Table 3.
Summary table of reviewed articles.
| References | Aim of study | Assessment task | Main findings |
|---|---|---|---|
| Blokland et al. (2013) | The study aimed to investigate the feasibility of using a combination of EEG-fNIRS in tetraplegia patients. Secondly, it aims to test the feasibility of motor execution instead of MI as a brain switch control task. | Subjects performed six sequential movement tasks; each task consists of six trials with visual aids. Each trial lasted for 15 s. Three different tasks were completed, i.e., “rest” (do nothing), “movement” (fingers and thumb tapping continuously), and “imagined movement” (imagine tapping of fingers and thumb continuously). | EEG-fNIRS system proved to be beneficial for users who lack sufficient control of current EEG-based brain switches. The average classification performance was 87% for the motor attempt and 79% for MI in tetraplegia patients. |
| Do et al. (2013) | The research aims to investigate the feasibility of a BCI system for lower extremity prosthesis patients. EEG, EMG, and gyroscope are used along with commercial robotic gait orthosis devices to investigate BCI's feasibility for motor disability of SCIP. | One orthosis and one healthy patient were considered in this study with 30-s idling and ten mint kinematic motor imagery tasks, respectively, and the additional walking task at a 2 km/hr speed for a healthy subject. | The offline accuracy for healthy and SCIP was 94.8 ± 0.8% and 77.8 ± 2.0% receptively. The BCI-RoGO system helped to regain brain driven basic ambulation with less training time. The results are quite convenient to use BCI based systems for rehabilitation and restoring overground walking. |
| Rea et al. (2014) | The study aimed to assess the measurement and classification of hemodynamic signals associated with lower limb motor movements for chronic stroke and its usage for future fNIRS-BCI rehabilitation applications of the lower limb. | The experimental paradigm consisted of two sessions with eleven left and right hip of movement preparation (9–11 s) followed by movement execution (left/right hip) (3 s), and then in between rest (15–25 s) performed in a pseudo-randomized order | Single-trial analysis indicated that specific hemodynamic changes associated with the left and right hip movement preparation could be measured with fNIRS with classification accuracies of 73 and 89%, respectively. These findings encourage further investigations of fNIRS suitability for BCI applications in rehabilitating patients with lower limb motor impairment after stroke. |
| Bulea et al. (2014) | In this study, EEG delta band is used to investigate the intention before movement execution to differentiate between three different classes, i.e., rising from sitting position to standing, lowering from standing position to sitting, and standing or sitting quietly. | Ten healthy adults participated in the experiment of performing three different tasks during two trials. One trail was self-paced, and the other is audio-triggered. Each trial has ten alternate sit to stand and stand to sit transition. Standing or sitting in the third task, which was random of 3–10 s. | Classification accuracy of 78% is achieved using features extracted from pre-movement epochs with no post-processing required and minimizing the classification delay. Result suggests that the primary motor cortex (M1) contains more discriminative information for standing and sitting intention when movements are self-initiated compared to cue. |
| Salazar-Varas et al. (2015) | The study focuses on the early detection of subsequent alertness after the obstacle's sudden appearance using EEG signals. This work's final application is to generate the STOP control command for exoskeleton control in obstacle detection. | Different responses regarding obstacle appearance were detected, such as reaction (stop walking after the appearance of an obstacle), delayed reaction (continuous few steps after the appearance of an obstacle), no reaction (ignore the obstacle), and free reaction (subject freely decides when to stop). A run consists of 180 s of Reaction condition, 240 s of Delayed reaction condition, 180 s of No reaction condition, and 120 s of Free reaction condition. | The results obtained for the majority of the subjects showed that polynomial coefficients achieved have the lowest false positive rate and a high true positive rate (mean accuracy of 79.5%), which shows the feasibility to detect the obstacle before the subject reacts. The slope feature offers acceptable performance for classification. |
