A Review of Motor Brain-Computer Interfaces Using Intracranial Electroencephalography Based on Surface Electrodes and Depth Electrodes

Brain-computer interfaces (BCIs) provide a communication interface between the brain and external devices and have the potential to restore communication and control in patients with neurological injury or disease. For the invasive BCIs, most studies recruited participants from hospitals requiring invasive device implantation. Three widely used clinical invasive devices that have the potential for BCIs applications include surface electrodes used in electrocorticography (ECoG) and depth electrodes used in Stereo-electroencephalography (SEEG) and deep brain stimulation (DBS). This review focused on BCIs research using surface (ECoG) and depth electrodes (including SEEG, and DBS electrodes) for movement decoding on human subjects. Unlike previous reviews, the findings presented here are from the perspective of the decoding target or task. In detail, five tasks will be considered, consisting of the kinematic decoding, kinetic decoding,identification of body parts, dexterous hand decoding, and motion intention decoding. The typical studies are surveyed and analyzed. The reviewed literature demonstrated a distributed motor-related network that spanned multiple brain regions. Comparison between surface and depth studies demonstrated that richer information can be obtained using surface electrodes. With regard to the decoding algorithms, deep learning exhibited superior performance using raw signals than traditional machine learning algorithms. Despite the promising achievement made by the open-loop BCIs, closed-loop BCIs with sensory feedback are still in their early stage, and the chronic implantation of both ECoG surface and depth electrodes has not been thoroughly evaluated.

Among these invasive devices, MEA-based BCIs were proved to be highly effective in motor decoding using animal models [17], [18], [19], [20].This effectiveness later was validated by several studies of movement [21], [22], [23], [24], [25] and speech [26] on human subjects.However, this line of research is limited by suitable participants because patients' willingness to accept invasive implantation with limited benefits is low.As such, the MEA-based BCIs will not be covered in this work.
On the other hand, the large population of patients with neurological diseases requiring invasive surface and depth electrode implantation provides unique opportunities for invasive BCIs.About 1% of the world's population has epilepsy, and approximately one-third of these patients have medicationresistant epilepsy [27].ECoG and SEEG are frequently used for seizure focus localization in patients receiving surgical interventions, and this provides opportunities for BCI Fig. 1.Common recording devices used in BCI research.EEG is recorded non-invasively using electrodes placed at the surface of the scalp.ECoG uses electrodes placed on the surface of the cortex.MEA is an array of tiny penetrating needles that record signals from 1-2 mm beneath the cortex.SEEG and DBS depth electrodes are placed along a penetrating depth shaft, both of which will be referred to as depth electrodes in this paper.
studies [28], [29].ECoG has drawn significant attention since its first proof-of-concept study [30].Though not as popular as ECoG, the application of SEEG in BCIs has been growing in recent years [13], [31], [32], [33], [34].Besides epilepsy patients, over 160,000 patients worldwide have undergone DBS for a variety of neurological and non-neurological conditions, with numbers increasing each year [35], [36].Similar to the SEEG electrodes, the DBS electrode is another possible way to record intracranial signals for BCI research.Though not intensively studied as other methods, their usage in motor BCI dates back to 2004 [14].Because of the similarity between electrodes used in SEEG and DBS and that both record signals from deep brain regions, they will be referred to as depth electrodes in this paper unless otherwise stated.
It is worth noting that, although the current BCIs research has widely used iEEG signals collected from epilepsy patients and Parkinson's patients, most of them are for research purposes and do not aim to treat or cure their motor dysfunctions and neurological disorders.

A. Scope of This Review
This review focuses on motor BCIs studies using ECoG, SEEG, and DBS electrodes, conducted on human participants.Although there are existing reviews on motor BCIs based on both ECoG [37], [38] and SEEG [28], the rapid development of invasive BCIs in recent years necessitates an update.In addition, previous reviews approached the problem from the perspective of signal or decoding methods.For example, a high-level review of invasive BCI based on ECoG was provided by Schalk, etc., ranging from the device, signal properties in the temporal and spectral domain, controlling strategy, and existing papers presented in chronological and task complexity order [37].Another review focused on the decoding strategy and presented BCIs by looking at the decoding approaches which can be grouped into arbitrary mapping, classification, and regression [38].Herff etc. introduced the SEEG-based BCIs in different domains including motor, visual speller, speech, navigation, and passive BCIs [28].From the aforementioned review papers, we learned that movement can be decoded using various algorithms (such as the advanced deep learning methods), but to what extent this information can be decoded and has already been decoded currently is not known.For example, the decoding of the kinematic information could be simply the classification of status (different gesture), or it could be of more complexity as decoding the trajectories in 2 dimensions or even 3 dimensions.To know exactly what the invasive BCIs are capable of currently, we looked at five aspects related to movement, including the decoding of kinematic, kinetic, identification of body parts, dexterous hand decoding, and motion intention decoding.Studies related to these five aspects are presented gradually becoming more difficult to reveal the current capabilities of the invasive BCIs (what they can do).
In these five aspects, kinematic decoding involves information about position, speed, or velocity, while kinetic parameter means the force magnitude; both may be further divided into discrete or continuous parameters.In addition, hand decoding is covered in a dedicated section because it is a critical requirement in daily life and many key studies have been conducted on this topic.
In this review, publications about motor BCIs based on ECoG, SEEG, and DBS were pooled from the Web of Science, IEEE Xplore, and Google Scholar between 1997 and December 2022.The following keywords were used to search the databases: brain-computer interface (BCI), brain-machine interface (BMI), electrocorticography (ECoG), stereo-electroencephalography, stereotactic electroencephalography (SEEG), deep brain stimulation (DBS).These search criteria retrieved 1021 papers.Then, papers will be excluded if they were not human experiments, used devices other than ECoG, SEEG, or DBS, or decoded non-motor information.In the end, a subset of 180 papers was kept in this review.The pooled studies were grouped by recording device and sorted by year, as presented in Fig. 2.
In the next section, detailed information about surface and depth electrodes will be presented along with their application in the clinical setting.

B. Electrocorticography (ECoG)
ECoG was pioneered in the early 1950s by Wilder Penfield and Herbert Jasper at the Montreal Neurological Institute [39].It has since been recognized as a popular tool in a variety of animal studies [40], [41], [42], [43], but has found limited application in human participants due to safety issues associated with intracranial surgery.Thus far, the majority of participants were epilepsy patients requiring invasive monitoring for seizure focus localization before surgical resection; approximately one-third of epileptic patients have medication-resistant epilepsy that responds well to surgical intervention [27].To locate seizure focus, scalp EEG, along with other imaging modalities (such as fMRI or PET), is performed to assess the laterality and possible origin of the seizure prior to surgery.ECoG is then used to identify the extent of the epileptic tissue, whilst SEEG is used if the laterality is unknown or if the focus of the seizure was identified in the deep area of the brain [44].In addition, ECoG is used in other situations, such as functional mapping of the eloquent cortex using electrical cortical stimulation intraoperatively during awake craniotomy for brain tumour resection surgeries [45], [46].During the implantation period, BCI studies can be conducted in a limited time window (∼ 1-2 weeks on epilepsy patients).The first ECoG-based BCI study was conducted in 2004 [30], and since then, many BCI studies have been undertaken, ranging from motor control, and force prediction to speech decoding.

