The relatively large number of complications reported by Levy et al. (2016) should be cautiously interpreted, as the authors monitored complications up to 24 weeks post-rehabilitation, and patients were explanted at Week 8. Indeed, a death was reported during follow-up, which was unrelated to the study.
T   Nair et al., 2020).Similar to the earlier reports, all but one of the infections involved soft tissue (most often related to skin flora) and no meningitis or brain parenchymal infection occurred (Weber et al., 2017;Wei et al., 2016).Notably, Bergey et al. (2015) state that the number of deaths during the study is not higher than normally reported for patients with medically intractable partial onset seizures.Of the minor complications that were described in the clinical results reported online (occurring in 228 of 230 participants), the most common ones included 'therapeutic agent toxicity' (n = 115), 'adverse drug reactions' (n = 71) and 'implant site pain' (n = 69), which apart from the latter are likely unrelated to the ECoG electrode implant.NeuroVista was a subdural warning system for epilepsy, which reported on only a limited number of participants (n = 15) due to early conclusion of the study.The authors documented a considerable number of complications reported within the first year after implantation (13 in total), of which four were serious (Cook et al., 2013).Most complications seemed to be either related to the surgical implant procedure or device failure and not due to a reaction or deficit induced by the presence of the ECoG grid.
Overall, chronic implantation of ECoG electrodes does not appear to cause a substantial number of serious complications (see Table 2 for overview of the studies included).Infection occurs both in the epidural and subdural studies, although none of the studies reported infections that spread to the central nervous system.Most serious complications in epidural studies appeared to be related to cortical stimulation, but the total number of reports on epidural ECoG is too low to draw definitive conclusions.

| DISCUSSION
The use of ECoG for the design of fully implanted BCIs has become a tangible reality in the last decades, with several studies showing its benefits in terms of stable, long-term recordings.When opting for ECoG-based BCIs, one must choose between subdural and epidural placement of the electrodes.In this review, we attempted to collate available evidence on epidural and subdural ECoG recordings with the goal of establishing the most optimal option for long-term BCI use in terms of signal quality and complications.We systematically evaluated all epidural studies that discussed signal quality and compared those with the most recent analogous subdural studies.In addition, we made an overview of studies reporting on complications of acute and chronic epidural and subdural recordings.Even though the total number of studies describing epidural recordings is still limited, 28 (25%) 9 (8%) Wiggins et al. (1999) 5 (13%) 1 (3%) 2 (5%) Wyler et al. (1984) 2 (7%)

Not assessed
Note: Although all serious adverse events are reported online, nonserious adverse events only are reported if they occurred in more than 2.5% of the patient population.
Abbreviations d Goldring & Gregorie (1984) discusses both new results and those already described in Goldring (1978).For clarity, we keep these groups separate in this table.
e Paediatric patient population.
f Ninety-eight per cent of patients were implanted with both subdural grids and depth electrodes. g The relatively large number of complications reported by Levy et al. (2016) should be cautiously interpreted, as the authors monitored complications up to 24 weeks post-rehabilitation, and patients were explanted at Week 8. Indeed, a death was reported during follow-up, which was unrelated to the study.
we were able to find some key effects of the dura on ECoG recordings (for a summary, see Table 3).

| Epidural and subdural are both viable options for BCIs
In general, available evidence suggests that epidural ECoG signals can be distinguished at a $3-mm scale and that epidural decoding can reach similar, albeit slightly inferior, performance as its subdural counterpart.Large differences in signal amplitude were found between epidural and subdural recordings, however, which seem particularly relevant for recordings with high-density grids of electrodes, likely influenced by the smaller electrode diameter (normally accompanied with smaller IED).Indeed, it has been shown that a smaller electrode radius comes with a decrease in signal-to-noise ratio (Wodlinger et al., 2011) and consequently a higher susceptibility to dural attenuation (Bundy et al., 2014).Yet, this difference does not seem to have substantial effects on actual BCI decoding performance, at least with the current generations and specifications of clinical and high-density grids.
The long-standing belief that epidural recording leads to fewer complications has so far not been substantiated by the literature, partly because of the limited number of epidural studies available.Yet, complications following chronic subdural implantations do not seem to be of a different nature than those reported for chronic epidural implantations.For a proper comparison between both techniques, however, more evidence on chronic epidural implants is needed.As shown above, there is limited evidence for the occurrence of tissue encapsulation of both epidural and subdural ECoG grids in NHPs.Histological studies are necessary to evaluate the occurrence of encapsulation in humans and its effects on signal quality.An interesting remark made by Schendel et al. (2013) while studying rats is that the presence of newly formed tissue may in fact play a crucial role in minimizing micromotions that could potentially result in cortical trauma and signal instability and that this encapsulation could therefore be potentially beneficial in the case of surface electrode arrays (Murphy et al., 2004).To what extent this applies to humans is yet to be demonstrated.

