Review of real brain-controlled wheelchairs

1 Review of real brain-controlled wheelchairs Á. Fernández-Rodríguez, F. Velasco-Álvarez and R. Ron-Angevin Department of Electronic Technology, University of Malaga, 29071, Málaga, Spain. [email protected]; [email protected]; rro[email protected] Abstract. This paper presents a review of the state of the art regarding wheelchairs driven by a brain-computer interface (BCI). Using a brain-controlled wheelchair (BCW), disabled users could handle a wheelchair through their brain activity, granting autonomy to move through an experimental environment. A classification is established, based on the characteristics of the BCW, such as the type of electroencephalographic (EEG) signal used, the navigation system employed by the wheelchair, the task for the participants, or the metrics used to evaluate the performance. Furthermore, these factors are compared according to the type of signal used, in order to clarify the differences among them. Finally, the trend of current research in this field is discussed, as well as the challenges that should be solved in the future. Keywords: brain-computer interface, wheelchair, real environment, review. 1. Introduction One of the main objectives of research groups working on assistive technologies is to improve the life quality and autonomy of people affected by motor neuron diseases (MND), such as amyotrophic lateral sclerosis (ALS). A brain–computer interface (BCI) is a tool to establish an additional communication channel between the user and a particular device, through their brain activity (1). Therefore, numerous applications are proposed for these users: e.g., managing a speller matrix (2), a robotic arm (3), a telepresence robot (4), or a domotic system (5), as well as applications focused on neurorehabilitation (6). Through these interfaces, people affected by a MND could gain some autonomy by means of the addition of this new communication channel that does not require the use of the motor system. Several types of BCI exist, depending on various features that will be briefly explained below. One of the most important features is the recording technique of the physiological signal, such as electroencephalography (EEG), functional magnetic resonance imaging (fMRI) or nearinfrared spectroscopy (fNIRS). However, the most used physiological signal is the EEG, mainly because of its adequate temporal resolution, portability, and relative low cost (7). There are also numerous types of EEG signal that can be recorded. Initially, we can distinguish between an endogenous signal, which is evoked at will by the participant, or an exogenous signal, which is evoked by an external stimulus presentation (8). Among the endogenous signals, the slow cortical potentials (SCP) are produced by a change in the level of cortical activity (9), Event-Related Desynchronization/Synchronization (ERD/ERS) changes elicited by motor imagery (MI) tasks, and other signals corresponding to different mental tasks, such as objects’ mental rotation or word association (10). On the other hand, regarding exogenous signals, the P300 signal is a positive peak that appears in the EEG approximately 300 ms after the presentation of a rare stimulus (2); the steady-state visual evoked potentials (SSVEP) are changes in the neural activity located at the visual cortex that occur at the same frequency as a blinking stimulus (11). Regarding the kind of electrode, there are two types: dry and wet. The second needs an electrolytic gel to record the signal. Most electrodes used in laboratories are wet due to a higher quality reception of the EEG signal because of lower impedance (12). These electrodes are placed at the scalp following the International 10/20 system which specific configuration and number of electrodes depends of technique factors, as the EEG signal to register, or practical factors, as the time required for the set up. Nevertheless, in recent years, many brands have marketed dry electrodes (e.g. Neurosky, Emotiv Systems or OCZ Technology) that have come to be used in some studies to prove their accuracy, and to see whether they might be suitable for safe operation for people with MND (e.g. (13,14)). The use of BCI systems for navigation of multiple devices – in both real and virtual environments – has been the subject of numerous investigations. Navigation in virtual environments can be for simple tasks such as moving a car to the right or left of the road (15), or more complex ones, such as the management of a character in a videogame (16). The use of these interfaces in real environments, which are less controlled and dependent on the use of devices that interact with them, may be more challenging, because mistakes can have real consequences for users. However, the development of these applications can be more useful, because they allow direct user intervention in their environment through their brain signal. Furthermore, as specified by Millán et al. (17): “BCIs must be combined with existing assistive technologies (AT), especially those they already utilize”, so a braincontrolled wheelchair (BCW) would be a perfect example of this combination. The first BCW was published by Tanaka et al. in 2005 (18). One could cite as a precursor the paper that of Millán et al. (19), in which the ability to manage a small mobile robot using an EEG signal was shown, setting a precedent for the control of a larger real system, such as a robotic wheelchair. Nevertheless, the leap from a remotely managed device to a wheelchair needs to be done carefully, as an accurate system is needed in which the safety of the user is not