C. Depth Electrodes
SEEG was introduced as a diagnostic tool for patients with epilepsy by the S. Anne Hospital, Paris, France, in the second half of the 20th century [47].SEEG is often used for epilepsy monitoring when the suspected seizure onset zone is located in deep brain structures, such as the insula or hippocampus [44].SEEG devices usually contain multiple recording contacts (typically 8-16 with a center-to-center distance of 3.5 mm) placed along a cylindrical shaft that is inserted via a burr hole in the skull.Signals recorded using SEEG have high amplitudes and broad bandwidths [48].In contrast to ECoG, which has a higher density coverage of the cortical surface, SEEG covers sparser and bilateral regions of the deep brain regions [44].It can record signals from deep subcortical regions, such as the medial temporal, orbitofrontal, cingulate, and insular regions.One of the earliest demonstrations of SEEG for BCIs emerged in 2011 in a work by Krusienski et. al., who demonstrated that a P300 speller can be achieved with nearly 100% accuracy [49].
Deep brain stimulation (DBS) is another widely used method in clinical neurosciences in the past two decades, which can directly measure pathological brain activity and can deliver adjustable stimulation for therapeutic effect.It has been found effective in conditions such as Parkinson's disease, depression, and Alzheimer's disease (AD).In DBS, the conventional lead is composed of 4 electrode ring contacts that are 1.5 mm in length arranged on a diameter of 1.27-1.36mm cylindrical shaft [50].Its usage as BCI dates back to 2004 by Loukas and Brown [14], in which the voluntary hand-movement onset can be detected at 95% accuracy using signals recorded from the subthalamic nucleus (STN) of patients with Parkinson's disease (PD).
The electrodes used in DBS are similar to those employed in the SEEG, both of which will be referred to as the depth electrode in the following contents.
Overall, SEEG-or DBS-based BCIs have received less attention than those based on ECoG, but the number of BCI using depth electrodes in the literature has been steadily increasing due to a recent trend of moving from subdural grid ECoG recordings toward SEEG deep electrodes for intracranial localization of seizures [51].Such shift is caused by several reasons, such as: • Safety concern.The SEEG electrode has a length of roughly 2 mm, less than 1mm in diameter, arranged on a long cylindrical shaft, with an inter-electrode distance of roughly 1.5 to 3.5 mm [52].The ECoG electrodes are circular plates with a diameter of roughly 2mm arranged in a grid or strip layout.Because of the flexible planar arrangement, craniotomy is required for the ECoG implantation and therefore SEEG is preferred as a minimally invasive approach in surgical fields.
• Surgery procedure.At present, the planning of the SEEG electrodes can be facilitated by 3-dimensional (3D) highresolution magnetic resonance imaging, and electrode placement can be aided using 3D-printed omnidirectional platforms, robotic assistance, or frameless stereotaxy.
• Treatment outcome.There is a lower rate of resection after SEEG but a decreased risk of complication and a higher rate of seizure freedom compared to ECoG [53].This could be because this minimally invasive method has lowered the bar for epilepsy practitioners to study challenging patients who require bilateral implantation.This trend also can be seen in Fig. 2, which shows the number of publications about motor BCI using ECoG or depth electrodes.In the figure, a clear decreasing trend can be seen for ECoG starting from 2012, while the number of SEEG studies increased compared to that before 2012.Although the reviewed publications only comprised a small part of epileptic studies using the intracranial method, because not all epileptic studies investigated BCI, such shift was indirectly supported by Fig. 2.
The pooled publications were grouped into five categories, including the kinematic decoding, kinetic decoding, identification of body parts, dexterous hand decoding, and motion intention decoding, as presented in Fig. 3, in which the respective number of publications related to ECoG and depth electrodes are also presented as the outer pie chart.In this pie chart, it is obvious that most intracranial BCIs study kinematic information, and dexterous hand and kinetic decoding also draw a lot of attention, while the investigation of the identification of body parts and moving intention are less studied.On the other hand, regarding the employed devices, it is clear that only very few studies related to kinematic, kinetic, and hand decoding use depth electrodes, while no Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.Fig. 3. Publications are grouped into five categories.The inner pie chart consists of five categories including kinematic decoding, kinetic decoding, identification of body parts, motion intention detection, and hand decoding, respectively.Note studies that decode kinematic and kinetic information for the hand will be categorized as hand decoding.Each category was further partitioned into studies based on the surface and depth electrodes, shown in the outer pie.
study exists using depth electrodes for body parts identification and moving intention decoding.Overall, studies using depth electrodes consist of 20% of reviewed studies.
In the next section, some important studies will be presented to outline the progress made by these three electrodes.

II. STATE OF THE ART
A list of representative studies using both depth and ECoG electrodes can be found in table I and table II.In the tables, we highlighted several important aspects, such as recording devices, device setup, location, implant regions, experiment paradigm, feature extraction, decoding algorithm, decoding target, and decoding result.The category of each study was indicated in column 'Decoding target', in which kinematic, kinetic, identification of body parts, dexterous hand decoding, and motion intention decoding was represented as A1, A2, A3, A4, and A5, respectively.In the content below, detailed interpretations of the listed studies were provided, from three aspects: various motor tasks, signal processing method, and location of the electrodes.

A. Motor Task
In this section, studies related to five motor decoding tasks will be presented, including the kinematic decoding, kinetic decoding, identification of body parts, dexterous hand decoding, and motion intention decoding.Within each of these five categories, studies were further divided by numerical type of target parameter (discrete or continuous) and recording devices (ECoG surface electrodes or depth electrodes).

B. A.1 Kinematic Decoding
The ability to grasp and manipulate objects requires the simultaneous decoding of both the kinematics and kinetic parameters.As shown in Fig. 4, kinematic decoding studies movement parameters, such as position, velocity, and acceleration, while kinetic decoding studies the force applied.
Kinematic decoding can be further categorized as either discrete (classification) or continuous (regression).The number of papers related to the ECoG surface and the depth electrodes, categorized into continuous and discrete groups, is presented in Fig. 5.In this pie chart, it is shown that most studies decoded discrete parameters, while continuous decoding was less investigated.On the other hand, for the employed devices, the depth electrodes were rarely used, especially in decoding continuous parameters.Overall, depth electrodes consisted of 13% of kinematic studies.

C. A.1.1 Discrete Kinematic Decoding Using ECoG
Previous studies have demonstrated the possibility of discrete 1D and 2D movement control using ECoG and depth electrodes.The representative studies will be presented below.
The first ECoG-based online BCI study was conducted on four patients with intractable epilepsy [30].Two experiments were employed in that study, an online 1D cursor control task and an offline joystick centre-out task.In the cursor control task, a two-stage procedure was used to achieve the online vertical position controlling of a computer cursor.The first stage was feature selection during which participants were asked to either open and close the right or left hand, protrude the tongue, say the word 'move', or imagine performing each of these three actions during which ECoG signals were collected simultaneously.The frequency bands that showed the highest correlations with any task described above were then identified as informative features and will be used for subsequent control.During the next online control stage, participants tried to control the vertical position of a computer cursor with online visual feedback by modulating the ECoG features identified in the previous stage.The cursor will move every 40 ms (based on the data of the previous 280 ms) either up with task execution or down with rest.They demonstrated that a cursor can be controlled vertically with an accuracy of 74-100% in a brief training period of 3-24 min.In an additional joystick center-out task, they showed directions predicted by a neural network decoder were highly correlated with the actual movement directions.From these two paradigms, it is suggested that ECoG signals recorded from selected electrodes can be used for rapid and accurate cursor control.
Another 1D cursor control study evaluated decoding with or without sensory information [64].In that study, a cursor was controlled to move to the right or left whilst the participants were actively moving a joystick in the corresponding direction.Besides, they showed that the same cursor controlling can be obtained using only a small subset of electrodes (only two neighboring electrodes for one participant).Next, they tried to eliminate signals from somatosensory input by sulci reconstruction to identify and discard electrodes from the sensory cortex, and they found the decoding accuracy dropped slightly from 81% to 76%.This suggests that kinematics decoding is possible, even without somatosensory feedback.Beyond 1D cursor control, another study using ECoG demonstrated cursor control in 2 dimensions [57].In their study, the participants were asked to move a joystick in one of four directions (up, down, left, or right) and a fifth trigger condition.With a Bayes classifier, not only the moving statue (movement vs idle) can be distinguished at high accuracy (83% to 96%), but four moving directions can also be well identified (58% to 86%) using only the high gamma component.
Another study proposed a different strategy for movement control by modulating frequency power in 7 ∼ 13 Hz as a switch to choose the desired direction [93].In their design, the participant would activate the switch only if they were not satisfied with the direction of the moving cursor at any moment.If the switch was not activated, the cursor would continue to move in the current direction.Each time the switch was activated, the cursor changed direction.They showed that the participant was able to obtain this brain-switch control after 15 minutes and was able to reach all targets.However, they only managed to achieve an information transfer rate of 6.3 bits/minute.
It is worth noting that there are other studies that decode discrete motor-related information but are not used for the restoration of movement [94], [95], [96].For example, a fully implanted ECoG device was used to control a computer typing program by attempting to move the hand contralateral to the implanted electrodes [97].

D. A.1.2 Discrete Kinematic Decoding Using Depth Electrodes
Signals from deep brain regions have also been shown to be modulated by both motor imagery and motor execution.For example, in Parkinson's disease (PD) patients, it has been shown that both movement-related frequency desynchronization (ERD) and synchronization (ERS) were found in the subthalamic nucleus (STN) during intermittent or continuous voluntary movements [98].Also, similar to the motor cortex, the deep structures exhibit oscillatory activity changes during a self-initiated movement [99].Direct evidence comes from a BCI study using DBS electrodes, in which Loukas et al. evaluated the possibility of decoding hand movement in participants with PD [14].With spectral features extracted via wavelet transformation, they were able to predict the onset of self-paced hand movement with 95% sensitivity and 77% specificity.Further, most predictions were made a second before the actual movement.In another DBS study, participants were visually cued to press a key with either their right or left index finger [85].With optimal feature selected using weighted sequential feature selection (WSFS) strategy, movement detection accuracy was obtained at 99.6 ± 0.2% and 99.8 ± 0.2%, and subsequent laterality (left or right) classification reached 77.9 ± 2.7% and 82.7 ± 2.8% using Bayesian and support vector machine (SVM) classifiers, respectively.
Breault et al. demonstrated that hand movement speed can also be decoded from the SEEG signals [90].In their study, they used signals from various deep brain regions across multiple frequency bands to predict speed in a center-out task.A linear model was trained while the features were selected by applying the Least Absolute Shrinkage and Selection Operator (LASSO) on the training dataset.Using the selected feature and the trained decoder, the decoding performance on the testing dataset had an average correlation coefficient of 0.38 ± 0.03 and an average mean squared error (MSE) of 1.07 ± 0.09 across all participants.Furthermore, the decoded speed matched the categorical representation of the test trials (correct or incorrect) with a mean accuracy of 70 ± 2.75% among participants.