| Consideration to choose epidural or subdural BCIs
Based on the above, epidural and subdural placement both seem to be suitable approaches for long-term fully implanted BCIs, with successful examples already been reported (Aarnoutse et al., 2019;Benabid et al., 2019;Moly et al., 2022;Pels et al., 2019;Śliwowski et al., 2022).In deciding where to place the electrodes for future studies, one may want to take into account the effects of the dura on signal amplitude, specifically in situations targeting high-frequency band features and in cases of limited amplifier abilities (i.e., high noise floor).Another possible consideration to choose the ECoG placement is the fact that epidural recordings may be less suitable for the BCIs that provide neurofeedback by stimulating cortical tissue (e.g., Gharabaghi, Naros, Walter, Roth, et al., 2014;Walter et al., 2013Walter et al., , 2012)), as electrical stimulation of the dura has been suggested to cause discomfort or pain (Caldwell et al., 2019;Wirth & Van Buren, 1971).Interestingly, however, the studies regarding a participant who received epidural electrical stimulation did not report pain related to such stimulation (Gharabaghi, Naros, Walter, Roth, et al., 2014;Walter et al., 2013Walter et al., , 2012)).If this can be further confirmed, epidural-based BCI could be a good candidate for short-term implants for stroke rehabilitation (Soekadar et al., 2015), although the therapeutic effect of this approach is yet to be demonstrated (Levy et al., 2016).Another point of concern only briefly reported is the epidural susceptibility to artefacts due to chewing (Oxley et al., 2016;Sauter-Starace et al., 2019;Shimoda et al., 2012), which have been suggested to cause poor epidural performance.
Lastly, there are also some additional arguments that can be of potential importance for the choice of ECoG placement, such as the less accurate localization of the exact epidural electrode position without neuronavigation (Benbadis, 2000;Wirth & Van Buren, 1971;Zumsteg & Wieser, 2000) the fact that aging and cortical atrophy may increase the distance between the cortex and the epidural space (e.g., McCluney et al., 1992;Yang et al., 2012) and that epidural surgery may be complicated by dura adherence to the skull in older people, making accurate electrode placement more difficult.However, to which extent these factors affect signal quality and BCI performance is still unknown and deserves dedicated attention in future studies.

| Emerging ECoG technologies
ECoG technology has been increasingly boosted in the last decade, with advancements not only in the biomaterials for clinical use but also in the manufacturing aspects of the electrodes, such as grid size, density and signal conduction.In fact, significant effort has gone to the development of ultra-high-density surface microelectrode arrays (Chang, 2015;Ha et al., 2016;Khodagholy et al., 2015;Paulk et al., 2021).Additionally, given the potential of ECoG for neural interfaces, the scientific community has recently invested in the development of ECoG grids that can be combined with imaging and stimulation techniques, such as functional magnetic resonance imaging (fMRI) (Fallegger et al., 2021) or optogenetics (Brodnick et al., 2019;Griggs et al., 2021).Moreover, the use of ever so small grids and the combination of ECoG with stimulation techniques such as optogenetics may require the choice for subdural placement for more precise localization and direct contact with brain tissue.

| ECoG and beyond
This review presents a body of evidence towards the stable and long-term use of (epidural and subdural) ECoG recordings for epilepsy monitoring and for BCI, with examples where the added benefit of the output application may outweigh the implantation risks (Rouanne et al., 2021;Vansteensel et al., 2016).Still, BCI solutions, and in particular implantable BCI systems, are not only dependent on ECoG technology.In the last years, the BCI field has seen a boost of development of new alternative technology that can read and/or write into the brain (Choi, Lee, Park, Cho, et al., 2018;Rapeaux & Constandinou, 2021), including Neuralink ® arrays (Musk & Neuralink, 2019), neural dust (Seo et al., 2016), neuralpixels (Paulk et al., 2022), wireless micro-electrode arrays (Simeral et al., 2021), stereoelectroencephalography (Herff et al., 2020), neurotrophic electrodes (Gearing & Kennedy, 2020), endovascular electrodes (Stentrode ® ; John et al., 2019) and the emergence of new nanomaterials for electrode design like graphene (Bramini et al., 2018).Each approach comes with a specific target user group, a target application, one or more target cortical areas and a minimum target performance.Whereas a communication-BCI for individuals with severe paralysis may benefit more from one system, a BCI to control exoskeletons may command an entirely different one.The future of BCI relies heavily on the continuous development of these techniques and the open sharing of results.