| Sburlea et al. (2015) | This study focuses on the design of an EEG-based decoder that combines temporal and spectral features to detect early movement states in stroke patients. The study also summarizes the patient's intrinsic motivation when performing gait rehabilitation. | Each trail consists of two-part, i.e., relaxation and movement. Relaxation time was 10 s, followed by a beep to start a movement which of patient dependent time length. The single trial consists of 20 repetitions. In a single week, the subject performs three trials that form 100 relaxation and movements. | It is concluded that using a decoder that combines temporal and spectral features can detect pre-movement state with an accuracy of 64% in a range between 18 and 85.2%, with the chance level at 4% in stroke patients. Furthermore, it was found a significantly strong positive correlation (r = 0.561, p = 0.048) between the motivation of the patients to perform the rehabilitation related task and the accuracy of the BCI detectors of their intention to walk. |
| Lopez-Larraz et al. (2016) | This study proposes an EEG based closed-loop BMI system for control of an ambulatory exoskeleton for gait rehabilitation of SCIP without any balancing and support. | The BMI session is performed consisting of two sessions; one is screening blocks, and the other is closed feedback blocks. The subjects performed 3–4 screening blocks each of 20 trials to calibrate the BMI decoder. In closed-loop feedback, the block consists of four intervals, i.e., rest, preparation, movement attempt, and movement. | Three out of four patients performed at least one successful BMI session with an average performance of 77.61 ± 14.72%. All the patients showed low exertion and fatigue levels during the experiments, which validate the closed-loop BMI system for gait rehabilitation. |
| Hortal et al. (2016) | In this work, a BMI based on the event-related desynchronization and event-related synchronization phenomena are developed to control a lower limb exoskeleton. The BMI can detect the gait starting and stopping pattern. | In each of these sessions, the subject performed several runs (8 or 10). Each run consists of 10 repetitions. Each run starts with 10 s of relaxation time, the 10 s of the walking task, and 5-s rest at the end. | In both preliminary optimization analysis and real-time tests, the results obtained are very similar. The true positive rates are 54.8 and 56.1%, respectively. Regarding the false positive per minute, the values are also very similar, decreasing from 2.66 in preliminary tests to 1.90 in real-time. Finally, the average latencies in detecting the movement intentions are 794 and 798 ms and preliminary and real-time tests. The existing system has the potential to use in real-time BMI for gait rehabilitation. |
| Li et al. (2016) | In this research, EEG signals are recorded to identify two different types of gaits, like movements and phase synchronization between brain regions. | The experimental procedure started with a resting period of 1 min; subjects were lying in a vertical position at an angle of about 55–75 degrees. Afterward, the subjects were instructed to perform automatic gait-like stepping movements (25–30 steps per min). | Our results suggested that brain activities were altered in different frequency bands after SCIP, which supported diverse neural networks with different resonance-like frequencies in the brain. In attempted/active movement, spatial function, and multi-modal integration with somatosensory information were crucial aspects of PMC, the function which needs to be considered separately in different EEG bands. |
| Gui et al. (2017) | This work aims to develop a lower-limb robotic exoskeleton with multiple gait patterns that can be controlled by users' intention. For subject's active participation enhancement, a multi-modal HRI system is established, which includes cognitive HRI and physical HRI. The BCI was used to identify four typical locomotion modes: stop, regular walk, acceleration, and deceleration. | A central pattern generator (CPG) is created to create joint trajectories. The relative state variables of the locomotion mode, i.e., amplitude, frequency, and offset, were transferred to the central pattern generator (CPG) for command generation. In this way, the rehabilitation system is expected to achieve desired assistive gait patterns regulated by EMG based pHRI and EEG based cHRI. | EEG and EMG based HRI to enhance the active participation of patients for gait rehabilitation has been designed. According to the user's voluntary intention, the state variables of CPG are changed through EEG-based cHRI and EMG-based pHRI. The results show that the proposed system incorporates voluntary and active movement consciousness of healthy subjects and stroke patients. |