E. A.1.3 Continuous Kinematic Decoding Using ECoG
To perform complex activities, such as grasping a cup, the decoding of movement trajectories or speed as a continuous variable is a critical requirement.Various studies have been conducted on continuous decoding of position and velocity in 2D and 3D space.Detailed works will be presented next.
In one example, Schalk.et al. evaluated the possibility of continuous kinematic decoding using ECoG signals in a cursor control task [11].In their study, 5 participants were asked to use a joystick to move a cursor to track a counterclockwise moving target on a computer screen.They observed a new component of the brain signal, the local motor potential (LMP), which exhibited a high correlation with the position and contained substantial information about the direction of movement.Then the frequency-based features and the LMP, which was calculated as the 333 ms running average of the raw signal, were used as input features for linear decoders.A linear model was derived for each of the four kinematic parameters (i.e., horizontal and vertical position, horizontal and vertical velocity).They achieved high correlation coefficient (0.49 to 0.81 for 5 subejcts) between the predicted and the actual trajectory which was within the range of those achieved using microelectrodes.They then attempted to understand the importance of different signals and different brain locations for decoding; they observed a wide range of brain areas that show substantial contributions to decoding, including not only motor and premotor cortical areas but also additional areas such as the dorsolateral prefrontal cortex and those that do not have obvious motor control relevance.In addition, they demonstrated that the amplitudes of different frequencies and LMP were cosine tuned, similar to the tuning behavior of the fire rate of neurons recorded with MEAs [100].
Pistohl et al. employed a similar target tracking paradigm wherein the target position was defined randomly [54].In this study, six participants were asked to perform continuous 2D arm movements to track a random target.A Kalman filter was used to predict position and velocity in both horizontal and vertical directions.Despite the stochastic movement, they obtained only slightly decreased decoding accuracy, measured as the correlation coefficient, compared with [11].
In most studies, the power of the ECoG signals within specific frequency ranges was computed in short-duration bins and instantaneously and linearly decoded.To investigate the optimal feature extraction and decoding strategy, Gunduz et al. designed a center-out and target selection task to study the effect of spectral resolution, time embedding, and decoder complexity on the decoding performance [101].To determine how finely the ECoG spectrum should be divided, they partitioned the ECoG spectral frequency bands into 8, 16, and 32 sub-bands.To test the decoder complexity, they used a Wiener filter as the linear decoder, while a neural network architecture called echo state networks (ESNs) was used as a nonlinear decoder.By showing that there were no significant differences in performance using different sub-band numbers, they concluded that the lowest resolution level at which the well-known neurophysiological rhythms are maintained (n = 8) is the optimal resolution.For the decoder, they showed that nonlinear models performed better and were robust in the presence of random interictal activity.They further proved that the high frequency between 300 Hz and 6 kHz, which was usually discarded or not available also contributed to the decoding.
To push this line of work forward, the prediction of the 3D position of the forearm and hand has been studied using ECoG [68], [69], [76], [79], [102].In one study, the position of the shoulder, elbow, and wrist and the joint angle of the elbow were decoded [68].In their study, three participants were asked to clockwise reposition three blocks at the corners of a square space one by one.The average correlation coefficient decoded using a sparse linear regression method achieved a correlation coefficient of 0.44 ∼ 0.73.In another 3D study, 5 epileptic patients were asked to perform the center-out reaching task to 8 targets positioned at the corners of a physical cube [76].Seven sub-bands frequency and LMP were used as features to reconstruct the hand position in 3D using a hierarchical decoding method.In the first stage, a logistic regression model using elastic net regularization was used to determine whether the patient was moving or not.The output from the first step was then used to switch between two PLS models: one to predict kinematic parameters at rest and the other for the moving period.With this 2-stages decoding schema, they achieved mean correlation coefficients between 0.31 and 0.80 for speed, 0.27 and 0.54 for velocity, and 0.22 and 0.57 for position.
Based on the preliminary studies above, a comprehensive study of six elementary upper extremity movements prediction was attempted [69].In this study, Wang et al. tried to determine whether the movement of 6 elementary upper extremities, contralateral to their ECoG implantation, can be decoded from three participants who underwent subdural electrode placement for evaluation of epilepsy surgery.Six elementary movements involved: 1) pincer grasp and release; 2) wrist flexion and extension; 3) forearm pronation and supination, 4) elbow flexion and extension; 5) shoulder forward flexion and extension; 6) shoulder internal and external rotation.A two-stage decoding approach was used.In the first stage, a binary decoder was used to distinguish between idle and moving states.Next, a Kalman filter was constructed to decode movement trajectories during the moving period.The state decoder classified idle and moving states with an accuracy of approximately 91%.The position and velocity of the upper extremities were decoded with average correlations of 0.70 and 0.68, respectively.The high decoding accuracy presented in this study suggested the possibility of a 6-degree of freedom (DOF) upper extremity prosthesis.
Although voluntary motor movements are believed to be controlled by the hemisphere contralateral to the moving limb, there is increasing evidence that the ipsilateral hemisphere also plays an active role.To determine whether 3D arm kinematics (speed, velocity, and position) could be decoded from cortical signals recorded from the ipsilateral hemisphere, Bundy et al. conducted a study wherein four epileptic patients were asked to perform reaching movements using both hands from the centre to the corners of a 50 cm long physical cube while ECoG signals were simultaneously recorded [103].They showed that both contralateral and ipsilateral hand movements can be reconstructed.Furthermore, in the evaluation of the temporal-spectral response of both hand movements, they found that the spatial and spectral encoding were similar for contralateral and ipsilateral representation.In particular, ERS/ERD can be observed over the sensorimotor cortex for both contralateral and ipsilateral reaches, but begin earlier and are greater in amplitude for contralateral reach.
Finally, to evaluate the decoding performance in the real world, a study was conducted on a patient with tetraplegia following a C4-C5 spinal cord injury, to test the possibility of cortical control of a virtual avatar or a suspended exoskeleton with up to eight DOF [82].In their work, a wireless epidural ECoG device was implanted bilaterally on the sensorimotor cortex, which was identified during the patient making real or virtual movements with their upper and lower limbs.Next, a decoder was gradually calibrated.At first, the participant was asked to mentally trigger on-off events of various effectors (switch decoding).Then, the participant made a series of continuous movements of their upper limb segments with gradually increased and up to eight DOF (four degrees of freedom for each upper limb).At the same time, the decoder was calibrated and updated regularly.Finally, the calibrated decoder was fixed for the online prediction.For performance evaluation, they used the ratio R to represent the trajectory decoding performance, which is the ratio of the distance traveled by the effector from the origin to the target over the shortest possible origin-to-target distance.In the final exoskeleton control test, the participant completed five 8D tasks over 7 weeks with an R of 9.8 and a switch decoding accuracy of 70.9%.

F. A.1.4 Continuous Kinematic Decoding Using Depth Electrodes
Continuous kinematic decoding is also possible using depth electrodes.For example, 2D cursor control has been demonstrated by Vadera et al. [86].They recorded SEEG signals from the STN when PD participants were cued to wiggle/rest their hands and feet corresponding to four targets on the screen.Local field potentials (LFPs) were used as features to train a decoder for the subsequent online cursor-control task.During the online 2D target acquisition task, participants have to wiggle/rest their hands to steer the cursor right/left and wiggle/rest their feet to steer the cursor down/up.They showed that the decoded trajectories were in the correct direction with some degree of spread.In most cases, the spread stayed within the correct quadrant of the workspace.

G. A.2 Kinetic Decoding
Alongside kinematic decoding, it is also important to obtain kinetic decoding to dexterously manipulate objects.The number of papers about kinetic coding using ECoG and depth electrodes is presented in Fig. 6.This pie chart demonstrated that most studies investigated decoding of continuous kinetic parameters, while only a few decoded discrete parameters.On the other hand, studies using depth and surface electrodes are relatively similar.Overall, studies using depth electrodes consist of 50% of intracranial kinetic studies.
Both methods have been proven to be possible to decode discrete and continuous kinetic parameters.It looks like most of the studies are continuous decoding, and the possible reason is that a grip dynamometer was used to record the applied force in the kinetic study therefore a continuously changing force (used in the continuous decoding task) can be obtained instead of several discrete statuses (used in the discrete decoding task).
Publications related to kinetic decoding were divided into discrete and continuous parameters in the inner pie, and each type was further categorized into surface and depth electrodes in the outer pie.
isometric force using their right hand in one of 6 orthogonal directions or grasped under 4 pressure conditions (0, 20, 30, 40 kPa).Using temporal-spectral analysis, they found a particular channel in the posterior parietal cortex exhibited approximately a positive linear relationship between mean high gamma power and absolute isometric kinetic magnitude.Further, using a Quadratic Discriminant Analysis (QDA) classifier, force magnitude classification achieved an accuracy of 41%.The decoding of the 6 isometric kinetic directions also achieved above-chance accuracy.

I. A.2.2 Discrete Kinetic Decoding Using Depth Electrodes
Kinetic decoding is also possible using depth electrodes.With SEEG electrodes, Murphy et al. showed that signals from deep cortical areas, including the insular cortex and the central sulcus, can be used to differentiate three different levels of hand grasping and rest state [88].Tan et al. from another study, recorded deep brain signals using the DBS electrodes from nine human participants with PD while performing a grasping task using different force amplitudes [15].They observed that the subthalamic nucleus (STN) behaved differently at different grasping levels.Specifically, they observed that as the force increased, beta suppression deepened, and then plateaued; however, gamma and high-frequency power increased monotonously.They further categorized the force amplitude into multiple discrete levels (1 to 10) and found that power modulation in the beta band was the only independent predictor of force when effort levels < 5, while gamma-band activity was the only independent predictor when effort level ≥ 5.