in danger. Therefore, the proposals to control a robotic wheelchair must 2 overcome some challenges to improve the quality of life of patients with severely impaired motor abilities. First, one of the issues to be faced is the use of reliable navigation systems to ensure the user safety and offer a flexible displacement in the environment so that the user can move comfortably and freely. In this way, it could be interesting to show the most useful and innovative proposals that could be used in future prototypes. Similarly, it should not be forgotten that the target population of these interfaces are patients with severely impaired motor abilities, so that the proposals should be adapted to these users and offer them an experience as pleasant as possible. Finally, due to the numerous proposals made in this area, it would be desirable to detail how the performance was evaluated and what were the participants’ characteristics (e.g. the number of users who tried to control the wheelchair and complete the tasks, if they were trained or had some kind of disease). In 2013, Bi et al. (20) published a survey of BCIcontrolled mobile robots; however, the present paper will focus on the particular field of BCWs, including the development and characteristics of the different proposals tested in real environments to date (figure 1), describing the type of signal that was used to control the navigation system device, users who handled it, the tasks performed by them, the navigation of the interface, and the metrics used to evaluate the performance. Therefore, this review includes papers that use a BCI system to control a wheelchair in a real environment, and that detail the BCW with enough data to be classified, based on the mentioned characteristics. The different interfaces are compared according to the signal used, in order to highlight the advantages and disadvantages between them. Finally, it is important to advise that in the case to find similar BCWs and authorship, we only include the most detailed paper. Figure 1. BCW papers compiled in this review. 1.1 Glossary In this brief section, we will define some terms used in the rest of the paper. Even when most of them are commonly used terms, it is worth defining them clearly in order to avoid any ambiguity. ● User’s tasks: In order to control a BCI, users perform different tasks whose consequences are predictable, so that they can be used as inputs. These control tasks include mental strategies (such as the MI of limbs or selective attention tasks) (21), but actual muscular tasks as well, being that some BCI systems are assisted by real movements. ● P300 BCW and SSVEP BCW: wheelchairs that only rely on P300 and SSVEP, respectively. ● ERD/ERS BCW: wheelchairs that depend on the users’ control of their electrophysiological activity through the execution of mental tasks that affect the EEG causing ERD/ERS changes. The used mental tasks include tasks such as MI, mental calculations, or word association (10). It should be noted that a BCW that analyzes the EEG caused by actual movements (not MI) will be classified in this paper as “muscle-assisted” (see next categories) for comparison purposes. ● Hybrid BCIs are commonly accepted as systems relying on one EEG input combined with one or more channels (that can be EEG, electromyography (EMG), electrooculography (EOG) or movement detection, among others). However, as one of the focuses of the paper is to compare parameters of similar systems, we have defined in this paper a subgroup of hybrids systems so that motor actions were excluded (see next definition). ● Hybrid-mental BCW: wheelchairs that are based on more than one kind of EEG signal (e.g. ERD/ERS and P300), as a consequence of different mental strategies, excluding any kind of real motor action. ● Muscle-assisted BCW. As mentioned above, we think that systems using motor actions (even when analyzing their consequent EEG) should not be compared with those that use purely mental tasks. For this reason, we have included in this group called Muscle-assisted BCW two kinds of wheelchairs: i) those that use EEG signals elicited by actual motor execution; and ii) hybrid wheelchairs that, in addition to purely cognitive tasks, use muscular activity as information input (detected by means of EEG artifact, EMG, or EOG). ● Low-level navigation: the control of the wheelchair is achieved through simple navigation commands, such as “move forward” or “turn right”, and basic supports as stopping the wheelchair when obstacles are encountered. In this way, users can perform any path they want to, having fine control of the specific movement. The system does not assist the execution of the selected command. ● High-level navigation: these systems let users have a rough control of the BCW, selecting high-level commands such as “take me to the kitchen” or “leave this room.” The BCW must be equipped with some intelligence so that the specific path to the selected objective is transparent to users (in other words, the user does not select specific low-level commands). ● Shared-control navigation: both the user and the system share the control of the BCW (22,23). This can be done in two ways: i) users generate low-level commands, while the system assists the navigation with features such as obstacle avoidance, or maximum likelihood command execution; and ii) users can switch between a lowand a high-level navigation mode. ● Discrete control: the selection of a navigation command implies a prefixed movement, e.g. a turn of 45 or 90 degrees or a fixed advance distance of 1 m. ● Continuous control: the user can control the extension of the movement after the selection of a navigation command, e.g. the turn amplitude or the advance length. 