| Liu D. et al. (2017) | This study describes the impact of different feedback modalities on the performance of an EEG-based BMI that decodes MI of leg flexion and extension. Firstly, an online decoder is built to classify MI, secondly, analyze the effect of visual and proprioceptive feedback on BMI performance, and discriminate features and brain modulations among and within-subjects in this paradigm. | Each participant performs three sessions with a break of a week in between. Each session was composed of 5 runs of approximately 10 min with a resting time of 5 min in between. Each run consisted of 60 trials with extension and flexion cue balanced and randomized inside. | The results suggest that proprioceptive feedback has an advantage over visual feedback. In real-time classification, the average accuracy was 62.33 ± 4.95 and 63.89 ± 6.41% for the two online sessions. The study reported a closed-loop brain-controlled gait trainer as a proof of concept for neuro-rehabilitation devices. |
| Zhang et al. (2017) | This research's primary goal is to classify different gait intentions, i.e., stop, walk, turn left, and turn right from the EEG signals. The other objective is to identify brain areas employed to classify different gait movements in both healthy and SCIP. | The first session consists of four different tasks, i.e., walking forward, turning left, turning right, and stopping. The second was walking and stop with multiple sessions and at least ten walks to stop and stop to walk transitions. | Using MKL and optimal kernel weights simultaneously prove the feasibility of classifying internal gait states from the EEG signals. It also helped to identify and learn through a group of features from relatively active brain areas. |
| Contreras-Vidal et al. (2018) | The study investigates the neural decoding and finds out the relationship between gait kinematics corresponding with neural changes while performing overground gait therapy for chronic stroke patients. | Six Chronic post-stroke hemiparesis patients participated in the experiment of 12 sessions for 4 weeks. H2 robot-assisted exoskeleton was used for training. | A significant relationship between decoding accuracy, total steps, and walking speed was found in the study. The synchronization of EEG signals from the brain, kinematics, and dynamics feedback from the exoskeleton can promote brain reorganization due to motor learning expected because of activity-dependent brain plasticity. |
| Tobar et al. (2018) | The study's objective is to decode cortical activity using EEG signals for ankle flexion and extension at two different force levels in both legs. fMRI is used to locate the brain's anatomical areas, contributing to motor execution and ankle movements. | Eight participants of age mean 29.67 ± 8.81 participated in the experiment. Different experiments on EEG and fMRI were conducted on different days. A total of 8 active tasks were performed in both experiments. | Classification accuracies of 65.64 and 22.19% for estimated current sources and EEG sensor signals with (11.11%) above chance level were obtained. fMRI recording helped to identify the specific areas to generate control commands. |
| Liu et al. (2018) | This research aims to decode the plantar flexion movement intention using continuous classification and asynchronous detection in a gait training paradigm for self-paced gait using movement-related cortical potentials. | Each experimental run consists of 3 consecutive tasks. Participants were asked to relax and fix their eyes on the cross in the monitor's center for 1 min for calibration. In the second task, participants performed self-paced plantar flexion five runs of 10 min, with a rest period of 3 min in between. Each run consisted of 60 trials, with left and right directional cues randomized and balanced inside. | With the proposed movement detection method, a higher true positive rate, lower false positives, and comparable latencies are achieved compared to the existing online detection methods. No significant differences were observed b/w left and right legs regarding neural signatures of movement and classification performance. |
| Hedian et al. (2018) | The study aims for intention detection and classification of fNIRS signals using two variables for motion intention detection, i.e., step length and walking speed. It aims to classify between three different states, i.e., small steps with low speed, small steps with mid-speed, and mid-step with slow speed. | All the subjects were asked to walk at a distance of 4.4 m with three walking states, i.e., the gait of small-step with low-speed, small-step with mid-speed, and midstep with low-speed. All three states were repeated twice with the rest of 30 s in between and backward process tasks to move back to resting position. | In this study, fNIRS-based automatic gait intention detection (walking speed and step size), with the classification accuracy of 78.79%, is achieved. The results confirm the use of fNIRS based BCI for rehabilitation. |