J. A.2.3 Continuous Kinetic Decoding Using ECoG
As with the continuous decoding of kinematic parameters, continuous decoding of the kinetic parameters is also critical to achieving fine control over external actuators.Flint et al. demonstrated that accurate prediction of isometric pinch force can be achieved by decoding ECoG signals from 10 human participants [71].Signals were recorded contralateral to the moving hand, while participants were asked to squeeze a force sensor between their thumb and litter or index finger.Classical ERS and ERD can be observed from the temporal-spectral response.Then, six sub-band frequencies and LMP were used to train a Wiener cascade filter.The predicted force explained 22% to 88% of the variance in the actual force, with the high gamma frequency band and LMP being the most informative features.Furthermore, they showed that the electrodes at the post-central sites were significantly worse at force decoding than the M1, precentral, or all electrodes combined, which implied that force encoding was predominantly motor rather than sensory.
To obtain a better understanding of the relationship between isometric force and high-gamma response, Branco et al. modeled the temporal dynamics of the high gamma power under three different grasp force tasks: fast impulse-like responses, continuous dynamic force, and isometric force contractions [104].Two different models were used, one captured the relationship between force magnitude and high gamma power, and another captured the relationship between force derivative and high gamma power.Temporal spectral analysis showed that the HFB power did not exhibit a sustained response during a constant force, but rather a transient response during finger flexion (contraction) followed by a 70% smaller response during finger extension (release).They showed that the model based on the force derivative fitted the HFB better than the model based on the force magnitude, regardless of the task and location of the electrode.They concluded that the gamma power was correlated with the force derivative instead of the force itself.This relationship is further confirmed by another ECoG study, in which they showed that the fluctuations of HFB-ERS primarily, and of LFB-ERD to a lesser extent, correlated with the time-course of the first time-derivative of force (yank), rather than force [105].

K. A.2.4 Continuous Kinetic Decoding Using Depth Electrodes
Tan et al. demonstrated that using local field potential (LFP) activities recorded from the subthalamic nucleus (STN) in patients with deep brain stimulation (DBS) electrodes, it is possible to reconstruct the temporal profile of gripping force [89].In another similar DBS study, Fischer et al. recorded signals from the STN of 11 participants who performed actual or imaginary hand grasping for five seconds at different amplitudes [106].They found that both beta and gamma activities changed in accordance with the level of force throughout the grasping period.
A recent work conducted by Wu et al. demonstrated that prolonged grasp force could be decoded continuously from SEEG recordings using an advanced deep learning method [31].In their study, participants were asked to grasp at two increasing rates to two target amplitudes, resulting in four combinations (tasks): slow ascending light grasp; slow ascending hard grasp; fast ascending light grasp, and fast ascending hard grasp.All tasks lasted for 15 seconds.Using a novel deep learning network that consists of a convolutional (CNN) and recurrent neural network (RNN), the decoding mean squared error (MSE) was achieved as low as 0.05 for one of the five participants.

L. A.2.5 Simultaneously Decoding of Kinetic and Kinematic Parameters
Finally, the ability to grasp and manipulate objects requires the decoding of both the kinematics and isometric parameters simultaneously.Previous work suggested that these two behavioral aspects were controlled separately.For example, it is suggested that motor learning of kinematics and kinetics in reaching movement occurs independently [107] and kinematic and kinetic control can be disrupted independently [108].
To better understand the relationship between the kinetic and kinematic parameters, Flint et al. studied ECoG response using a finger-moving paradigm [109].Seven patient participants were asked to repeatedly execute a one-finger flexion movement.The participants moved their index fingers to a force sensor worn on the thumb, and after touching, the isometric force was applied by pressing the index finger to the force sensor.Sub-band frequency features were input to a Wiener cascade decoder to reconstruct the continuous movement kinematics and subsequently applied continuous isometric force.Across participants, the overall median fractional variance accounted for (FVAF) was 0.7 ± 0.2 for force decoding and 0.7 ± 0.3 for movement decoding.Next, the location of the peak decoding performance for movement and force was calculated and projected to the cortex surface.They found that the peak-performance electrode location was different for the two decoding tasks.The mean distance (between movement and force) in all participants of the peak decoding performance is 9.9 ± 2.0 mm.They concluded that spatial representations of movement and force on the cortical surface are different, which supported and extended the findings of early work by Venkadesan and Valero-Cuevas [110], who inferred that the human motor system using two separate control strategies for movement and isometric force control.

M. A.3 Identification of Body Parts
The identification of actual or imagined movement of different body parts, such as the hands, tongue, or feet can be used in the subsequent control task, such as directing a wheelchair and cursor control using an arbitrary mapping strategy.
Body part movement can be well identified using intracranial signals.For example, a study was conducted using ECoG to discriminate fingers, wrists, or elbows when performing flexion and extension movements [84].The obtained off-line decoding accuracy can be achieved at 62% -83% from 4 human participants using a linear model using high gamma frequency.
Although high specificity (ability to discriminate different body parts) was achieved in the above study, a balance has to be obtained when both sensitivity (ability to distinguish between task and idle state) and specificity were considered simultaneously.It has been shown that a single decoder might be inefficient to simultaneously maintain the high performance for both [111].In their study, the subject was asked to make 6 elementary upper extremity movements involving finger, wrist, forearm, elbow, and shoulder [111].A decoder was used to discriminate movement vs idle and the 6 different movements.However, they found that the decoder can achieve a high sensitivity of discriminating movement state, but the specificity is low.One possible solution is the two-step decoding method proposed in an ECoG study [69].In the two-step method, a binary classifier was first employed to differentiate idle and movement states, while a second decoder will be used for the real classification task.
The identification of different body parts can be used for cursor control.For example, four directions of cursor control were achieved by arbitrarily mapping brain responses modulated by actual or imagery movement of different body parts involving the hand, tongue, jaw, shoulder, legs, and finger [3].In the above study, 2D control was obtained in three stages.In the first screening stage, informative ECoG features with the largest task-related amplitude changes were identified.Next, the participants were trained first on horizontal and then vertical cursor control by modulating the feature identified in the previous step.Finally, control was achieved by combining sets of ECoG features that the participant had previously learned to control independently.The average accuracies achieved by five participants ranged from 53%-73%.Participants usually acquired substantial control over identified ECoG features in a short period (12-36 min).In a similar cursor control experiment, Reddy et al. achieved an accuracy of 58% to 86% using only high gamma as the feature.
To test this line of research on locked-in patients, Wang et al. conducted experiments using ECoG signals on a tetraplegia patient caused by C4 level spinal cord injury (SCI) [4].In their study, the participant achieved robust control of a computer cursor in 3D space by imaging the movement of segments of the upper limb including the thumb, elbow, and wrist.A co-adapting strategy was used during the training of the decoder.Specifically, in the stage of decoder adaptation (the decoder learned from the participant), the weight of a linear model was updated by the observed brain signal.In the following stage of human adaptation (the participant learned from the decoder), the decoder weight was fixed, and the human participant learned to modulate the cortical activity by imaging the movement of different segments of the upper limb.To further ease the training process, the decoding process started from 1D control and then gradually moved to 2D and then 3D control.It demonstrated that the participant can obtain 3D control rapidly by blending in control of the third dimension while maintaining control of the first two dimensions.

N. A.4 Dexterous Hand Decoding
Dexterous hand decoding is a particularly important requirement of functional BCIs.To restore fine hand or finger control to paralysed patients or amputees, two requirements must be met: a prosthetic device with a high degree of freedom (DOF) and the ability to recognize rich motor intentions from neural signals.Both ECoG and depth electrode studies demonstrated that it is possible to identify different hand gestures and moving digits.Some ECoG studies also showed high accuracy in the reconstruction of finger trajectories, while the SEEG studies proved that signals from white matter also can be used for decoding.The number of consisting papers, categorized into continuous and discrete groups, is presented in Fig. 7, and more detailed representative works will be described below.In this pie chart, it is shown that most studies decoded discrete parameters, while continuous decoding was less investigated.On the other hand, for the employed devices, the depth electrodes were rarely used, especially in decoding continuous parameters.Overall, depth electrodes consisted of 17% of kinematic studies