3 Usually, the movement continues as long as the user keeps the command active. 2. State of the art in BCW The use of invasive methods for capturing signals in BCI systems is less extended than non-invasive methods, i.e. the EEG (7). In the case of BCW, all the references cited in this review used EEG (figure 2). The different BCW papers have been divided regarding the signal used as input, so that five categories can be established: ERD/ERS, P300, SSVEP, hybrid-mental and muscle-assisted BCW. The number of BCWs found for each category is: 9, 6, 6, 5 and 9, respectively. It is worth mentioning the absence of SCP-based systems, possibly due to the low information transfer rate (ITR), and the need for a longer training time to acquire control, even compared with ERD/ERS (24). Figure 2. General model of a BCW. 1) The EEG is captured through electrodes at the Acquisition stage. 2) The raw signal is analyzed to obtain some significant characteristics called Features. 3) The main part of the Signal Processing uses the previous features to establish a Classification of the signals into a minimum of two groups; at this point the BCI system decides what state corresponds to the current EEG. 4) After classifying the signals, the system actuates in the Control stage moving the wheelchair. This movement is a Feedback for the subjects that helps them on the control of their EEG signals and, consequently, on the control of the BCW. Within each type of EEG signal, the navigation system used by the BCW (which may be low-level management, high-level or shared control) is detailed. Then we will describe the most important features related to the participants (e.g. users affected by MND) (table 1). Next, the characteristics of the interface will be explained, referring to the commands available to the user (e.g. turn left/right, go forward or stop) and the tasks that must be performed for the selection of a command; i.e., voluntary mental tasks, selective attention to the stimulus, or muscle action (table 2). Regarding the number of user tasks, these should be as few as possible, because the handling by the user will be simpler and, in the case of tasks that modulate ERD/ERS signals, fewer mistakes will be made in the classification (25,26). On the other hand, it is of interest to maximize the number of commands, as they are the options available to the user to handle the BCW and move around the environment autonomously. Therefore, using a ratio between the number of available commands (AC) and the number of user tasks (UT) it is proposed, the command to task ratio (CTR): 𝐶𝑇𝑅 = 𝐴𝐶 𝑈𝑇 (1) It is worth mentioning that this ratio is not suitable for comparing BCI systems based on different EEG signals, as the tasks to be performed are be different cognitive processes. However, in the case of systems with similar EEG input, this ratio points out how easily new commands can be added. To continue, the extraction features and classification methods used to categorize the different user’s task in each signal category will be mentioned (table 3). Finally, the metrics used for the evaluation of user performance in handling the BCW, such as the accuracy or the usability, are mentioned (table 4). 2.1 ERD/ERS BCWs ERD/ERS signals are those elicited at the user’s will by certain tasks that cause a variation in the amplitude of the neuronal rhythmic activity (8). However, learning to modulate these signals on a voluntary basis is usually complex, requiring more learning time for users than other systems (7). Despite this difficulty, as Nicolas-Alonso and Gomez-Gil (7) showed, the use of these signals has several advantages that should not be overlooked, such as: i) they are independent of any stimulation; ii) they can be operated by free will; iii) they are useful for users with affected sensory organs; and iv) they are suitable for cursor control. 2.1.1 Navigation. Five BCWs based on ERD/ERS signal had a low-level navigation system (18,34,43,44,57), three used a shared management (28,35,46) and one with a highlevel navigation system (59). Therefore, it can be seen that the handling of low-level is used to a greater degree than shared control or highlevel. The ERD/ERS signals may be of particular interest in low and shared navigation because they can offer continuous control of the BCW with a few low-level commands (e.g. forward, back or stop, turn left and turn right), to allow the free movement of the chair through the environment (60). However, thanks to the shared-control systems presented here, the management of the chair is assisted by various tools, such as obstacle detection and dodge, or assistance in selecting the most appropriate command, depending on the specific situation in the environment. 2.1.2 Participants. Referring to publications with ERD/ERS signal BCWs, these studies had an average of 3.22 participants (σ = 1.39). It is noteworthy that no study presented participants affected by MND. This low number of participants may be related to the difficulty in acquiring a proper and safe operation of a BCW which could imply an extensive training based on the ERD/ERS modulation of the EEG signal, as well as difficulties aggravated in the case of users with MND (e.g. security, communication, mobility and placement of instrumentation). Furthermore, the control of ERD/ERS signals may require a long training for some users. 