These are measured as a difference in voltage between two different cerebral locations over time. The signal is recorded by EEG electrodes positioned on the scalp. The recorded potential difference is reflected as the postsynaptic potential in the cell membranes of cortical neurons (Olejniczak, 2006; Nguyen and Hong, 2013). These signals are most effective for BCI, as they are detected immediately (e.g., P300 signals are detected 300 ms after stimuli are given). These signals also contribute in the detection of brain drowsiness state (Qian et al., 2016, 2017).

The hemodynamic response is a process in which the blood releases glucose to active neurons at a greater rate than inactive ones. The glucose with oxygen delivered through the blood stream results in a surplus of HbO in the veins of the active area as well as a distinguishable change of the ratio of local HbO to HbR. These changes are detected by functional magnetic resonance imaging (fMRI) and fNIRS (Boas et al., 1994, 2001, 2014; Huppert et al., 2009, 2013; Nicolas-Alonso and Gomez-Gil, 2012). These signals have an inherent delay in hemodynamic response generation. However, the most recent discovery on initial dips makes the HbO signals a viable candidate for BCI (Hong and Naseer, 2016; Zafar and Hong, 2017).

The eye can be modeled as a dipole, with its positive pole at the cornea and its negative pole at the retina. Assuming a stable corneo-retinal potential difference, the eye is the origin of a steady electric potential field. The electrical signals generated from this field are measured by EOG (Bulling et al., 2011). Sometimes an eye tracker also is used for the detection of eye movements. Mostly, these signals are used for the investigation of vigilance and drowsiness activities.

These signals are an indication of muscles’ electrical activity, which arises whenever there exists a voluntary or involuntary contraction (Chowdhury et al., 2013; Patil and Patil, 2014; Xie et al., 2014). These signals are recorded by EMG electrodes, which are most widely used in neuro-prostheses (Ravindra and Castellini, 2014; Chadwell et al., 2016; Chen et al., 2016). Table 1 summarizes possible combinations that can be used in the development of hBCI hardware.

EOG-based BCIs are useful for people who have control over their eye movements, as by this means, multiple commands can simultaneously be generated. Combinations of eye movement signals (blink, wink, frown, etc.) with neuronal signals usually are utilized for hybrid EEG–EOG-based BCIs (Ma et al., 2015). In this section, we also include hybrid studies that have used eye-tracking with EEG to develop hybrid systems for BCI. We discuss EOG and eye tracker-based studies together, as both use eye movements for classification. For command generation, signals are decoded simultaneously, and for control of a BCI system, they are fused using a combined classifier (Jiang et al., 2014). Although EOG is used to remove ocular artifacts from EEG data (Li et al., 2015), drowsiness detection (Khushaba et al., 2011) and wheelchair control (Ramli et al., 2015) are also among the most common applications of EEG–EOG-based systems. Figure 3 shows the method used to acquire simultaneous EEG and EOG data for analysis.

Electromyography signals are generated and detected as a result of muscular movement (Trejo et al., 2003; Foldes and Taylor, 2010; Cler and Stepp, 2015). These act as an artifact in EEG signals, resulting in the false detection of brain signals (Fatourechi et al., 2007; Bhattacharyya et al., 2013). The purpose behind a hybrid EEG–EMG-based hBCI is to combine EEG and EMG signals in hBCI. This incorporation of EMG signals is user specific and depends on the activity or task performed by that user. The applications of hybrid approaches vary from a simple game control application for an able-bodied person through to a prosthetic arm control application for an amputee. Figure 4 shows a typical strategy used for incorporating EEG and EMG signals into an hBCI system.

The applications of EEG–EMG-based hBCI are found in the control area of assistive devices (Leeb et al., 2011; Kiguchi et al., 2013). In the early work using EEG with EOG and EMG (Kennedy and Adams, 2003), the EMG signals were used to categorize different “locked-in” patient types. In their study, six types were defined, the first three of which were categorized using EMG as follows:

Patients capable of movement (e.g., eye movement and finger movement).