O. A.4.1 Hand Decoding Using ECoG
To regain hand movement, a modular prosthetic limb (MPL) from Johns Hopkins University was developed, which permits actuation with 17 degrees of freedom in 26 articulating joints [112].An example study was conducted to evaluate the MPL device on a human participant [113].In their study, the human participants were asked to perform 5-finger flexion while cortical signals were recorded with high-density ECoG.Then the informative electrodes were identified and used by a hierarchical linear discriminant analysis (LDA) decoder to predict: 1) if any finger was moving and, if so, 2) which digit was moving.Using the trained decoder, participants try to use the same prediction framework to achieve immediate control of individual MPL fingers.Movement detection and different finger movement classification achieved accuracies of 92% and 76%, respectively, during online control.
Apart from the hardware, the decoding of rich motor intentions has also been demonstrated.For example, evidence showed that there is a relatively large finger somatotopy in the human motor cortex [114], [115], and evidence from both ECoG and fMRI studies have found a degree of separability in the peak population responses for different fingers in the precentral gyrus [116], [117].These differential finger responses suggested the potential of hand gesture classification and finger trajectory reconstruction, and some important studies will be presented below.
For example, a finger-moving ECoG data set was collected from three epileptic participants and was made public in the BCI Competition IV [11].The data were collected when the epileptic participants performed individual digit flexion and extension repetitively.Various studies have been conducted on this data set [62], [81], [118], [119], [120], [121].For example, in one study, five fingers can be classified at 84.6% averaged across three participants [62].
Apart from the moving finger classification, a common and useful decoding target is hand gesture [60], [70], [122].For example, one study classified 4 hand gestures that were taken from the American Sign Language fingerspelling alphabet with accuracies of 97% and 74% for two participants [70].Another group achieved 73% for the 5-digit movement classification [122].Further, an online control over a prosthetic arm to perform three different hand gestures in real-time was conducted on one stroke participant [60].First, features showing characteristic power modulation during the movements were identified.Then, in the free-run stage, the participant controlled the prosthetic arm to mimic the actual hand movement using a hierarchical decoding model trained with features identified in the previous stage.The first decoder distinguishes between moving and idle state, while the second decoder discriminates which gesture was performed.The first decoder detected 61.0% movement onsets before 1 second of the actual movement onset, the second decoder achieved 69.2% accuracy of hand gesture decoding during the free-run stage.
Simultaneous neural control of reach and grasp movements was also possible using the MPL device [123].Two participants, holding a bulb with a pressure sensor, performed reach or grasp or reach&grasp tasks.To accomplish simultaneous control of reach and grasp, different channels were used to classify reach and grasp status separately, and then, the decoded status (commands) were sent simultaneously to an MPL device.
In contrast to the commonly used contralateral control, ipsilateral control is necessary when the contralateral cortex is damaged.In an example ipsilateral ECoG-based BCI study, a stroke patient was asked to repeatedly move his left or right thumb, or index finger [124].Using ipsilateral ECoG signals, Scherer et al. achieved a classification accuracy of 88% (four movements) and found that time-frequency patterns over sensorimotor areas for contralateral and ipsilateral movements overlap to a large degree.In addition, they observed less pronounced activity during ipsilateral movement compared to the contralateral movements.In another similar study, researchers tried to classify up to 7 hand movement intentions from servery paralyzed chronic stroke without residual hand movement using ECoG placed on the cortex ipsilateral to the paralyzed hand [73].With an SVM decoder, they achieved an accuracy of 61% for 7 hand movement intentions.
To date, most motor-based BCI studies used signals recorded from the sensorimotor cortex as a whole without differentiating the sub-regions.On the other hand, parts of the sensorimotor cortex may be damaged in some patients, and it is important to examine the performance of BCI using the motor or sensory cortex alone.This is particularly true for patients who have lost sensory input due to amputation, or paralyzed patients whose sensorimotor cortices may have undergone extensive reorganization.To demonstrate the feasibility of hand gesture decoding using only the motor or sensory cortex, various ECoG studies have been carried out [64], [67], [77], [104], [113], [125].For example, a study was conducted using a high-density ECoG to record from a small patch of Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
motor and sensory cortex in five patients with epilepsy [77].In the study, four complex hand gestures, taken from the American Sign Language fingerspelling alphabet ('D', 'F', 'V', and 'Y'), were performed.By analysing the cortical temporal and spatial response of finger movement, the primary somatosensory cortex (S1) high gamma component exhibited enhanced activation before movement onset.To compare the cortex areas, three decoding tasks were performed separately using signals from different locations: S1, the primary motor cortex (M1), and S1 and M1 together.They showed that, with a chance level of 25%, S1 achieved a classification accuracy of 76%, similar to those obtained using M1 (74%) and the sensorimotor cortex as a whole (85%).In addition, S1 exhibited characteristic spatiotemporal neuronal activation patterns.They further concluded that it is possible to obtain gesture decoding from a very small patch of the cortex using subdural HD ECoG grids.
Another similar study examined the decoding of hand gestures without sensory information [67].ECoG signals were recorded while participants were asked to perform five isometric hands and four finger movements.To remove the sensory information, the researcher tried different methods.For example, the signal from the sensory cortex was excluded during the decoding, or the participant performed motor imagery instead of motor execution.When sensory information was included, the classification accuracy of five hand gestures reached 68 -84% for different participants.However, the error rate increased substantially on average when sensory information was removed.
Finally, dexterous neuroprosthetic control during awake craniotomy is also important to verify optimal electrode placement during surgical implantation.This intraoperative decoding of hand gestures during awake brain surgery has been tested on the ECoG signals [126], [127].For example, Xie et al. investigate the possibility of gesture decoding intraoperatively during awake craniotomy [127].In their study, participants were asked to perform 4 types of hand gestures contralateral to the motor cortex covered by ECoG electrodes.The decoding accuracies of 4 participants were 90%, 96%, 91%, and 93%, respectively.This high-performance BCI during craniotomy is important to reduce the risk of reimplantation if the original cortical site is not optimal.

P. A.4.2 Hand Decoding Using Depth Electrodes
Hand gesture decoding is also possible using SEEG signals [12], [13], [32], [91].For example, Li et al. performed a hand gesture decoding on epileptic participants who had SEEG implanted for the epilepsy monitor [32].In their study, three participants were asked to perform three repeated hand gestures (scissors, rock, and thumb) while SEEG data were recorded simultaneously.The most informative channels were selected based on the modulation of frequency power in relation to task state (rest vs. task), and an LDA was used as a classifier.They showed that using only five channels, all participants were able to command a prosthesis hand to make three different hand gestures with an average accuracy of 78.70 ± 4.01%.Using the same paradigm, Wang et al. demonstrated that signals from PPC contained useful information to discriminate three hand gestures [13].In their study, 25 epileptic patients were recruited.The electrodes were partitioned into three regions of interest, including the posterior parietal cortex (PPC), the postcentral cortex (POC), and the precentral cortex (PRC).Using all channels or channels outside of PPC, they obtained the highest decoding accuracy of approximately 60% and 70%, respectively.The higher decoding accuracy by including PPC suggested that PPC contains useful information for decoding.They further study the activation sequence of three ROIs and found that the activation of ROIs was significantly sequential along the time course, where PPC activated first, PRC second, and POC last.
In another SEEG study, the role played by white matter in gesture decoding was evaluated [12].In their study, 30 epileptic patients were recruited and asked to perform 5 types of upper limb and hand gesture movements repetitively while SEEG signals were recorded simultaneously.By investigating both the alpha-and gamma-band power of electrodes from both white and grey matter, they demonstrated that both grey and white matter were activated.They further showed that the high gamma and alpha power magnitude in white matter was significantly lower than that of grey matter.When using signals from white matter, they obtained decoding significantly above the chance level, which demonstrated that white matter also contains useful information for decoding.

Q. A.4.3 Finger Trajectory Decoding Using ECoG
Finger flexion and extension are important for daily activity.To determine whether it is possible to faithfully decode the time course of flexion of each finger, a study was carried out on five participants who were asked to repeatedly flex each of the five fingers in response to the visual cues [56].Using five sub-band frequencies and local motor potential (LMP), a linear multivariate decoder was trained for each finger using data from 50 to 100 ms preceding the actual movement.When averaged across fingers, they achieved correlation coefficients ranging from 0.41 to 0.58 for different participants.Next, they demonstrated that gamma, high gamma, and LMP are highly informative for the decoding task.
The efficiency of LMP in decoding hand movement was further evaluated in another study in which participants were asked to open or close their hands in a slow grasping motion [58].A generalized linear model (GLM) was implemented to decode the time course of finger positions using LMPs of the selected ECoG electrodes.They succeeded in finger position prediction for 3 out of 4 participants.They further showed that the spatial distributions of the LMPs in the brain and their correlations with the kinematics of hand grasping were stationary across multiple sessions, involving variations in wrist angle, elbow flexion, and arm location.
In contrast, another reach-grasp decoding study showed that the LMP was inferior in prediction compared to all other subband features [59].In their study, features, the same as [58], were extracted.Contrary to the previous studies, LMP exhibited the worst decoding accuracy among all features using a generalized linear model (GLM).It is possible because the participant performed movement randomly, resulting in a nonperiodic movement.The contradicted conclusion suggested that further studies are still needed to better understand the nature of LMP.
In addition, 3D reconstruction of finger trajectory is also possible using ECoG [128].In their study, the participant was asked to perform extension/flexion tasks with three fingers.Then, offline trajectory reconstruction was performed using a sparse linear regression.They managed to obtain an average Pearson's correlation coefficient as high as 0.83-0.90.

R. A.5 Motion Intention Decoding
Most of the studies mentioned so far are synchronous BCIs, in which the participants were restrained to communicate in a predefined time frame.Asynchronous BCIs, on the other hand, would allow the user to operate BCI devices spontaneously.These asynchronous BCIs may support more practical applications as they are self-initiated and self-paced.However, asynchronous BCIs require the detection of the event in addition to identifying the properties of the event, such as the previously described study [69] and several other studies presented below.One example ECoG study used a centre-out paradigm in which five epileptic patients were instructed to move a cursor to 8 periphery targets using a joystick [129].Using historical data from the previous 1-second window, sliding every 100 ms, they showed that it was possible to predict movement onset for 2 out of 5 participants using an SVM decoder.Another ECoG study used signals from the prefrontal cortex and showed that different movements can be classified, with 74% averaged across all participants using signals 2 to 0 seconds before the actual movement [72].
Another possible way for intention identification is using connectivity between brain networks.With regard to the motor network, recent evidence suggests that movements arise from a distributed network [130].For example, besides the classic sensorimotor cortex, the dorsolateral prefrontal cortex (DLPFC) was proven to play a major role in behavior control, and it has a central integrative function for motor and behavior control [131], [132].An example ECoG study demonstrated the intention identification by investigating the temporal dynamics of connectivity, measured by mutual information, between and within the DLPFC and the primary motor cortex (M1) [133].They obtained classification accuracy of 94.9%, 93.0%, and 94.2% for three participants, respectively, which are higher compared to using the M1 region only.