4 Table 1. Compiled papers and main characteristics in chronological order. BCW Year Signal Navigation system Control Subjects (18) 2005 ERD/ERS Low-level Discrete 6 (27) 2007 P300 High-level Discrete 5 (28) 2009 ERD/ERS Shared Continuous 3 (29) 2009 P300 Shared Discrete 5 (30) 2009 SSVEP Shared Discrete 9 (31) 2010 ERD/ERS and P300 High-level Discrete 5 (32) 2010 P300 High-level Discrete 1a (33) 2010 P300 Low-level Discrete 1 (34) 2011 ERD/ERS Shared Mixedb 4 (35) 2011 ERD/ERS Shared Discrete 2 (36) 2012 ERD/ERS and EMG High-level Continuous 3 (37) 2012 ERD/ERS Low-level Not specified 1 (38) 2012 Alpha band and EEG artifact Low-level Mixedb 7 (39) 2012 ERD/ERS and P300 Low-level Continuous 2 (40) 2012 P300 and EEG artifact The user can choose low or high-level Discrete 4 (41) 2012 SSVEP Low-level Continuous 2 (42) 2013 EEG artifact High-level Discrete 1 (43) 2013 ERD/ERS Low-level Discrete 1 (44) 2013 ERD/ERS Low-level Continuous 3 (45) 2013 ERD/ERS and EMG Low-level Continuous 1 (46) 2013 ERD/ERS Shared Continuous 4 (47) 2013 P300 Shared Discrete 11c (48) 2013 SSVEP Low-level Discrete 1 (49) 2013 P300 and SSVEP Low-level Continuous 5 (50) 2013 SSVEP Low-level Continuous 13d (51) 2014 Alpha band Low-level Discrete 8 (52) 2014 ERD/ERS, EOG and P300 Low-level Continuous 4 (53) 2014 ERD/ERS and SSVEP Low-level Mixedb 3 (54) 2014 ERD/ERS and SSVEP Low-level Continuous 3 (55) 2014 SSVEP Shared Continuous 4 (56) 2014 EOG (embedded in EEG) Low-level Continuous 5 (57) 2015 ERD/ERS Low-level Discrete 3 (58) 2015 SSVEP High-level Discrete 37 (59)a 2016 ERD/ERS High-level Discrete 3 (59)b 2016 P300 High-level Discrete 6 a affected by Guillain-Barre Syndrome b discrete turns and continuous advance and recoil c 1 participant with cerebral palsy and motor impairment d 1 paraplegic participant 2.1.3 Task and interface. The average of tasks and commands used was 2.67 (σ = 0.71) and 5.77 (σ = 8.06), respectively. So the CTR was 2.11 (σ = 2.67). It must be taken into account that the BCW of the study of Varona-Moya et al. (57) Zhang et al. (59) was the only ones with a CTR greater than the unit. The rest of the studies have a CTR equal to the unit, where every task served to execute a single command. The specific tasks to be performed by the user were fairly homogeneous across studies. Most of the papers (80%) used hand MI – left, right, or both – as one of its tasks. It was also common to use feet MI or to maintain a state of rest. On the other hand, the most common commands were to move forward, and to rotate the chair to the left and right. Just one of the BCWs used a graphical user interface (GUI), the proposal of Zhang et al. (59) that was the one with the largest number of commands through the use of a successive dichotomy method. By other side, other article includes the presence of an audio interface, indicating different commands serially that users could select through righthand MI (57). Hence, the only two proposal based purely on ERD/ERS signal whose CTR outperformed the unit were those that needed a specific graphical or auditory serial interface. 2.1.4 Feature extraction and classification methods. The feature extraction methods used were quite heterogeneous; however, the power spectral density (PSD) can be highlighted as being the most used by the proposals with ERD/ERS 5 Table 2. User’s tasks, used commands, and command tot task ratio for the brain-controlled wheelchairs compiled in this review. Paper User's task Commands CTR ERD/ERS (18) 2; left/right thinking 2; forward in diagonal line left/right 1 (28) 3; MI left hand, word association, relax or arithmetic operation 3; forward and turn left/right 1 (34) 4; MI left/right hand and foot, and idle state 4; forward, turn left/right and stop 1 (35) 3; MI left/right hand and idle state 3; turn left/right and then forward, and stop 1 (43) 2; MI right hand and feet 2; forward and turn right 1 (44) 3; MI right/left hand and feet 3; forward and turn left/right 1 (46) 2; MI right/left hand or feet 2; turn left/right 1 (57) 2; MI right hand and idle state 7; forward, backward, turn left/right, maintain position and turn on/off the system 3.5 (59)a 3; MI right/left hand and idle state 27; 25 locations, validate and stop 9 P300 (27) 1; selective attention 9; 7 locations, an "application button" and lock 9 (29) 1; selective attention 18; 15 locations, turn left/right and validate selection 18 (32) 1; selective attention 15; 6 for the BCW (not specified) and 9 for the robotic arm 15 (33) 1; selective attention 4; forward, backward and turn left/right 4 (47) 1; selective attention 7; forward, backward, turn left/right 45º or 90º and stop 7 (59)b 1; selective attention 41; 37 locations, validate or delete selection, stop and show extra locations 41 SSVEP (30) 1; selective attention 4; forward, backward and turn left/right 4 (41) 1; selective attention 5; forward, turn left/right and turn on/off the system 5 (48) 1; selective attention 4; forward, turn left/right and stop 4 (50) 1; selective attention 4; forward, backward and turn left/right 4 (55) 1; selective attention 5; forward, backward, turn left/right and stop 5 (58) 1; selective attention 5; 4 locations and a "return to the previous window" command 5 HYBRIDMENTAL (31) 2; MI hand fingers tapping or MI walking and making left/right turns, and selective attention 10; 9 locations and stop 5 (39) 4; MI right/left hand and feet, and selective attention 4; accelerate, decelerate and turn left/right 1 (49) 1; selective attention 4; forward, stop and turn on/off the system 4 (53) 3; MI right/left hand movement and selective attention 8; forward, turn left/right, accelerate, decelerate, maintain an uniform velocity and turn on/off the system 2.67 (54) 3; MI right/left hand movement and selective attention 4; accelerate, decelerate and