The remaining three types of patients were categorized using EOG and EEG signals. For EEG–EMG-based BCI, a neuro-electric interface was developed for real-time applications (Trejo et al., 2003). In 2005, a BCI that removes EMG artifacts from EEG for mouse-cursor control was developed (McFarland et al., 2005). In 2007, a detailed survey on EMG artifacts in EEG signals was presented (Fatourechi et al., 2007). In 2010, a study by Brumberg et al. (2010) combined EEG and EMG for tetraplegic patients: their results showed that communication, however slow, can be achieved using EEG–EMG-based hBCI. In that same year, jaw muscle contraction and EEG signals were used to generate commands for an assistive neuro-prosthetic device (Foldes and Taylor, 2010). Also, the use of EMG with EEG was explored in a review article on the operation of robotic and prosthetic devices (McFarland and Wolpaw, 2010). In 2011, an investigation was conducted for the prediction of voluntary movements before their occurrence using hBCI (Bai et al., 2011). Vehicle steering (Gomez-Gil et al., 2011) and determination of muscle fatigue levels (Leeb et al., 2011) using EEG–EMG-based BCI have also been reported in the literature. In 2012, simultaneous measurement of EEG–EMG signals led to the achievement of an assistive control of exoskeletons for locomotion (Cheron et al., 2012), wherein a surface-tactile stimulation device was used for the training of brain signals, and dynamic recurrent neural network-based classifiers were used for training and testing of brain signals. Single-trial decoding of reaching movement also was conducted using EEG–EMG-based signals (Demandt et al., 2012). A similar study (Kiguchi et al., 2013) in 2013 proposed EEG–EMG-based control of an artificial arm for above-elbow amputees. An EEG–EMG-based motion estimation method was proposed for the control of the forearm and the supination/pronation motion of the artificial arm. In 2014, signals produced by jaw clenching were removed from EEG signals for two-dimensional cursor control on a computer screen (Costa et al., 2014). This study was later extended to the control of a robotic arm in bi-dimensional workspace. In 2015, EMG was used in rehabilitation applications for robot-assisted exercise tasks (Fels et al., 2015). In this case, neuro-feedback was used for intensive motor training and EEG–EMG was employed to predict the workload profiles for the experience of frustration. A review article by Rupp et al. (2015) also discussed the application of EMG for a hybrid neuro-prosthesis, proposing the use of functional electrical stimulation (FES) for therapy. The scheme used EEG–EMG to record brain activity and to investigate, on that basis, recovery in muscles and the brain. Most recently, new work has been done, which combines SSVEP-based tasks with EMG for choice selection (Lin et al., 2016). A 60-target hybrid BCI speller was built in this study. A single trial was divided into the following two stages: a stimulation stage and an output-selection stage. In the stimulation stage, SSVEP and EMG were used together. Every stimulus flickered at its given frequency to elicit SSVEP. CCA and mean filtering were used to classify SSVEP and EMG, respectively. In the result, 81% of accurate results were obtained by hybridizing EMG with SSVEP activities.

The research completed on hybrid EEG–fNIRS is still very limited. This technology is used mostly to improve classification accuracy (Fazli et al., 2012) or increase the number of control commands (Khan et al., 2014) in a BCI system. Although the research has shown good results for the combination of fNIRS with bio-signals (Zimmermann et al., 2013), hybrid EEG–NIRS has shown the best results thus far for BCI. In this case, two brain signal acquisition modalities are combined using neuronal signals (recorded using EEG) and hemodynamic signals (recorded using NIRS). One important disadvantage of the use of hemodynamics (either fMRI or fNIRS), however, is the inherent delay in the response (Huppert et al., 2013), which renders the generation of commands slow in comparison to EEG. However, in the case of combined EEG–fNIRS, this kind of disadvantage can be removed. Also, the detection of initial dip (i.e., the phenomenon that HbO decreases and HbR increases with neural firing) instead of hemodynamics might lead to a better time window selection for the combined modalities. Figure 5 shows an approach used to combine the EEG–NIRS modalities for BCI.

The first study on hybrid EEG–NIRS for application to BCI appeared in 2012 (Fazli et al., 2012). It showed that the combination of fNIRS’s features (HbO and HbR) and EEG features increases the classification accuracy. In this case, a multi-class classifier that combined the NIRS and EEG features for classification was used. The results showed that using EEG + HbR features, the average classification accuracy for motor execution tasks was improved from 90.8 to 93.2%. Similarly, for MI tasks, the average classification accuracy using EEG + HbO features was increased from 78.2 to 83.2%.