S. Signal Process Methods
Invasive devices have the ability to record signals from a wide range of frequencies.While different frequencies represent very different neurophysical processes, the choice of band feature during decoding is an important task.Though the sampling rate of modern recording devices can be as high as 30k Hz [134], most invasive studies extracted features from multiple frequency bands below 500 Hz.To present a clear review of the extracted features, studies were categorized into four classes including those using low-frequency bands (LFB) feature (below gamma), high-frequency bands (HFB) feature (including gamma and beyond), a combination of LFB and HFB, and raw signals.The number of studies using these four features is presented in Fig. 8.This pie chart demonstrates most studies used both low and high-frequency features.On the other hand, high frequency was most frequently used compared to low frequency and raw signal.Overall, studies using raw signals consist of 4% of all studies.
It is evident that most studies used a combination strategy, and the HFB was more frequent.Most studies extracted features from several different bands, such as 8-12 Hz, 18-26 Hz, 70-100 Hz, and 110-150 Hz [3], in which the HFB changes were more focal than the LFB changes [11], [135], [136].LFB showed promising decoding results [54], and it has been proved that it was coupled with HFB during the motor control [137], [138].HFB was proved to be the most informative [3], and it demonstrated better decoding accuracy [30], [57], [139].This might result from the fact that HFB (gamma band in particular) was strongly correlated with the firing rate of individual neurons [140].HFB is advantageous also because of its spatial specificity which is crucial for differentiating motions controlled by closely located cortical areas [135], [141].Another feature, termed local motor potential (LMP), which can be calculated as the running average of the raw signal, was also proved to be highly informative in ECoG signals [11].In a following study, Gunduz et al. further demonstrated that the LMP feature yielded higher classification for decoding preparation, whereas HGA yielded higher accuracy for execution.However, the origin and nature of the LMP are still under debate and it seems this component is only effective in decoding slow-paced movement [58], or regular repetitive movements [59].
Once the features have been extracted, various decoding algorithms can be used, which can be classified into two categories: classification and regression.The simple regression methods can linearly model the relationship between brain signals and limb movements, such as the linear regression or its variants, including pace regression [56], ridge regression [142] or linear Wiener filter [71].To enhance the performance, nonlinear methods can be used, such as the time-varying dynamic Bayesian network and step-wise multilinear regression [38], [129], [143], Naïve Bayes model [69], Kalman filter [54].
In the classification task, there are also many decoders available.For example, regularized linear discriminant analysis (RLDA) was used to classify movement directions [64].A support vector machine (SVM) classifier was used to classify 5 hand movements [13].The Fisher Discriminant Analysis (FDA) classifier was used on features extracted by Discriminative Canonical Pattern Matching (DCPM) and Common Spatial Patterns (CSP) [144].
Besides the 'traditional' methods above, another type of nonlinear algorithm, the deep learning method, which has been proven to be state-of-the-art in various domains, is increasingly reshaping BCI research.This method is capable of achieving comparable, and most times better, decoding results without any manual feature engineering.This method has been extensively evaluated in non-invasive BCIs and has gained increasing attention in the invasive domain [31], [145], [146], [147].It is worth noting that many more deep-learning methods have already been evaluated using non-invasive signals, which can be readily adopted by invasive studies [148].

T. Location of Recording Electrodes
The electrode positions used in the above-mentioned studies vary widely.This is mainly because most of them were conducted on patients in which the locations of the electrode were determined solely based on the clinical needs.On the other hand, the opportunities to record signals from different brain regions boost our understanding of motor control.For example, even though many studies have demonstrated that signals recorded from the motor cortex were most effective for decoding [104], other exploratory studies have expanded our understanding of motor control and provided new methods for future BCIs.Such as the ipsilateral controlling [103], and regions other than the motor cortex were also the valid signal source [149].Therefore, it is of interest to know the location of the electrodes, to obtain an overview of potential brain regions for BCIs.
In this section, regions recorded by the reviewed studies were reported, as shown in Fig. 9.In detail, regions used for motor control were extracted from each paper.For each region, the number of studied publications was counted.Noted that multiple regions could be recorded in one study, therefore the total number of studied regions would be bigger than the total studies.
Extra care should be taken to interpret this plot.The granularity of regions in different studies is different.For example, while many studies specify the recording region as the primary motor cortex (MI) [61], [69] or primary somatosensory cortex (SI) [77], other studies referred to these areas as the sensorimotor cortex as a whole [113], [150].Without further information from the paper to better discriminate the recording sites, this paper used different granularity in the report.For example, while there are 7 papers recorded from the premotor cortex (PMC), they are not accounted for the paper recording from the motor cortex.
It is evident from this section that very different regions were recorded in the reviewed papers.Most existing studies recorded from the Frontal, Temporal, and Parietal lobes, while the most depth electrodes were located in STN, Hippocampus, the cingulate cortex, and the white matter.In Fig. 9, the sensorimotor cortex as a whole is the most frequently used Counts of papers recording from different brain regions.The inner pie represented the four lobes, while the outer pie denoted the sub-regions in each of the main lobes.Each annotation was in the format of region/count.For example, PMC/7 denoted there are 7 studies conducted from the premotor cortex.PFC: prefrontal cortex; PMC: premotor cortex; SMA: supplementary motor cortex; M1: primary motor cortex; Borca: Broca's area; PPC: post parietal cortex; STG: superior temporal gyrus; Wernicke: Wernicke's area; STN: subthalamic nucleus.
region.This is in line with our understanding that such a region contains the richest information about movement.In addition, many other regions were also used for the decoding, such as SMA, PMC, and PPC, which implies that movement is distributed in various regions.On the other hand, subcortical regions were also used for the decoding, including STN, hippocampus, and white matter.Overall, the sensorimotor cortex is most frequently used, comprising 59% of all reviewed studies.
The signals recorded from various positions have advanced our understanding of motor control.On the other hand, the variety of locations implied a distributed motor control network, which will be discussed in the discussion section.

III. DISCUSSION
This review has presented an overview of ECoG-, SEEG-, and DBS-based motor BCIs on human participants.Contrary to previous non-invasive or invasive reviews, this review was presented from the perspective of the decoding tasks, which were categorized into kinematic decoding, kinetic decoding, identification of body parts, dexterous hand decoding, and detection of motion intention in this review.Among the studies reviewed, there are significant differences in the brain regions recorded.In this section, the distributed cortical representation of motor control and its implications for future BCIs implantation will be discussed.Then the difference between surface and depth electrodes will be further compared.In addition, challenges faced by current motor BCIs will be discussed, including closed-loop BCIs and BCIs on patients Two brain networks engaged during motor control.The left subplot depicts a path that starts from the basal ganglia and prefrontal cortex to the supplementary motor area (SMA) and the pre-supplementary motor area (preSMA) and finally arrives at the M1.The right subplot shows another path that starts from the sensory cortex (S1) to the parietal cortex, and premotor cortex and finally reaches M1.
with dysfunction which is the primary target of BCIs.Finally, opportunities brought by recent developments to enhance overall BCI performance will also be highlighted.

