User:MusaSMahmood/sandbox



= Motor Imagery-based Brain Machine Interfaces =

Motor-imagery-based brain machine interfaces (MI-BMI) are devices that translate brain activity information into commands for controlling external software or hardware by using motor imagery. Brain-machine interfaces allow users to bypass normal motor function by gathering neural data directly from the brain, and have a number of applications in research, mapping, augmentation and rehabilitation. Motor imagery or motor imagination involves mental simulation of motor actions without activation of muscles. Motor imagery provides many advantages over visually-evoked paradigms such as SSVEP and P300, as it does not require the use of stimulation devices, which often obstruct the view of the user. Therefore, motor imagery provides a promising communication and control channel between the brain and external targets that do not require the use of peripheral nerves and muscles, making it useful for individuals with serious motor disability. Despite the clear advantages and convenience such a system would provide, no commercially available motor imagery-based BMI exists, and development of such an interface faces many issues, such as performance variability between subjects and limited spatial resolution of EEG. It is likely that implanted electronics will demonstrate a more feasible strategy for MI-BMIs as they provide direct access to the neuronal clusters that are most involved in motor imagery activity.

History
Pioneering work on motor imagery-based BMIs was performed by Gert Pfurtscheller using EEG, who detected event-related desynchronization (ERD) localized over the contralateral hemisphere during separate movements of either hand in human subjects. This discovery opened up the possibility of controlling brain machine interfaces with non-invasive EEG. Prior to this, study of motor imagery was performed using timers, measurement of regional cerebral blood flow (rCBF) using positron emission tomography, magnetic resonance imaging (MRI), and functional magnetic resonance imaging (fMRI). In 2001, Pfurtscheller and Neuper demonstrated online discrimination between left-hand versus right-hand movement with a tetraplegic subject. Later, they demonstrated asynchronous detection of these same activities along with a null class (no activity) using common spatial filters and linear discriminant analysis. Later work provided incremental improvements to classification strategies, and introduced new motor imagery tasks, such as movement of feet and tongue. Since then, motor imagery-based BMIs have been demonstrated to be effective in restoring motor control for stroke patients, with better functional outcomes observed in a split-intervention study.

There has been limited performance improvements to EEG-based MI-BMIs due to physical limitations such as spatial resolution, with minor only minor improvements in number of classes. For example, separation of left foot versus right foot is challenging due to the close proximity of the regions in the brain. In 2013, Yi et al. demonstrated the ability to classify both simple and compound limb motor imagery. In 2019, separation of left and right foot kinaesthetic motor imagery was demonstrated using common spatial pattern. The most common classes demonstrated are the left and right hands, feet (usually together, but occasionally separable), and tongue; for up to 5 motor imagery classes.

Mechanisms
Motor imagery is a dynamic state during which a subject mentally simulates a given action, implying that the subject feels themselves performing a given action, usually corresponding to a first-person perspective. Imagined motor tasks can be characterized as either kinesthetic (first-person process) or visual (third-person process), with kinesthetic imagery demonstrating a clear spatial pattern in the sensorimotor regions, and visual showing less decipherable EEG patterns. Brain activity during motor imagery is involved with fluctuations in sensorimotor rhythms; oscillations generated during preparation, execution and imagination of motor activity. In this case, imagination or preparation behavior can be observed as desynchronization or synchronization events in sensorimotor rhythms. Specifically, event-related desynchronization (ERD) involves a reduction in power during movement preparation or execution, and an increase in power (event-related synchronization) occurs after completing the movement. During imagery, there are contralateral ERD and ERS effects on the central and parietal lobes, with independent component analysis demonstrating localization and strong activity in the primary motor cortex. Although the primary motor cortex is the main source of signals that communicate with the spine and execute movements, preparation and imagery are not exclusive to this region. The premotor cortex and supplementary motor area (SMA) are also implicated in the planning and coordination of motor actions.

Motor Simulation Theory
Motor simulation theory is a concept proposed by Marc Jeannerod as a unifying mechanism for motor cognition; explaining how action-related cognitive states relate to motor execution. The theory suggests that MI works by rehearsing motor systems off-line via a hypothetical simulation process, and are based on the theoretical concept of forward modeling or embodied/grounded cognition, where MI are considered to be embodied mental states.