In 2013, a study of Safaie et al. (2013) analyzed the steps involved in the development of hardware for a hybrid EEG–NIRS system, among which was the design of a wireless wearable module for simultaneous decoding of brain activity. In 2014, an optimal time window for hybrid EEG–NIRS features selection was investigated using a SSVEP-based paradigm (Tomita et al., 2014). The results showed that the optimal window for EEG and fNIRS is 10 s. Another study by Khan et al. (2014) showed that the number of control commands can be increased by simultaneously decoding the EEG and fNIRS activities from different brain locations, in which LDA was used as a classifier for both EEG and fNIRS, MA and mental-counting tasks were decoded using fNIRS, and left- and right-hand tapping were coded using EEG. An offline study on tetraplegia patients (Blokland et al., 2014) showed that combined EEG–NIRS can be used to decode motor attempts and imagined movement with high accuracies up to 87 and 79%, respectively. A logistic regression classifier based on the Bayesian theorem was used for classification, specifically by obtaining the EEG data in a 0–15 s window and the fNIRS data for classification in a 3–18 s window. The study showed that the highest accuracy for tetraplegia patients was obtained by combining EEG with HbR. Kaiser et al. (2014) investigated the effect of training on cortical brain areas using EEG–NIRS with MI as a task. The study compared the subjects with high BCI performance (accuracy > 70%) with those with low BCI performance (accuracy < 70%), employing LDA for acquisition of the classification accuracies. The results showed that training with MI-based BCI affects cortical activations, especially with those subjects showing a low BCI performance. In another work, hybrid EEG–NIRS showed higher classification accuracies in discriminating auditory and visual stimuli (Putze et al., 2014). In this work, SVM was used as a classifier to discriminate the visual and audio cues and, thus, to develop an hBCI; the accuracy achieved was as high as 94.6%.

In 2015, a study used threshold-based discrimination for fNIRS signals and SVM-based classification for EEG signals to achieve 88% accurate results for self-paced MI tasks (Koo et al., 2015). In another work, a sensory motor rhythm-based paradigm was used to investigate the superiority of multi-modality for idle state detection (Lee et al., 2015). The LDA-based classification (achieving a 3.6% increase in the accuracy for MI tasks using the hybrid modality) showed that the NIRS signals contributed to the detection of the active/idle state as well as to the detection of active classes to confirm early activity detection. In two other similar studies, the MI of both the force and speed of hand clenching was decoded using hybrid EEG–NIRS (Yin et al., 2015b,c). In the first case, the extreme learning machine classifier was used to decode the responses associated with the force and speed imagery of the hand with an accuracy of 76.7%, whereas, for the second case (Yin et al., 2015c), the features of EEG and NIRS were combined and optimized using the joint mutual information selection criterion, again utilizing the extreme learning machines, in which case, the resulting average classification accuracy for the force and speed of hand clenching was 89%.

Several studies on the applications of hybrid EEG–NIRS have emerged in 2016, showing the trend of growing research in this area. The studies related to BCI applications were also discussed. In the case of active tasks, four motor tasks, namely right- and left-arm movement and right- and left-hand movement tasks, were decoded using the hybrid EEG–NIRS system (Buccino et al., 2016). Employing the LDA-based classification, the features using common spatial patterns (CSP) were compared with the signal mean (SM) and signal slope (SS). In the rest-task classification, the SM–SS average accuracy was 94.2% and the CSP average accuracy was 86.2%. The SM and SS, meanwhile, also performed better for right–left classification. SM–SS achieved an average accuracy of 72.2%, whereas with CSP, only 67.1% was possible. In the case of arm-hand classification though, CSP showed a better performance (83.6% average accuracy) than SM–SS (79.9%). In the case of passive tasks, the neural and hemodynamic correlates were estimated to measure drivers’ mental fatigue levels (Ahn et al., 2016). An average accuracy of 75.9% was achieved using an LDA-based classifier combining EEG, fNIRS, and ECG modalities. A new hybridization concept for combined EEG and fNIRS was introduced by Keles et al. (2016). In their study, different EEG bands (α, β, θ, and Δ) were estimated for the resting state. The correlation of an EEG band was convoluted with the modeled hemodynamic response (generated using two gamma functions) to generate the expected response with the incorporated neuronal activity for the hemodynamic signal. Since this was a pioneer study, its BCI-related role is yet to be investigated.

The most recent work on hybrid EEG–NIRS-based BCI demonstrates decoding of eight commands from the prefrontal and frontal cortices (Khan and Hong, 2017). In this work, four commands were decoded using EEG, in which two commands were generated using an eye blinking-based tasks and two commands were decoded using eye movement tasks. The interesting part of this study was the decoding of fNIRS commands in a 2 s window. For the selection of an optimal window size, the difference between the baseline and the first trial was used for channel selection. The signal mean and minimum values of HbO were used to detect the brain activity in that window. Four commands were generated using MA, mental counting, word formation, and mental rotation-based tasks, respectively. An average accuracy of 86% for EEG and that of 75.6% for fNIRS were achieved using an LDA-based classifier. This study was tested for the control of a drone, whose results showed the feasibility of using prefrontal-based commands for BCI.