A. Distributed Motor Network
There is evidence suggesting that movement arises from a distributed network that is centered in the primary motor cortex (M1) [130], [151].In this distributed view, M1 collected information from two paths during motor control, as depicted in Fig. 10.One path starts from the basal ganglia and prefrontal cortex to the supplementary motor area (SMA) and the pre-supplementary motor area (preSMA) and finally arrives at the M1.Another path starts from the sensory cortex (S1) to the parietal cortex, and premotor cortex and finally reaches M1.
The convergence of these two paths at M1 means it is an ideal recording target for motor BCIs.More evidence showed that M1 exhibited the highest tuning strength by the arm movement direction, velocity, and speed [61], and was assigned the highest weight by a linear decoder in a 2D cursor control task [11].On the other hand, the sensorimotor cortex as a whole is a closely connected network consisting of the primary motor cortex (M1) and the primary somatosensory cortex (S1), both of which are somatotopically organized.This makes S1 a logical alternative target for BCIs.In addition to the reflection of proprioceptive feedback from muscle and joint receptors [152], S1 was also proven to be activated in the absence of movement.For example, spinal cord injury fMRI studies indicated both M1 and S1 were activated during the attempted movement [153], [154].Direct evidence from an ECoG study demonstrated a clear broadband  increase in S1, and a decrease in the beta band one second before movement onset [125], [155].These results indicate that there is information over the S1 hand area that is decoupled from sensory feedback.Furthermore, predictive theories of motor control have recently hypothesized that S1 might play a role in generating an efferent copy, which is an internal prediction of the sensory consequence of a volitional movement [155], [156].Direct evidence of the efferent copy came from an ECoG study, which demonstrated there was a sequential high gamma (HG) activation in pre-motor (PM), S1, and M1 regions during movement preparation, in which high-gamma (HG) activation in PM preceded S1 by an average of 53 ± 13 msec while S1 activation preceded M1 by 136 ± 40 msec [155].
The distributed motor-related network is also bilateral which involves both the contralateral and ipsilateral hemispheres during voluntary motor movements [157], [158], [159].Such a bilateral encoding mechanism could be advantageous for BCI applications when one hemisphere is damaged.Using ECoG, it is demonstrated that the spatial and spectral encoding of contralateral and ipsilateral limb kinematics were similar, while ipsilateral movement produced less pronounced activity compared to the contralateral movement [124].The possibility of cross-prediction of kinematics between arms has also been proved using ECoG signals [103].The above studies suggested that information about movements is more bi-hemispherically represented in humans.The ability to use ECoG to decode the kinematics of the ipsilateral hand underscores the possibility that patients can use signals from their unaffected hemisphere to control a BCI.
Besides the sensorimotor cortex, two brain circuits exist in the volitional movement control and both converge on the primary motor area [151], [160], [161].One of the pathways starts from the supplementary motor area and receives input from the basal ganglia and the prefrontal cortex (PFC) [162], implying the possibility of decoding using PFC signals.What's more, the dorsolateral prefrontal cortex (DLPFC) was proven to play a crucial role in exerting control over behavior and has a central integrative function for motor and behavior control [132].Indeed, the participants of PFC were evident in the ECoG recordings, during both the intention of moving or performing wilful actions [11], [163].In another example ECoG study, the prefrontal area has been found to be a source of slow cortical potential, while high gamma activations are observed in the premotor and parietal areas during movement preparation and in the primary motor areas during execution [143].The early appearance of the slow cortical potential in PFC could imply the presence of a cognitive process such as motor planning or motor preparation before actual movement.Based on this hypothesis, using a self-paced hand grasping and elbow flexion task, Ryun et al. showed that the prefrontal area was sufficient to allow the classification of different movements using signals from PFC (2.0 s to 0 s) before actual movement in four human participants [72].In another ECoG study of connectivity dynamics within and between M1 and DLPFC, the researcher demonstrated that it is possible to decode motion intention using intra-or interregional connectivity [133].
Aside from PFC, the posterior parietal cortex (PPC), has also been used to decode the intended endpoint of reach [164], and continuously control the trajectory of an end effector in a monkey study [142].The saccade, reach, and grasp areas have also been identified in PPC using a monkey model [165].An MEA-based BCI using human participants also demonstrated signals from PPC can be used to control a cursor on a computer screen and a robotic limb.In addition, the signal in PPC showed different activation patterns when participants moved the same hand to mouth, cheek, or forehead [166].Such neural encoding of behaviourally meaningful Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
actions opens the possibility of high-level intention decoding which enables patients to perform complex movements without requiring attentionally demanding moment-to-moment control.One SEEG study showed that PPC signals improved the performance of hand movement classification tasks [13].
The distributed motor-related signals, presented in this work based on human experiments, argue for a distributed neural encoding that extends the idea of neural ensemble physiology proposed by a review paper based mainly on animal studies [167].By investigating the neuron-dropping curves (NDCs), the author argued for the 'distributed-coding principle' which stated that a single motor parameter was processed within multiple cortical areas.This necessity of a large-scale and distributed recording in BCIs was also supported by recent findings in simultaneous large-scale neuron recording using more advanced high throughput devices, such as the Neuropixels probes [168], as argued in another recent review [169].In that review, Urai et al. argued for a distributed and sparse nature of neural responses.Then, they further call attention to a potential criticism of large-scale recording studies: they are observational rather than hypothesis-driven and lack the ability to distinguish concrete mechanistic models.With an example evidence accumulation model in a decisionmaking experiment, the author showed that the ability to connect the vast heterogeneous signals to theoretical mechanisms may come short, at least momentarily, and argued for new theoretical frameworks to integrate brain and behavior.
In summary, with the convergent consent of a distributed motor network, challenges have been recognized as to the simultaneous recording of the distributed large-scale neural ensemble, the corresponding theoretical frameworks, as well as the decoding algorithms that are capable of handling high dimension data.

B. Surface Vs Depth Electrode
The intracranial electrode can be broadly classified into two types: surface and depth electrodes.This section will compare these two devices in terms of the recording sites and chronic implantation.
1) Recording Site: The most obvious difference between these two electrodes is the recording site.While ECoG records from surface areas of the cortex, depth electrodes record from sparse locations within white matter and subcortical structures.Such surface electrodes are advantageous as cortical surfaces contain the richest information [130], [151].On the other hand, electrodes placed within white matter were believed to record mostly the volume-conducted signals from the grey matter during motor tasks [170], [171], and contain less useful information.Another study comprehensively compared neural response in different cortex and subcortical regions during hand/arm movement [12].From 32 epileptic patients performing different hand/arm motions, the SEEG signals were collected and classified into one of five movements using an SVM classifier.They concluded that while sensorimotor areas carried the richest discriminative information, cortical regions including the posterior parietal, prefrontal cortex, temporal occipital cortex, and subcortical regions including white matter and insula also contain useful information for decoding.
Although SEEG devices can record discriminative information from the subcortical regions, another critical question is how much information can be decoded, i.e. how complex a BCI driven by the SEEG signals can be.As in [12], although signals from subcortical regions, such as the white matter, were proved to be useful, the classification (3 classes, 33.3% at chance level) accuracy was still low: 42%.From the previous review, it is also evident that the ECoG-based BCIs achieved a higher degree of freedom (DOF) than the SEEG-based BCIs.Therefore, considering the recent discoveries about the topographic organization of motor-related brain areas, as opposed to ECoG, SEEG alone may not be the ideal technique for motor decoding due to its sparse sampling of eloquent cortical surfaces.Another possible reason for the reported inferior performance in motor decoding using the SEEG signals is the limited studies about SEEG-based motor BCIs.With the interest in SEEG-based BCIs increasing, it is anticipated that more and richer motor information will be decoded from the SEEG signals.
Despite the inferior performance in motor decoding, because of the less invasiveness (safer), depth electrodes might still be preferred when placed in the preferred locations and only a few electrodes were needed to control the BCI system.In an example SEEG-based study, the visual motion response extracted from the middle temporal visual area (MT) was used to drive a spelling BCI [172].With preoperative functional magnetic resonance imaging (MRI), the responsive areas and the best electrodes were identified from the epileptic patients.An information transferring rate of 62 bits/min can be achieved using only three identified electrodes over the MT area.Although the electrode locations were not decided by the preoperative MRI, but solely by the clinical need for epileptic treatment, this study showed that it is possible to identify the task-related regions, either cortical or subcortical, before surgery and use the mini-invasive SEEG electrodes for signal recording and controlling.In this situation, only a few electrodes were needed to control the spelling system and they can be implanted mini-invasively through a small burr hole, which is much safer than the craniotomy required for the ECoG implantation.
Another piece of evidence of the preference for the SEEG electrode comes from a speech decoding study, which used both ECoG and SEEG signals.In their study, the SEEG electrodes achieved as good decoding accuracy as the ECoG [173].This may be because the SEEG electrodes penetrated Heschl's gyrus (HG) which contains the primary auditory cortex.The good performance and the less invasiveness (small burr holes) make the SEEG preferred under this circumstance.In conclusion, while ECoG is superior in most cases, SEEG might be preferred when the electrode locations are optimal.
Besides the high performance in motor decoding for the ECoG signals, recently fast-growing high-density, high throughput µ ECoG, or micro-ECoG, will further enhance its performance.For example, high-density ECoG has been proposed using electrodes with microwire below 10 µ m and spacing of 4 mm [109], 1 mm [174], [175], or even 600 µ m [176].The high spatial density was proved to be necessary by various studies [177], [178].For example, motor representations of different fingers were shown to be located within an area of approximately 1 cm 2 [116], which means that standard clinical grids will not be able to capture detailed information to distinguish different fingers [77].A similar conclusion came from another study that demonstrated that details of hand/wrist movements can be inferred from ECoG signals recorded over the motor cortex at a resolution of 1 mm [68].From this point of view, ECoG would be a better candidate for motor BCIs.
2) Chronic Implantation: Another main consideration of invasive motor BCIs is chronic implantation.While there is a lack of systematic evaluation of long-term ECoG-based BCIs, some studies suggested the potential of ECoG electrodes for chronic implantation.For example, Larzabal et al. found the signals were stable on two patients implanted with epidural ECoG for three years [179].Nurse et al. demonstrated implantation of subdural ECoG for 766 days [180].They showed that high-gamma signals could be recorded throughout the study, though there was a decline in signal power for some electrodes.
Similarly, there has been no study on the chronic viability of SEEG electrodes.However, since the leads employed in SEEG and the associated surgery are similar to those used for DBS procedures, the chronic implantation of the Deep Brain Stimulation (DBS) electrodes might suggest similar long-term safety for SEEG electrodes.DBS is widely used as a treatment for tremors, dystonia, Parkinson's disease (PD), obsessive-compulsive disorder (OCD), anorexia nervosa, and depression [36].DBS has been clinically approved for longterm implementation.For example, in a 6-year follow-up study conducted by Kennedy et al., they observed no significant adverse events [181].Castrioto et al. reported a 10-year followup study of DBS treatment of 18 PD patients [182].These reports on the long-term usage of DBS might suggest that the SEEG electrodes are safe in chronic BCIs.However, the long-term effectiveness of stimulation does not necessarily mean long-term stability of the recordings, as the growth of encapsulation tissue might have a different effect on the recording than stimulation.Therefore, the chronic BCIs based on both surface and depth electrodes still require further evaluation.