turn left/right 1.33 MUSCLEASSISTED (36) 4; MI right/left hand and feet, and cheek movement 4; forward, turn left/right and stop 1 (37) 2; left/righ hand movements 2; turn left/right 1 (38) 3; attention, idle state and eye-blinking 15; 13 directions, forward and stop 5 (40) 4; selective attention and 2-4 eye-blinkings 8; 4 locations, forward, backward and turn left/right 2 (42) 4; raise eyebrows, teeth clench on the left/right side and both 4; forward, backward and turn left/right 1 (45) 4; MI right/left hand and teeth clench on the left/right side 4; forward, turn left/right and stop 1 (51) 2; close the eyes and keep them open 4; forward, backward and turn left/right 2 (52) 4; MI right/left hand, selective attention and eye-blinking 8; forward, backward, turn left/right, stop, accelerate, decelerate and maintain position 2 (56) 3; blink twice, close the eyes and keep them open in six different directions 8; forward and backward in three directions, validate and stop 2.67 6 Table 3. Feature extraction and classification methods used for the brain-controlled wheelchairs compiled in this review. Paper Feature extraction Classifier ERD/ERS (18) FFT Recursive training algorithm for pattern recognition (28) PSD Gaussian (34) Logarithmic values of six band power components Recurrent neural network (35) Logarithmic band power LDA (43) Learning vector quantization in µ and β bands LDA (44) CSP SVM (46) PSD Gaussian (57) PSD LDA (59)a CSP SVM P300 (27) Raw signal SVM (29) Moving average technique LDA (32) Signal averaging Linear classifier (33) Signal averaging SVM (47) Optimal statistical spatial filter Binary Bayesian (59)b Signal averaging SVM SSVEP (30) Frequency band power Threshold method not specified (41) CCA Bayesian (48) Frequency peaks Decision tree method (50) PSD Statistical maximum (55) FFT and CCA CCA coefficient (58) Amplitude of the fundamental frequency Threshold method not specified HYBRID-MENTAL (31) Raw signal SVM (39) One versus the rest CSP (ERD/ERS) and band-pass filter 0.1-20 Hz (P300) LDA (49) Statistical average (P300) and the minimum energy combination (SSVEP) SVM (53) CSP (ERD/ERS) and canonical correlation analysis (SSVEP) Radial basis function kernel SVM (ERD/ERS) and the canonical correlation coefficient (SSVEP) (54) CSP (ERD/ERS signal) and canonical correlation analysis (SSVEP) SVM MUSCLEASSISTED (36) PCA and a modified type of CSP (ERD/ERS) and signal averaging (EMG) SVM (ERD/ERS) and a threshold method (EMG) (37) Not specified LDA (38) Raw signal Not specified (40) Magnitude summing Maximum detection (42) Integral of energy in different bands Linear classifier not specified (45) CSP LDA (51) Signal averaging Threshold method detailed in the paper (52) One versus the rest CSP (ERD/ERS), CSP (P300) and a band-pass filter for the eye signal (0.1-15 Hz). SVM and CCA with a thresholding for the eye-blinking (56) Hidden Markov model SVM Note: when an extraction feature or classification method only affects to a specific signal, it will be indicated between parentheses. 7 Table 4. Used metrics for the evaluation of the BCW. Paper Evaluation ERD/ERS (18) Success rate (28) Success rate (34) Time required, path length, percentage of hits (35) Time required, used commands and collisions (43) Percentage of hits (44) Percentage of hits (in a standard way and in 4 seconds time windows) (46) Path length and time required (57) Time required, time optimality rate, used commands, incorrect selections, corrective actions, extra actions and false negatives (59)a Concentration time, incorrect selections, response time to stop, success rate, error distance of the stop area and false activation rate P300 (27) Error rate, selection time and false acceptance rate (29) Task success, path length, time required, path optimality rate, time optimality rate, errors, collisions, used commands, errors caused by a misunderstanding of the interface, obstacle clearance, number of missions, workload, learnability and confidence (32) Selection time (33) Time required (47) Task success, path length, time, path length optimality ratio, time optimality ratio, collisions and success rate (59)b Concentration time, incorrect selections, response time to stop, success rate, error distance of the stop area and false activation rate SSVEP (30) Success rate, best time required and used commands (41) Time required (48) Qualitative evaluation (50) ITR, positive predictive value (PPV) and usability measures (55) Unrecognized rate, path length, time required (58) Mission: get to reach 4 destinations HYBRIDMENTAL (31) Selection time and false positives (39) Path length, path optimality rate, time required, classification accuracy, wrong speed control time and collisions (49) Missions: to send a "go" command and keep the chair in place, both tasks for 30 seconds (53) Time required, useful and useless commands of switch control, selection time and collisions (54) Time required and frequency of use of the auxiliary button (to manually avoid collisions) MUSCLEASSISTED (36) Percentage of hits (37) Qualitative evaluation (38) Time required (40) Success rate, time required and transfer rate (commands per minute) (42) Task success, path length, time, used commands, collisions and obstacle clearance (minimum and average distance to the obstacles) (45) Time optimality rate (51) Success rate and error rate (specified in false positives and false negatives) (52) Task success, path length, time required, path length optimality ratio and time optimality ratio (56) Task success, path length, time required, path length optimality ratio and time optimality ratio, collisions, mean velocity, workload, learnability, confidence and difficulty 8 signal (28,46,57). Other papers used methods such as learning vector quantization in mu and beta bands (43), the logarithmic value in the bands of interest (34,35), the common spatial patterns (CSP) (44,59) or the fast Fourier transform (FFT) (18). Referring to the classification method, the most used was the followed by support vector machines (SVM) (44,59), Gaussian classifier (28,46), artificial neural networks (34) and a recursive training algorithm for pattern recognition (18). 