Non-Invasive
Due to the cognitive nature of MI, it is difficult to verify adequate performance of MI tasks by a subject, as well as inability to provide feedback on what mental states provoke an adequate, predictable response. Therefore, attempts have been made to combine MI with noninvasive technologies such as electroencephalography (EEG), functional magnetic resonance imaging (fMRI), and magnetoencephalograpy.

Electroencephalography (EEG)
The most common non-invasive method for acquisition of MI brain activity for use in BMIs is electroencephalography (EEG), which demonstrates some promise for use in rehabilitation of upper limb function after stroke. EEG is able to adequately detect sensorimotor rhythms in the range of 13-15 Hz, along with the corresponding desynchronization and synchronization events that are characteristic to MI. Sensorimotor rhythms are highest in power when the corresponding sensorimotor areas are idle, and decreases when the corresponding sensory or motor areas are activated during motor imagery or motor activity. The advantages to EEG are that it is completely noninvasive, only requiring electrodes to be placed on the scalp in order to measure relative potential differences between various scalp locations as brain activity data. However, EEG is limited by the power of signals that can be recorded at the scalp surface, and also has fairly limited spatial resolution due to the size of the electrodes required for measurement and how closely they can be spaced on the scalp. Despite these limitations, EEG has been shown to be the preferred technology for uses in brain-machine interfaces due to its safety, reliability, and ease of use. Recent technological advances have allowed for improved signal acquisition through the use of lightweight, flexible electronics and dry electrodes. The latest EEG designs display a trend toward wireless, wearable devices, which are preferable for day-to-day monitoring, with compact battery-powered designs over conventional bench-top amplifier boxes and hair-cap-based systems. For mobile systems, dry electrodes are preferred due to short setup times, minimal skin irritation, and excellent long-term performance. Additionally, they perform better than gel-based EEG sensors while providing long-term wearability without reduced signal quality.

Additional methods
Detection of sensorimotor rhythms is also possible with magnetoencephalography, although the applications are limited due to the restrictions of magnetoencephalography as a technology. Functional MRI has also been used with EEG, but has not been adequately demonstrated as an adequate BMI candidate. The primary limitations are the nature of MRI and magnetoencephalography as requiring large, bulky equipment, while also requiring the operator to remain stationary in an uncomfortable position.

Invasive
Use of EEG and other non-invasive technologies such as fMRI or magnetoencephalography are limited in their applications. The primary limitations for EEG is its bias towards peripheral neurons (neurons closest to the scalp surface), and lacks access to deeper cortical tissue. This is simply due to the inverse square law, where electrical impulses generated by neurons further away from the electrodes on the scalp surface diminish at a rate square to the distance. In addition to the surface-bias, EEG is also limited by it's spatial resolution, and is unable to target neurons individually, or even as groups or clusters. The EEG gathered at the scalp surface include the sum of activity across the brain, although mostly representative of post-synaptic potentials in the superficial cortex. This means that any single point on the scalp represents a blurred image of what's going on under the surface. A common means to get around this limitation is to use large clusters of electrodes; as many as (but not limited to) 256 electrodes, and then attempt to use advanced preprocessing and feature extraction methods to extract the most relevant information, or attempt to heuristically invert the data to determine what parts of the brain contribute to what signals.

Alternatively, in order to achieve the best possible signal, researchers may attempt to bypass the barriers protecting the brain, and attempt to place electrodes on or within the brain itself. This results in highly localized and high quality signals with consistent impedances, as well as access to deeper tissue. Depending on the electrode sizes, they may be able to capture action potentials from single neurons or small neuronal clusters. This results in a significantly higher resolution and the capability to target individual clusters with great precision. The major drawbacks from such invasive systems are the requirements for costly and risky surgeries, as well as the use of expensive cutting edge devices that will need to survive the harsh environment of the brain; where the fluid can often cause corrosion and deterioration in many electronic and housing materials. Building robust implantable electronics is the primary challenge to development of future invasive brain-machine interface solutions.

Electrocorticography (ECoG)
Electrocorticography involves the use of electrodes placed on the exposed brain surface, providing localized, high quality signals of brain electrical activity. Due to the surface placement of the electrodes, ECoG is limited by the surface neurons. In the case of motor imagery-based BMIs, surface access to the motor cortex provides adequate coverage of important motor pathways to allow for implementation of MI as a paradigm.