C. Challenges and Opportunities
In the above-reviewed studies, invasive BCIs based on surface and depth electrodes have already shown great potential.However, there is still a long way to go before practical usage in the real world.In this section, the challenges and prospects faced by the current invasive BCIs will be discussed.
1) BCIs on Potential Target Patients: Although signals recorded by ECoG-, SEEG-, and DBS-electrodes from epilepsy patients and Parkinson's patients have been used for BCI studies, the current BCIs on these patients do not aim to treat or cure their motor dysfunctions and neurological disorders.The BCIs are for research purposes, which are neither a treatment nor an assistive method.In addition, most of the invasive BCI studies rely on very few patients which limits the generalization abilities of invasive BCIs.The number of patients recruited and the corresponding number of studies was presented in figure 11.It is clear that most of these invasive studies use very few patients.
It also should be noted that most of the presented studies are about the upper limb and only a few of them decode lower limb movement.This is mainly because of the location of the related brain area.In the human brain homunculus, the eloquent cortex of the low limb resides inside of the longitudinal fissure between two hemispheres and is not readily reachable during the treatment of neurological disease.Therefore, the decoding of lower limb movement using invasive BCIs is still under-investigated, and further studies are needed.
That said, BCIs still have the potential to be an assistive method for certain disabled patients whose motor functions are lost or impaired due to lateral sclerosis (ALS), locked-in syndrome (LIS), stroke, etc.In this section, the challenges and opportunities faced by BCIs for this group will be presented.
Patients with motor dysfunction mainly rely on motor imagery to modulate neural signals and control external devices.However, long-term damage to the brain, such as amyotrophic lateral sclerosis (ALS), locked-in syndrome (LIS), or stroke, will potentially affect the generation of neural activity.Direct evidence came from an ECoG study concluding that the condition of a patient with LIS may have significant effects on the spectral components of the low-frequency band (LFB) in the sensorimotor cortex [137].Another ECoG comparison study also demonstrated significantly inferior decoding accuracy in patients with chronic motor impairment [183].
Despite the difficulties, it is still possible for invasive motor BCIs to be used by motor impairment patients.For example, an ECoG study of four chronic stroke patients (with no sign of residual hand movement) demonstrated an average accuracy of 61% (chance level 15.6%) in motion intention decoding using ipsilateral ECoG [73].More evidence of preserved motor physiology came from another ECoG study on a patient whose left arm was lost for three years [184].Their study found classic frequency response (spatially focused increase in HFB and distributed decrease in LFB) reported in the healthy subjects.Then, they showed that decoding three different imaged finger movements can reach up to more than 90% based on the first 1-3 s of the HFB power.In summary, despite the impaired neural modulation, invasive BCIs are still possible for chronically disabled patients.
2) Closed-Loop BCIs: Most of the current BCI studies are open-loop, in which BCIs users would rely heavily on visual feedback during the interaction without somatosensation.This might preclude fine motor control over objects, or interaction with objects outside the line of sight.On the other hand, sensory feedback is important to both user preference and actuator performance, such as completing the transition from reaching to grasping immediately after touching the object or modifying the grasping strength to prevent slipping.
To restore somatosensation, two aspects are necessary: 1) tactile perception, such as texture, temperature, or vibration, extracted from the sensors equipped on the actuator; 2) transmitting the perception to the brain.
Various sensors can be used to meet the first requirement.For the second aspect, a common practice is to apply electric stimulation on the surface of the cortex.Previous research has demonstrated that human participants can discriminate the perceptual intensity elicited with either varied frequency or varied amplitude using both ECoG [185], [186], [187], [188], [189], [190] and SEEG [191].Despite these results, many issues still exist that have complicated the implementation of artificial sensation in BCIs.For example, despite that electrical stimulation of the cortical surface can produce somatotopically localized sensations, these sensations are not natural [186], [188], [192], [193].
Also, the electric field generated from the electrical stimulation may contaminate the decoding signal and deteriorate the BCI performance.Besides, the possible adaptation or habituation to the stimulating signal over time means higher stimulating intensity will be required over time [194].Finally, the possible adverse side effects of long-term electric stimulation, such as electric burn and alteration of the normal neural network [195], are still unknown.This is a particularly critical consideration for chronic closed-loop BCIs.
3) Regulatory Challenges: As the development of BCIs speeds up and more and more advanced applications emerge, there is also an increasing urgency for regulation [196].However, many challenges exist because of the associated risks and the fragmented regulatory framework.One of the highest possible risks behind this technology is the fact that the underlying neural mechanism of the brain is still not fully understood, which means users might experience chronic side effects and unexpected results.Besides safety, there are also broader ethical considerations.For example, there is a potential privacy risk because there are might much more can be decoded by the advanced algorithms in the future from the collected data than what's expected during the experiment.What complicates the matter even further is that the line between therapy and enhancement for BCIs is difficult to draw precisely.For example, a device designed to correct cognitive impairment could result in cognitive enhancement [197].In addition, BCIs's potential in human augmentation might widen the gap between social classes if accessible only to the wealthy.This can be extended to the national level: if used in the military setting, the technology might pose a national security risk [198].On the other hand, the fragmented regulatory framework represents another potential risk.For example, Johnson, etc. argues that the presence of multiple administrative agencies (the U.S. Food and Drug Administration (FDA) and the Federal Trade Commission (FTC)) with partial power to regulate BMIs creates potential governance problems [199].In sum, there is a discrepancy between the fast development of invasive BCIs and the regulation, and improved interagency coordination is urgently needed.
4) Opportunities: Despite the challenges faced by current invasive BCIs, there are still opportunities.For example, new materials, intelligent devices, and algorithms are now accelerating and reshaping the development of BCIs.For more information about the hardware (electrode, material, etc.) and the advanced manufacturing process, readers are referred to existing reviews [200], [201].
For example, high density and high channel number ECoG devices [177], [202], [203], as well as wireless, fully implanted devices [204], [205], have been proposed and proved to be beneficial in various studies.
In addition, the fast-growing algorithms, especially machine learning and deep learning methods, are boosting the decoding performance to a higher level.Compared to the simple linear models, which were widely used in current invasive BCIs, the non-linear model was proved to be more robust and resilient to noise [101].Especially, deep learning methods were proved to be effective and comparable (most times superior) to the 'traditional' method for both invasive [81], [206] and noninvasive BCIs [207], [208], [209], [210].
What's more, the performance can be further enhanced with other strategies, such as hybrid BCIs, shared control, etc.The hybrid BCIs often achieve better results by combining different brain signals or different controlling paradigms [211].It is also possible to combine other biological signals, such as electromyography (EMG), or eye movement [87].On the other hand, the shared control strategy has the potential to achieve a more stable system by incorporating and fusing external information, such as computer vision and sensor recordings [212], [213].In addition, it is also beneficial to use intelligent robotics in which only several simple commands are needed to trigger the robotics to complete complex tasks [214].
Therefore, it is reasonable to believe that the new algorithms and advanced materials will further enhance the performance and safety of invasive BCIs.

IV. CONCLUSION
This review provided an overview of the current research into ECoG-, SEEG-, and DBS-based motor BCIs studies conducted on human participants.Uniquely from previous invasive and non-invasive work, this review was presented from the perspective of the decoding tasks, which were categorized into five groups: kinematic decoding, kinetic decoding, identification of body parts, motion intention decoding, and dexterous hand decoding.The reviewed literature revealed a distributed motor network, in which the contralateral primary motor cortex (M1) contains the richest information while other regions, such as the ipsilateral M1, sensorimotor cortex as a whole and many other regions also contain useful information.It is also demonstrated that ECoG-based BCIs were superior by recording a large area of the cortical surface.However, the safety issue in long-term usage of both surface and depth Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.

Fig. 2 .
Fig. 2. The number of publications about motor BCIs using three ECoG, SEEG, and DBS in recent years (until December 2022).The blue and orange bars represent publications made by ECoG surface electrodes and depth electrodes, respectively.Since 2012, there has been a clear decreasing and increasing trend for BCIs based on surface and depth electrodes.

Fig. 4 .
Fig.4.Illustration of the kinematic and the kinetic parameters when moving an object with a prosthetic hand.Kinematic decoding studies the trajectory parameters, such as velocity, speed, and location, in a 3D space.The kinetic decoding studies forces applied to the object.

Fig. 5 .
Fig. 5. Publications related to kinematic decoding were divided into either discrete or continuous parameter types in the inner pie chart, and each type was further categorized into surface and depth electrodes in the outer pie.

Fig. 7 .
Fig. 7. Publications related to dexterous hand decoding were grouped into discrete and continuous types, and each type was further categorized into surface and depth electrodes in the outer pie.

Fig. 8 .
Fig. 8. Pie plot of the number of studies using different features.LFB, HFB, Raw, and Comb in the plot mean low-frequency bands, highfrequency bands, raw signal, and the combination of both LFB and HFB, respectively.

Fig. 9 .
Fig. 9.Counts of papers recording from different brain regions.The inner pie represented the four lobes, while the outer pie denoted the sub-regions in each of the main lobes.Each annotation was in the format of region/count.For example, PMC/7 denoted there are 7 studies conducted from the premotor cortex.PFC: prefrontal cortex; PMC: premotor cortex; SMA: supplementary motor cortex; M1: primary motor cortex; Borca: Broca's area; PPC: post parietal cortex; STG: superior temporal gyrus; Wernicke: Wernicke's area; STN: subthalamic nucleus.

Fig. 10 .
Fig. 10.Two brain networks engaged during motor control.The left subplot depicts a path that starts from the basal ganglia and prefrontal cortex to the supplementary motor area (SMA) and the pre-supplementary motor area (preSMA) and finally arrives at the M1.The right subplot shows another path that starts from the sensory cortex (S1) to the parietal cortex, and premotor cortex and finally reaches M1.

Fig. 11 .
Fig. 11.Counts of publications according to the number of participants recruited.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.
Authorized licensed use limited to the terms of the applicable license agreement with IEEE.Restrictions apply.