These reactive signals are generated from either the occipital brain area or the temporal brain area by the provision of either visual stimuli (Liu and Hong, 2017) or auditory stimuli (Santosa et al., 2014; Hong and Santosa, 2016). Although mostly such stimulations are intended for the generation of brain activity from the corresponding lobes (An et al., 2014; Tidoni et al., 2014; Wu et al., 2016; Xu et al., 2016), some audio/video stimuli are given to generate P300 signals (Rutkowski, 2016). For healthy individuals, these stimulations can be effective in generating multiple commands. However, they can also be beneficial for patients with no motor or eye movements.

These signals are detected mostly in the occipital brain region. They are generated by gazing at a stimulus, which causes an increase in neural activity in the brain. VEPs are elicited by sudden visual stimuli, the repetition of which leads to a stable voltage oscillation pattern in EEG that is known as SSVEP. The stimulus used for these signals is light flickering at different frequencies (sometimes in the “checker board” pattern with changing colors). Using SSVEP signals, multiple reactive commands can be generated. The drawback of this activity is the need for the continuous focus on flashing light, which might not be possible or an ineffective approach for some patients (Muller-Putz et al., 2005). The signal detection time for these signals has been reduced to less than 1 s using spatio-temporal features with a reduced number of channels (Zhang et al., 2013, 2014; Chang et al., 2016; Wang et al., 2016).

This signal is detected mostly from the parietal brain region. They are the ERPs that indicate the responses to specific cognitive, or sensory, or motor events. The presentation of a stimulus in an oddball paradigm can produce a positive peak in EEG signals. This peak appears 300 ms after the onset of the stimulus. The stimulus can be visual, auditory, or somatosensory. This evoked response in EEG is the P300 component of ERP. These, most widely utilized in speller applications, also can generate multiple commands for BCI. However, being reactive, these signals are mostly used only for healthy subjects (Bayliss et al., 2004; Piccione et al., 2006; Turnip et al., 2011; Turnip and Hong, 2012).

These signals are detected from the prefrontal and dorsolateral prefrontal brain regions. They are a good choice for BCI, as they require less training effort. In the case of fNIRS, they are especially suitable in that they incur fewer motion artifacts and less signal attenuation due to detector slippage in hair. Also, given their non-utilization of motor activities, they are more effective on patients with severe motor disabilities. MA, mental counting, and other tasks can be detected as active-type brain signals for BCI (Kim et al., 2012; Naseer et al., 2014, 2016a,b). The passive activity of drowsiness also can be detected from this cortex (Khan and Hong, 2015). Another research has reported the detection of music imagery, picture imagery, word generation, etc., from the prefrontal cortex (Naseer and Hong, 2015b; Ma et al., 2017). The most common task used for BCI purposes is MA.

These signals are detected mostly from the primary and central brain regions (mostly the motor cortex). They are most suitable for active BCI applications, as they are natural means of providing BCI control over external devices (Naseer and Hong, 2013). These signals are also targeted for the investigation of motor recovery or neuro-rehabilitation. They have wide applications in both EEG- and fNIRS-based BCI systems. Two different types of brain signals detected from the motor cortex are motor execution and MI. Motor execution is performed by the movement of muscles (mostly in hands or feet). Also, some eye-movement-related activities stimulate the motor region. Some BCIs use motor execution for the generation of commands, whereas in other cases, these signals are removed from EEG signals using EMG/EOG for better detection of MI. Motor imaging can be defined as a covert cognitive process of kinesthetic imagining of the movement of one’s own body without the involvement of muscular tension, contraction, or flexion (Naseer and Hong, 2015b). The MI signal is weak relative to the motor execution task. Also, not all subjects can perform this activity for BCI. In fact, since this activity is generated by the imagining of hand or foot movements, it is best suited for active-type BCI. A list of the different activities that can be used with a hybrid paradigm is provided in Table 2.

The simultaneous decoding of SSVEP and MI signals constitutes a hybrid system consisting of active (based on MI) and reactive (based on SSVEP) commands. Although only a few relevant studies have appeared, they have successfully demonstrated the significance of the hybrid paradigm for control (Horki et al., 2011; Cao et al., 2014; Li et al., 2014) and rehabilitation applications (Daly et al., 2013; Yu et al., 2015).