1.1.1 Evaluation. ERD/ERS systems usually have low-level navigation or shared control, so that the evaluation of a BCW in real environments is a complex issue, since in most tests the users are asked to go from point A to point B, but they are not given the specific commands to reach the destination (figure 3). Therefore, it is difficult to know for sure which commands have been selected or rejected at will (true positives and true negatives, respectively). The metrics most commonly used were success rate, path length, time required, path length optimality ratio, time optimality ratio, number of used commands, and number of collisions, as well as other less common metrics, such as obstacle clearance and command selection time. It is worth mentioning two metrics used by Li and Liang (44), and Varona-Moya (57), which do not take into account the number of hits (or errors) regardless of the runtime, but depend on the user’s ability to select each of the commands in a given time, thereby inferring the presence of false negatives if no command is selected. In the proposal of Li and Liang (44), they used a success rate in 4 seconds time window, in which they assumed the intention of generating a command, counting as a mistake not selecting any command. Whereas in the paper of Varona-Moya (57) it was used a metric called missed opportunities, in which false negatives were collected; i.e. cases where the user did not select the optimal command to help them to efficiently complete the course. By other side, in contrast to other BCWs, the proposal of Zhang et al. (59) had a high-level control, so their metrics were a bit different from the rest, may be measured more clearly variables as the success rate or useful in this control as the time to make a selection. Figure 3. Participant of Varona-Moya et al. (57) during the execution of the path. 1.2 P300 BCWs P300 is a positive deflection in the voltage of the EEG signal, generally registered from the parietal lobe of the cortex, with a delay of about 300 milliseconds after the presentation of an uncommon target stimulus using an oddball paradigm (2). This paradigm allows the use of a matrix of numerous stimuli, which are selected by visual fixation to execute the command with which they are associated. The main advantages of these systems are: i) they do not require extensive training for management, only a small calibration to adjust the system settings for each user system (7); ii) they tend to have high success rates and iii) high number of available commands, due to the large number of stimuli that these systems allow (61,62). However, a P300 system usually has a low ITR (7) and some studies have highlighted that performance may be reduced in the long term, as the P300 wave amplitude produced by the rare stimulus decreases due to habituation effects (63). 1.2.1 Navigation. Only one P300 BCW used a lowlevel navigation system (33), while three of them used high-level (31,32,59) and two used shared control (29,47). On the one hand, a highlevel system allows the selection of the destination to which the BCW will go autonomously, so the P300 is a great candidate to serve as a communication signal, due to its high success rate, and the possibility of offering many destinations in a GUI simultaneously. On the other hand, the only BCW with a low-level navigation system, which was fairly similar to the low-level navigation systems with ERD/ERS signal, based its management on four navigation commands (forward, backward, turn left and turn right), selected through four stimuli in a GUI. The disadvantage of these systems is that they did not allow continuous control of the mobility. By other side, while the shared control proposal of Lopes et al. (47) was quite similar to a low-level navigation system (figure 4), the wheelchair showed by Iturrate et al. (29) had one of the most innovative interfaces, which conducted a mapping of the environment, in which each point was represented in the GUI by a stimulus that could be selected by the user as destination, thus guiding the BCW to it autonomously. With this system the user gains the flexibility of a low-level navigation in close displacement, with the comfort of a high-level system. 9 Figure 4. Participant of Lopes et al. (47) in test scenario. 1.2.2 Participants. The average number of participants was 4.83 (σ = 3.71). Two papers had at least one disabled user: a patient affected by Guillain-Barre Syndrome (32), and one with cerebral palsy who was severely motorimpaired (47). This latest contribution had the highest number of total users for testing (N = 11) in P300 BCWs. 1.2.3 Task and interface. P300 interfaces proved very homogeneous, as all BCWs had a GUI to offer their stimuli. The average was 15.5 stimuli (σ = 13.07). The only task for the user consists of selective attention to certain stimulus that represents the desired command to be selected and executed by the system. That is, although the number of stimuli may be high, the required task is the same. The average number of commands presented was 15.67 (σ = 13.44) so, because the only task used in all studies for selection was the selective attention to the stimulus, the obtained CTR is equal to the number of stimuli (i.e. CTR = 15.67; σ = 13.44). 