| Review limitations and future directions
In this review, we only included scientific output where ECoG signal properties were in any way described and which addressed humans or NHP.As such, our research did not include the complete body of search outputs mentioning epidural and subdural recordings, such as those describing pre-clinical tests in rats, sheep and dogs.Yet, some of the excluded articles were included in the discussion to put the results found in humans and NHP into perspective.Although humans' and NHPs' brains can have similar features (Herculano-Houzel, 2009;Jerison, 1973), caution should be taken when directly comparing results, especially because the thickness and nature of the dura can vary considerably (Galashan et al., 2011;Kinaci et al., 2020).
Studies assessing the histological effect of epidural and subdural electrodes in humans after long-term implants are scarce.Additionally, complications or adverse events related to ECoG in humans and NHP are inconsistently reported and scarce in NHP studies.More elaborate and standardized reporting would benefit a better understanding of the long-term implications of epidural and subdural implants, in particular the implantrelated complications in ECoG studies.The case could likewise be made for the description of the outcome parameters of the decoding studies.The heterogeneity of the outcomes hinders the accurate comparison of decoding performances.
The assessment of the influence of IED and electrode diameter on signal quality is limited by the IED and diameter typically used for clinical and research practices so far, which have a very confined relationship: the smaller the IED, the smaller the diameter.Studies that explore separately the IED and electrode diameter may also be relevant to decouple the actual influence of this factors on the signal quality.For example, a recent study has predicted from fMRI that densely packed, large-size electrodes could potentially provide optimal decoding of hand gestures (Van den Boom et al., 2021).However, such configuration is not typically tested in clinical or research settings.
Given the target user group of implanted BCIs, the choice of electrode placement should also depend on individual traits such as aging and cortical atrophy.
T A B L E 3 Key findings.
• Electrocorticography (ECoG) signals recorded epidurally can be discerned at $3-mm precision, which is comparable with subdural recordings.• The dura especially attenuates neural signals over a broad range of frequencies for recordings from high-density ECoG grids with small electrode diameters.• Epidural ECoG offline decoding can perform at similar or slightly inferior levels as decoding based on subdural ECoG, and epidural ECoG shows promise for real-time accurate control of complex prosthetics and exoskeletons.• The nature of the serious complications associated with subdural ECoG is comparable with that of epidural ECoG, but the number of studies on epidural complications is too low for a detailed comparison.
Future studies should also consider looking at the effect of these two variables on the epidural and subdural signal quality and longevity.Lastly, an emerging direction of study for implantable BCI technology has been the restoration speech for people with severe paralysis (Rabbani et al., 2019).Intriguingly, to date, the attempts to decode offline (Anumanchipalli et al., 2019;Ramsey et al., 2018;Salari et al., 2018aSalari et al., , 2018bSalari et al., , 2019) ) and real-time (Moses et al., 2019(Moses et al., , 2021) ) forms or speech have only employed subdural (and mostly high-density) ECoG.Given the evidence presented above, one could speculate that speech can potentially be decoded with epidural ECoG as well.This, however, still needs to be demonstrated.

| CONCLUSION
In this review, we systematically compared epidural and subdural recordings and their potential relevance for the design of implantable ECoG-based neural interfaces.The current body of literature shows that epidural and subdural ECoG both allow for high-fidelity recording.
Although signal amplitude may be lower in epidural recordings, electrode placement relative to the dura does not seem to be a significant factor for decoding performance, at least for the current generations of clinical and high-density grids.Moreover, show that, despite the limited reports on epidural complications, the nature of serious complications observed in epidural studies is comparable with that of subdural studies.Future studies should focus on a systematic reporting of complications and direct comparison between epidural and subdural long-term implants in large cohorts.

AUTHOR CONTRIBUTIONS
SG performed the literature search and wrote the first draft of the manuscript.MPB and SG wrote the manuscript.MPB supervised the literature search and produced the figures.MJV, EJA and NFR reviewed and edited the manuscript.

F
I G U R E 1 Definition of size and placement of electrocorticography (ECoG) recordings.(a) Schematic representation of subdural and epidural ECoG recordings.The cerebrospinal fluid (CSF) circulates between the dura and the brain surface.(b) Schematic representation of ECoG grids of N Â M electrodes with [N, M] > 1 (both classic clinical ECoG grids with an interelectrode distance of 10 mm and an exposed surface diameter of 2.3 mm and high-density ECoG grids with ≤4-mm inter-electrode distance and <1.3-mm diameter) and strips of 1 Â N electrodes.

(
: aDBS, adaptive deep brain stimulation; BCI, brain-computer interface; CS, cortical stimulation; CLIS, completely locked in state; DBS, deep brain stimulation; ECoG, electrocorticography. a These studies are not part of our database, as they do not discuss the ECoG signal.They are included here as high-powered studies that illustrate the complications most often occurring with ECoG.b Two notes on Bergey et al. (2015): (1) The authors only report serious complications that occur in ≥2.5% of the participant population.(2) For more information on serious adverse events in this patient population, please see Wei et al. (2016), for a case report on implant site infection and bone flap osteomyelitis, Weber et al. (2017), for an investigation in infection and erosion rates, and Nair et al. (2020), for a comprehensive follow-up study on the efficacy and safety of the NeuroPace ® device.c Minor complications in the NeuroPace ® studies are described online for only those partaking in the long-term treatment clinical trial (n = 230).
Selection of studies that report complications in acute and chronic epidural and subdural ECoG recordings in humans.Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/ejn.15941by Utrecht University Library, Wiley Online Library on [11/05/2023].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License Although all serious adverse events are reported online, nonserious adverse events only are reported if they occurred in more than 2.5% of the patient population.Abbreviations: aDBS, adaptive deep brain stimulation; BCI, brain-computer interface; CS, cortical stimulation; CLIS, completely locked in state; DBS, deep brain stimulation; ECoG, electrocorticography.
T A B L E 2Note: Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/ejn.15941by Utrecht University Library, Wiley Online Library on [11/05/2023].See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions)on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License