Additional Methods
Deep brain stimulation and cortical implants are at the forefront of BMI development, allowing for access to neurons deeper in the cortex, as well as allowing for neural feedback through neurostimulation. Companies such as Neuralink are actively developing implantable brain-machine interfaces that can actively target various neurons using up to 1500 electrodes, and will likely be able to support a greater number of tasks when compared with its EEG-based counterparts.

Classification of Motor Imagery EEG Data
EEG is the foremost method for integration and classification of motor imagery due to its safety and ease of use. The primary challenge when integrating an EEG-based brain-machine interface is the processing and classification of raw EEG signals. A number of feature extraction, feature selection and classification techniques have been developed for the classification of EEG-based motor imagery, with varying degrees of success. A sampling of research papers demonstrating classification of MI-based BMIs is shown in the table below. Processing of raw EEG data begins with selection of signal processing techniques. Signal processing and feature extraction techniques include filtering, autoregressive modeling, fast-Fourier transforms, and other frequency-domain based techniques. Additionally, time-frequency domain techniques are also used as it reveals spectral information about the EEG; including the use of short-time Fourier transform (STFT), wavelet transform, and discrete wavelet transform. Decomposition methods using wavelets are useful as EEG contains various frequency bands containing different information about motor imagery. These methods are excellent for deriving dynamic features due to EEG being non-linear and non-Gaussian. Another commonly used method for extracting features are common spatial patterns (CSP). CSPs are commonly used to classify MI EEG, as different frequency bands of EEG contain different information, and CSP enables extraction of this information from different bands. Variations of CSP when combined with strong classification methods such as linear discriminant analysis or support vector machines have demonstrated some of the best classification accuracies in published literature. However, since the emergence of convolutional neural networks as a powerful feature extraction and classification strategy for EEG signals, it has been demonstrated to have some of the best performances when compared with conventional feature extraction and classification methods.

Advantages and Disadvantages
Motor imagery (MI) is a greatly advantageous paradigm for persistent BMI when compared to evoked paradigms as it does not require the use of external stimuli; its classes are based on imagined motor activities such as opening and closing a hand or moving feet. Additionally, trained subjects are able to rapidly switch between motor imagery tasks, allowing for lower latencies between tasks when compared with visually-evoked paradigms. Motor imagery can also be acquired with relatively few electrodes when compared with other BCI methods. Finally, use of improved training techniques such as neurofeedback and immersive virtual reality training may allow for more efficient and consistent training with less responsive subjects.

Performance Variation
The performance of a MI-based BCI is heavily dependent on the state of the user, with high intra- and inter- subject variability in classification performance. There is a strong relationship between signal quality and the mental state of the user, with fatigue or drowsiness in test subjects yielding poorer results. Evidence suggests that low-performance groups have a less-developed brain network for motor imagery, resulting in poorer performance. Due to limited intra-subject and clinical studies, it is unclear what strategies may be used to improve performance with BMIs.

Applications
Applications for motor imagery-based BMIs are limited by the number of classes that can be classified, and are used in both control and monitoring applications. Control applications involve the manipulation of external devices, such as a vehicle or a prosthetic device, while monitoring involves the determination of the mental or emotional state of the user. Biomedical applications of BCI include the replacement or restoration of central nervous system (CNS) functionality that is lost due to disease or injury. Illnesses such as stroke or amyotrophic lateral sclerosis (ALS), paralysis, amputations, and traumatic injury may result in permanent loss of motor capability. BCI-based prosthetics have previously been demonstrated using paradigms such as SSVEP. Motor imagery-based prosthetics with up to 5 degrees of freedom have also been demonstrated, but are limited by performance variations, with subject-to-subject accuracy ranging from 56% to 100%. Motor imagery and neurofeedback-based BCIs have also shown promise for rehabilitation of motor function after injury or illness due to the active involvement in neural plasticity and development. MI-BMIs have been demonstrated to promote clinical and neurophysiological changes in stroke patients through the use of neurofeedback training over extended periods. Further uses in therapy and assessment have shown that incorporation of virtual reality (VR) to provide visual feedback to the user, allowing them to safely navigate virtual environments, creating a more immersive neurofeedback environment.