1.2.4 Feature extraction and classification methods. For P300 signal, the extraction methods were more heterogeneous than the previous ERD/ERS signal. Three proposals used the signal averaging technique (32,33,59), and the others used the moving average (29) and a statistical spatial filter (47). On the other hand, only one BCW used raw signal (27). Referring to the classification methods, the SVM was used by half of the P300’s proposals (27,33,59) while the other systems used LDA (29), a binary Bayesian classifier (47) and a not specified linear classifier (32). 1.2.5 Evaluation. Unlike what happened in the BCWs with ERD/ERS signal, in P300 systems the most prevalent navigation was high-level. However, the metrics used in the various articles were very similar, regardless of the navigation system. The most used metrics were: command selection time, hit rate and time required. The paper of Iturrate et al. (29) has the largest number of metrics, which has been referenced many times. This paper used the metrics proposed in Montesano et al. (64) to evaluate the performance of an autonomous system, and had a total of 22 metrics evaluating various factors such as overall performance, command selection, GUI, navigation system, and cognitive variables of the user. Also, in this article, they used two routes: one in which they tested the performance of the BCW on curved paths, and another for straight and long paths. 1.3 SSVEP BCWs SSVEP signal are cerebral activity modulations produced in the visual cortex by blinking stimulus visualization, at a frequency higher than 4 Hz (11), eliciting a larger intensity in 5-20 Hz interval, where most SSVEP BCIs work (62). The main advantages of these systems are: i) high ITR (11); ii) they require short training (65); and iii) they allow an adequate number of commands with good accuracy rates (66). In contrast, some disadvantages are: i) it can provoke fatigue after long-term use; ii) it requires some control of the eye muscles (1); and iii) it can cause epileptic seizures (1,67). The SSVEP signal has been the last kind of signal to be used in a BCW, since it started to be tested with users in 2009 (30). 1.3.1 Navigation. There are three papers about BCWs with low-level navigation (41,48,50), one with high-level navigation (58), and two with shared control (30,55). Although the SSVEP BCWs are based on the selection of a visual stimulus, such as the P300-BCW, we found a clear predominance of prototypical low-level navigation and shared-control systems with four or five commands. This fact could be related to a smaller number of allowed targets in the case of a SSVEP interface, compared to a P300 system. An SSVEP-based BCI can detect how much time users keep their attention on certain stimuli, allowing to maintain the movement as long as users desire; i.e., continuous control (e.g. (41,50,55)). Otherwise, the high-level navigation BCW was similar to the high-level navigation of P300-BCW, but with a considerable smaller amount of commands. 1.3.2 Participants. The participant average in SSVEP BCW papers was 5.33 (σ = 4.68). The study of Ng (58) was excluded to calculate the mean, because its number of participants was not representative of the rest (N = 37). By other side, the paper of Diez et al. (50) should be highlighted because it was the only one to include at least one user with MND (a paraplegic patient, with severe paralysis of upper limbs, due to a lesion at the fifth cervical vertebra), and a considerable number of healthy users, with a total of 13 participants. 1.3.3 Task and interface. The number of stimuli is limited with this type of signal, being the average 4.14 (σ = 0.38). Most SSVEP-BCWs possessed a low-level navigation system, so the number of commands was relatively low, with an average of 4.67 (σ = 0.52), and was generally reduced to five: forward, backward, stop, turn right, and turn left. As was the case in P300 16 mapping for safe and comfortable navigation of a brain-controlled wheelchair. Lect Notes Comput Sci (including Subser Lect Notes Artif Intell Lect Notes Bioinformatics). 2013;8102 LNAI(PART 1):307–17. 43. Carra M, Balbinot A. Evaluation of sensorimotor rhythms to control a wheelchair. In: Biosignals and Biorobotics Conference (BRC), 2013 ISSNIP. 2013. p. 1–4. 44. Li J, Liang J. Design of assistive wheelchair system directly steered by human thoughts. Int J Neural Syst. 2013;23(3):1–12. 45. Li Z, Lei S, Su C-Y, Li G. Hybrid brain/muscleactuated control of an intelligent wheelchair. In: Robotics and Biomimetics (ROBIO), 2013 IEEE International Conference on. 2013. p. 19–25. 46. Carlson T, Millán JDR. Brain-controlled wheelchairs: A robotic architecture. IEEE Robot Autom Mag. 2013;20(1):65–73. 47. Lopes AC, Pires G, Nunes U. Assisted navigation for a brain-actuated intelligent wheelchair. Rob Auton Syst [Internet]. Elsevier B.V.; 2013;61(3):245–58. Available from: http://dx.doi.org/10.1016/j.robot.2012.11.002 48. Müller SMT, Bastos TF, Filho MS. Proposal of a SSVEP-BCI to Command a Robotic Wheelchair. J Control Autom Electr Syst [Internet]. 2013;24(1):97– 105. Available from: http://dx.doi.org/10.1007/s40313-013-0002-9 49. Li Y, Pan J, Wang F, Yu Z. A hybrid BCI system combining P300 and SSVEP and its application to wheelchair control. IEEE Trans Biomed Eng. 2013;60(11):3156–66. 