The combination of SSVEP and P300 signals results in a reactive hBCI. These brain activities are simultaneously recorded from the occipital and parietal brain areas. This paradigm is used repeatedly in several control and rehabilitation applications.

The MI- and P300-related tasks have been widely designed for applications in real-world environment (Li et al., 2010; Su et al., 2011; Long et al., 2012a,b; Yu et al., 2012, 2013; Bhattacharyya et al., 2014; Naito et al., 2014; Kee et al., 2015; Zhang et al., 2016a). The corresponding signals are obtained by positioning electrodes around the motor and parietal brain regions. In the work of Li et al. (2010), MI and P300 were combined for the control of a cursor on the screen using SVM as a classifier (90.75% average accuracy). In the work of Long et al. (2012b), target selection or rejection (mouse-clicking) features were added resulting in the average accuracy of 92.8%.

Navigation and target selection in a virtual environment was demonstrated in Su et al. (2011) using the hybrid MI–P300-based paradigm. In their work, MI was used for navigation and P300 for target selection (three control buttons). Overall, five commands were generated using Fisher LDA- and SVM-based classification (84.5% for MI and 81.7% for P300).

In the work of Long et al. (2012a), the same hybrid paradigm was used to control the direction and the speed of a simulated and afterward a real wheelchair. The turning (left and right directions) controls were associated with left- and right-hand imageries. The wheelchair was decelerated using a foot-movement imagery. The acceleration was represented using P300 signals. In this case, the LDA-based classification resulted in 71.6% accuracy for MI and 80.4% for P300. Another study of Yu et al. (2012) demonstrated the utility of hybrid MI–P300 for cursor-controlled on-screen feature selection-based Internet surfing. In that study, SVM-based classification resulted in an average accuracy of 93.2%.

In 2013, a real-time electronic mail communication system was implemented to enable clients/users to receive, write, and attach files to their email (Yu et al., 2013). According to their hybrid MI–P300 paradigm, the SVM-based classifier yielded a high accuracy (average: >90%) for the system.

In 2014, robot arm movement control for rehabilitation was implemented using the hybrid method (Bhattacharyya et al., 2014). Arm movement was controlled by MI signals, while P300 was used to detect the stopping intention. A 95% success rate was achieved in the SVM-based classification. The recent work on this paradigm has focused on optimal feature selection and simultaneous classification methods (Naito et al., 2014; Kee et al., 2015). Also, the use of optimal feature selection for the enhancement of output accuracy was another issue in Naito et al. (2014).

In 2015, GA-based strategy was proposed for the optimization of channel selection in the process of simultaneous recording of MI and P300 (Kee et al., 2015). A recent contribution of 2016 proposed an autonomous wheelchair navigation system that acquires the destination and waypoint based on the existing environment (Zhang et al., 2016a). In this case, the BCI module selects and determines the destination using MI and P300, respectively.

For mental VS motor task purposes, paradigms are designed to obtain the brain activities from the prefrontal and motor cortices. Both EEG and fNIRS have provided good results for these tasks. In these cases, mostly a working memory-related task is combined with a motor task to achieve the BCI system. For the localization of neuronal sources, EEG was applied using six cognitive tasks (arithmetic, navigation imagery, auditory recall, phone imagery, and MI of the left and right hands) and compared against the idle state to localize the brain location (Dyson et al., 2010). The spatial areas suggested a clear discrimination between the arithmetic- and auditory-related tasks, while the MI-related tasks were discriminated according to the baseline.

As for fNIRS, MA has been combined with music imagery for simultaneous decoding of brain activity (Power et al., 2010, 2011; Stangl et al., 2013). In this case, however, the prefrontal channels were averaged and the activities were differentiated using an HMM-based classification method. MA and MI have also been reported to be combined in the case of fNIRS. Four commands have been generated by simultaneous decoding of mental counting-, arithmetic-, and imagery-related tasks (Hong et al., 2015; Naseer and Hong, 2015a). In these cases, LDA-based classification was used to decode the activities from the prefrontal and motor cortices.

In a multi-modality case, meanwhile, four commands were generated by decoding mental tasks (MA and mental counting) using fNIRS and motor tasks (left- and right-hand tapping) using EEG for hBCI (Khan et al., 2014). This work was later extended for the decoding of eight commands using eye movement and mental tasks (Khan and Hong, 2017). In this case, EEG was employed to decode two and three eye blinks and left- and right-eye movements, whereas fNIRS was used to decode mental arithmetic-, mental counting-, word formation-, and mental rotation-based tasks. The decoded eight commands were used to operate a quadcopter in the 3D space.