| Khan R. A. et al. (2018) | The work introduces a novel fNIRS-based BCI system that can be used to control prosthetic leg and further utilize it to rehabilitate patients with gait disorders. The study focuses on optimal feature extraction and feature extraction to enhance the classification accuracy using different MLA. | The experimental paradigm consists of a baseline rest of 30 s before the start of the experiment. Subjects were asked to walk on a treadmill for 10-s followed by a rest of 20 s in a single trail. Each subject was asked to perform ten trials with the rest of 30 s between the trials. | The classification accuracies obtained for SVM was higher (75%) relative to other classifier using the hrf were significantly higher (p <0.01). Subject-wise accuracy was 77.5, 72.5, 68.3, 74.2, 73.3, 80.8, 65, 76.7, and 86.7% for the nine subjects, respectively. |
| Costa-Garciacutea et al. (2019) | The study aims to perform online classification of EEG signals, minimize the possible artifacts during gait, reduce classification time, and enhance the system's online accuracy. | Dual tasks were performed, with the primary task was asking the participant to walk on a treadmill with a speed of 2 km/h, while the secondary task was designed to change the participants' attention level. The secondary task was composed of three 30 s trials having mental arithmetics followed by a regular walk and then walking, followed by markers that correspond to low, medium, and high attention levels. | The noisy electrode was removed using MVT, instant kurtosis, and spectral power. The average success rate was enhanced to 69% for healthy subjects while 57% for SCIP using LDA. |
| Elvira et al. (2019) | The study aim was the detection of an unexpected obstacle during normal gait using EEG signals. The study also improves accuracy and reduces the false positive rate in comparison to the previous studies by using IMUs and improvement in feature extraction. | Each subject performed ten trials, which last for 2 min. Each subject is asked to complete a walk at a constant velocity of 2 km/h. During this time, the laser line is projected randomly for one second. The interval between two successive stimuli varies between 6 and 9 s–pa total of 12 and 14 lasers appear on each trial. The subject is instructed to suddenly stop when the laser is visualized and then resume gait afterwards. | The pseudo-online results of the BMI for detecting the appearance of obstacles, with an average percentage of 63.9% of accuracy and 2.6 false positives per minute, showed a significant improvement compared to previous studies. |
| Li et al. (2020a) | The study focuses on fNIRS-BCI for dynamic regulation of two different motion intention states in a realistic environment and detecting movement intension during self-regulated states instead of a resting state. It uses the inter-subject BCI instead of within-subject BCI with improvements in MLA to enhance inter-subject BCI performance. | Subjects were asked to walk at four different self-adapted states, i.e., speed increase, speed reduction, step increase, and step reduction. At the end of every gait, state subjects stop and take a rest for a minimum of 30 s. Every walk state is repeated twice during a single run of the experiment. | GBDT performed well in detecting the onset intention. The 2-layer-GA-SVM model increased the average accuracy of four types of intention from 70.6 to 84.4% (p = 0.005) from the single GA-SVM model. It uses inter-subject BCI instead of within-subject BCI with improvement in MLA. |
| Li et al. (2020b) | The study aimed to generate control command from fNIRS signals obtained from the brain to control assistive devices using walking intensions. | Each trail consists of two-part, i.e., relaxation and movement. Relaxation time was 10 s, followed by a beep to start a movement which of patient dependent time length. The single trial consists of 20 repetitions. In a single week, the subject performs three trials that form 100 relaxation and movements. | TKE is used to extract the features and GBDT to detect the walking intention at each sampling point. The walking model recognition results proved that it is feasible to detect the self-paced intention based on NIRS technology. |
BMI, Brain-Machine Interface; CPG, Central Pattern Generator; EEG, Electroencephalography; EMG, Electromyography; fNIRS, Functional Near-Infrared Spectroscopy.
GBDT, Gradient Boosting Decision Tree; HRI, Human-Robot Interaction; LDA, Linear Discriminant Analysis; MI, Motor Imaginary; MKL, Multiple Kernel Learning.
MLA, Machine Learning Algorithms; MVT, Maximum Visual Threshold; PMC, Primary Motor Cortex; SCIP, Spinal Cord Injured Patients; TKE, Teager-Kaiser Energy.