50. Diez PF, Müller SMT, Mut VA, Laciar E, Avila E, Bastos-Filho TF, Sarcinelli-Filho MM. Commanding a robotic wheelchair with a high-frequency steadystate visual evoked potential based brain-computer interface. Med Eng Phys [Internet]. Institute of Physics and Engineering in Medicine; 2013;35(8):1155–64. Available from: http://dx.doi.org/10.1016/j.medengphy.2012.12.005 51. Ming D, Fu L, Chen L, Tang J, Qi H, Zhao X, et al. Electric wheelchair control system using braincomputer interface based on alpha-wave blocking. Trans Tianjin Univ [Internet]. 2014;20(5):358–63. Available from: http://dx.doi.org/10.1007/s12209014-2235-5 52. Wang H, Li Y, Long J, Yu T, Gu Z. An asynchronous wheelchair control by hybrid EEG-EOG braincomputer interface. Cogn Neurodyn. 2014;399–409. 53. Cao L, Li J, Ji H, Jiang C. A hybrid brain computer interface system based on the neurophysiological protocol and brain-actuated switch for wheelchair control. J Neurosci Methods [Internet]. Elsevier B.V.; 2014;229:33–43. Available from: http://dx.doi.org/10.1016/j.jneumeth.2014.03.011 54. Li J, Ji H, Cao L, Zang D, Gu R, Xia B, et al. Evaluation and Application of a Hybrid Brain Computer Interface for Real Wheelchair Parallel Control With Multi-Degree of Freedom. Int J Neural Syst [Internet]. 2014;24(04):1450014. Available from: http://www.worldscientific.com/doi/abs/10.1142/S0 129065714500142 55. Duan J, Li Z, Yang C, Xu P. Shared control of a brain-actuated intelligent wheelchair. In: Intelligent Control and Automation (WCICA), 2014 11th World Congress on. 2014. p. 341–6. 56. Aziz F, Arof H, Mokhtar N, Mubin M. HMM based automated wheelchair navigation using EOG traces in EEG. J Neural Eng [Internet]. IOP Publishing; 2014;11(5):056018. Available from: http://www.scopus.com/inward/record.url?eid=2s2.0-84907908970&partnerID=tZOtx3y1 57. Varona-Moya S, Velasco-Álvarez F, Sancha-Ros S, Fernández-Rodríguez Á, Blanca MJ, Ron-Angevin R. Wheelchair Navigation with an Audio-cued , TwoClass Motor Imagery-based Brain – Computer Interface System . 7th International IEEE/EMBS Conference on Neural Engineering (NER). 2015;22– 4. 58. Ng DW-K, Soh Y-W, Goh S-Y. Development of an Autonomous BCI Wheelchair. 2014 IEEE Symp Comput Intell Brain Comput Interfaces [Internet]. 2014;1–4. Available from: http://www.scopus.com/inward/record.url?eid=2s2.0-84923042171&partnerID=tZOtx3y1 59. Zhang R, Li Y, Yan Y, Zhang H, Wu S, Yu T, et al. Control of a Wheelchair in an Indoor Environment Based on a Brain-Computer Interface and Automated Navigation. Neural Syst Rehabil Eng IEEE Trans [Internet]. 2015;PP(99):1. Available from: http://ieeexplore.ieee.org/lpdocs/epic03/wrapper.ht m?arnumber=7115145 60. Fazel-Rezai R, Allison BZ, Guger C, Sellers EW, Kleih SC, Kübler A. P300 brain computer interface: current challenges and emerging trends. Front Neuroeng [Internet]. 2012;5(July):14. Available from: http://www.pubmedcentral.nih.gov/articlerender.fcgi ?artid=3398470&tool=pmcentrez&rendertype=abstr act 61. Townsend G, LaPallo BK, Boulay CB, Krusienski DJ, Frye GE, Hauser CK, et al. A novel P300-based brain-computer interface stimulus presentation paradigm: Moving beyond rows and columns. Clin Neurophysiol [Internet]. Elsevier; 2010;121(7):1109–20. Available from: http://dx.doi.org/10.1016/j.clinph.2010.01.030 62. Wang Y, Wang R, Gao X, Hong B, Gao S. A practical VEP-based brain-computer interface. IEEE Trans Neural Syst Rehabil Eng [Internet]. 2006;14(2):234– 9. Available from: 17 http://www.ncbi.nlm.nih.gov/pubmed/16792302 63. Ravden D, Polich J. On P300 measurement stability: habituation, intra-trial block variation, and ultradian rhythms. Biol Psychol [Internet]. 1999 Oct [cited 2016 Apr 29];51(1):59–76. Available from: http://www.sciencedirect.com/science/article/pii/S03 01051199000150 64. Montesano L, Minguez J, Alcubierre JM, Montano L. Towards the adaptation of a robotic wheelchair for cognitive disabled children. Intell Robot Syst 2006 IEEE/RSJ Int Conf. 2006;710–6. 65. Gao X, Xu D, Cheng M, Gao S. A BCI-based environmental controller for the motion-disabled. IEEE Trans Neural Syst Rehabil Eng. 2003;11(2):137–40. 66. Bin G, Gao X, Yan Z, Hong B, Gao S. An online multi-channel SSVEP-based brain–computer interface using a canonical correlation analysis method. J Neural Eng [Internet]. 2009;6(4):46002. Available from: http://stacks.iop.org/17412552/6/i=4/a=046002 67. Fisher RS, Harding G, Erba G, Barkley GL, Wilkins A. Photicand pattern-induced seizures: A review for the epilepsy foundation of america working group. Epilepsia. 2005;46(9):1426–41. 68. Diez PF, Mut V a, Avila Perona EM, Laciar Leber E. Asynchronous BCI control using high-frequency SSVEP. J Neuroeng Rehabil [Internet]. BioMed Central Ltd; 2011;8(1):39. Available from: http://www.jneuroengrehab.com/content/8/1/39 69. Takeda S, Yamada M, Kawasaki K, Oyanagi K, Ikuta F, Arai M, et al. Motor neuron disease with multisystem involvement presenting as tetraparesis, ophthalmoplegia and sensori-autonomic dysfunction. Acta Neuropathol [Internet]. 1994;88(3):193–200. Available from: http://dx.doi.org/10.1007/BF00293393 70. McMenamin BW, Shackman AJ, Greischar LL, Davidson RJ. Electromyogenic Artifacts and Electroencephalographic Inferences Revisited. Neuroimage [Internet]. 2011 Jan 1;54(1):4–9. Available from: http://www.ncbi.nlm.nih.gov/pmc/articles/PMC296 2711/ 71. Lotte F, Congedo M, Lécuyer A, Lamarche F, Arnaldi B. A review of classification algorithms for EEG-based brain-computer interfaces. Journal of Neural Engineering. 2007;4(2) 72. Holz EM, Höhne J, Staiger-Sälzer P, Tangermann M, Kübler A. Brain-computer interface controlled gaming: Evaluation of usability by severely motor restricted end-users. Artif Intell Med [Internet]. Elsevier; 2016 Apr 30;59(2):111–20. Available from: http://dx.doi.org/10.1016/j.artmed.2013.08.001