In this category, there is not much hybrid research. In one work (Putze et al., 2014), hybrid EEG–NIRS was used to discriminate the auditory and visual stimuli. The details on this study can be found in the above “EEG + NIRS” section. Tidoni et al. (2014) used audio feedback to improve the performance of the BCI system. Six commands were generated using SSVEP-based tasks for a robot’s pick-up and placing tasks. It is found that audio–visual synchrony between footstep sounds and actual humanoid walking reduces the time required for steering the robot. This demonstrates the possibility of using auditory feedback congruent with humanoid actions to improve the motor decisions of the BCI. An et al. (2014) used six auditory stimuli for the selection of 36 visual symbols using EEG recording for a gaze-independent BCI. When a subject focuses on a sound, a selection screen consisting of six symbols is shown. The subject can then choose one from the six visual symbols for choice selection. This paradigm increased mental fatigue, as the user has to focus on the audio cues. An LDA-based scheme was used, and 87.7% of accurate results were generated. In another work of Barbosa et al. (2016), in contrast to An et al. (2014), a P300-based BCI was used for combining visual and audio stimuli. In this case, the audio stimuli were natural spoken words, which reduced the mental work load. The average online accuracy for their hybrid approach was 85.3%, which represents an improvement of 32% relative to independent audio or visual stimuli.

The BCI research has demonstrated the use of EEG in several communication and control applications. In most cases, MI was used to generate the commands for communication (Machado et al., 2010; Bi et al., 2013; Hwang et al., 2013; Ahn and Jun, 2015; Maria Alonso-Valerdi et al., 2015). However, it is difficult for patients to perform an MI activity. Also, the detection of MI signals requires extensive processing, and false detection can result in severe consequences in real environments. Long-term use of SSVEP and P300 can also increase visual fatigue of the subjects, thereby incurring a false detection of signals for BCI. Thus, proper measures are required to increase the accuracy of the system by minimizing the false detection rate. This can be achieved by combining multiple modalities. In such a case, simultaneous feature decoding results in a better system accuracy. The most common example is the hybrid EEG–fNIRS.

The number of active commands used for system control is a central issue for BCIs. The main problem in EEG-based BCI systems is the loss of accuracy resulting from an increased number of (active) commands. Although the number of commands can be increased using reactive tasks, it is difficult to make the patient concentrate on reactive tasks for a long duration. Although strategies are being designed to overcome this problem, the results have not yet been proved sufficiently effective for BCI adoption (Lesenfants et al., 2014). In this context, the hBCI plays an important role in providing the potential for an increased number of commands without undo-influence on classification accuracy. The approach most widely employed in the brain signal-based control of wheelchairs is to increase the number of commands by decoding the features from two modalities separately.

Tables 3–5 list all of the important studies from 2010 to 2016 that have combined two modalities to decrease false detection and enhance classification accuracy for BCI. The tables show the relevant hBCI studies for the enhancement of accuracy and increase in control commands. Ideas on increasing accuracy and the total number of commands can be deduced from them. We divided the tables into active, passive, and reactive BCI categories. This information can be helpful to the selection of brain signal acquisition modalities based on the types of activities. Also, we have incorporated the classifier and window size information in each table, which might be helpful to prospective researchers looking a method to enhance classification accuracy and increase the number of commands.

The ultimate goal of a BCI system is to provide assistance to patients (Kim et al., 2011). This assistance can be in the form of formulating a methodology that can be used to communicate with the environment. The patient should be able to express his/her thoughts through the use of the BCI system. Regarding rehabilitation, a BCI system has the capacity to distinguish improvement from non-improvement as a result of therapy and brain stimulation. Detection of seizures (Nguyen et al., 2012, 2013), epilepsy (Peng et al., 2014; Pouliot et al., 2014; Visani et al., 2015), and estimation of improvement in motor functions after stroke (Das et al., 2016) are such examples pursuing hBCI for patients. The current BCI system, however, lacks the potential to provide detailed functions. In fact, an hBCI can be a more powerful tool than a single-modality BCI, as it can provide more reliable information for control and rehabilitation applications for patients. Indeed, the recent hybrid combinations have shown successes. The hBCI can be more successful for patients in the following areas.

As per the above discussion, it seems that hBCI/BCI is more suited to patients; however, most studies have used healthy subjects for experimentation in a lab environment. This might be due to the fact that hBCI is still in its developmental phase. However, in our opinion, hBCI has several aspects that are most suited to healthy people. The following are the three major applications of hBCI.