Brain-Computer Interface technology in motor rehabilitation of stroke
Giulia Cisotto¹, Francesco Piccione², Silvano Pupolin¹
¹University of Padua - Department of Information Engineering, Italy
²Dept. Of Neurophysiology - I.R.C.C.S. S. Camillo Hospital Foundation, Venice, Italy
Keywords: stroke, neuroplasticity, operant-learning, EEG, BCI, motor recovery
Abstract
Brain Computer Interface (BCI) technology has recently been indicated by literature as one of the most promising tool for promoting the functional recovery from stroke and other neurological disease. Implementing a BCI platform for such applications requires the detailed design of algorithms that could allow patients to use the machine in a very friendly way by-passing its inherent complexity thanks to its intrinsic and wise adaptability to the user's behaviours and skills.
In the paper some fundamental issues to cope with in the realization of an effective electroencephalogram (EEG) based BCI system for mildly motor impaired stroke survivors is presented and discussed, and original solutions are provided in order to achieve a higher level of quantification and automatic detection of the cerebral activity that are basic requirements for a suitable and customized rehabilitative tool like BCI has recently become.
Introduction
One of the main technology pervaded field of healthcare nowadays is neuroscience. Disease like Amyotrophic Lateral Sclerosis (ALS), Spinal Cord Injury (SCI), Stroke, Persistent Vegetative State Syndrome (PVS), Parkinson Disease (PD), other severe cerebral injuries, Attention Deficit Hyperactivity Disorder (ADHD) and Epilepsy unfortunately often occupy the vertexes of health annual reports in most the Countries all over the world. Since a recent past, advancements in medicine have led to notable improvements in the lives of these kinds of patients. Unfortunately, many other gaps have to be filled to allow this injured people regain a quasi-normal and care independent life. An emblematic case that represents very well this evolutionary history of progress and new challenges to cope with is stroke. Indeed, stroke is statistically classified as the second or third most common cause of mortality and, with the global ageing of population, its incidence is constantly increasing. Besides that, more than half stroke survivors suffer from several kinds of impairments that limit the quality of their life. Language and dysphasic problems, cognitive deficits, gait, grasping and holding difficulties are only examples of that. With the advancements in medicine and technology, more effective rehabilitative treatments could be offered to patients and a more reliable clinical assistance has allowed patient’s families to live quasi normal lives again.
To this purpose, international guidelines for a good clinical practice have been defined to guarantee the most effective rehabilitative treatment to all stroke patients in the world.
As the motor rehabilitation of the upper limbs concerns, several kinds of therapies have been shown to be effective: goal direction, high intensity and repetition are the key features of most recovery programs. The fundamental reason of their success is that all of them promote neuroplasticity processes which allow brain to regain or bypass damaged neural paths and, consequently, the patient to recover lost motor functions.
Neuroplasticity [1] is one of the most amazing, exciting, controversial and studied property of the brain: research groups all over the world are currently deeply investigating this capacity of brain to modify its functionalities to ensure the best and widest range of performance to the human being. Mechanisms like neural sprouting, dendritic branching and synapto genesis are research topics mainly for neurophysiologists, biologists and physicians while applications and promotion of neuroplasticity is matter of interest for psychologists, physical therapists, neurologists and engineers. In fact, neuroplasticity is largely exploited – although throughout non fully explained mechanisms – in the standard as well in the more innovative stroke rehabilitation programs. It has been already shown that physical therapies, goal directed exercises, repetitive and high intensity training as well as biofeedback, bilateral training, constraint induced therapy and even motor imagery practice promote neuroplasticity changes.
Recently, the idea to exploit neuroplasticity to provide brain, through operant conditioning strategies, an external mirror of its own activity and induce it to change erroneous behavior has been realized in the BCIs [2].
In the paper a brief overview about BCIs, their original applications, their most recent successful examples and their common structure will be presented in the first section.
Afterward, the main goal and challenges of an EEG based BCI system for motor rehabilitation or motor substitution scopes will be highlighted along with the description of a particular such a BCI platform implemented at I.R.C.C.S. San Camillo Hospital Foundation at Lido of Venice with the collaboration of the Department of Information Engineering of Padua.
Finally, a solution to cope with some of the issues previously presented will be provided and some results will confirm its effectiveness and its promising character.
The paper will be concluded with a discussion about the outcomes mentioned before and the sketch of some perspectives and still open issues to address in the future work on the topic.
BCIs overview
Originally implemented for providing an alternative communication channel for people with no residual verbal capacities, BCIs have recently extensively entered the motor recovery research field.
Jacques Vidal, the BCI pioneer and its world-wide recognized father, in 1973 for the first time proved the feasibility to detect brain signals in real-time and use them to control the movement of a cursor on a computer screen [3]. Ten years later, in 1988, Farwell and Donchin implemented a BCI system [4] that could be trained and used to write up to 2.3 characters per minute.
But the first claimed success in the BCI community was reached by Niels Birbaumer and his team in Tubingen in 1999 [5] [6] when one of their patients suffering from a locked in syndrome of an ALS disease, after quite a long training phase, was able to operate a BCI and communicate again with the external world by means of that technology.
Despite of the highly significant historical value of that step in the advancement of medicine, that kind of systems did not become a gold standard for the health care of these severe impairments, anyway. This was mainly due to some consistent disadvantages affected such systems: actually, not much higher than the 2 words/min rate of the first prototypes has been reached yet; moreover, long training periods are still required to satisfactory using such devices. Since no further progresses were achieved, in the recent decades the focus of the BCI community, as hinted at it before, gradually moved from the alternative communication targeting complete paralysed people towards the motor rehabilitation of severe impaired patients suffering from stroke, SCI and tetraplegia [7] [8] [9].
To this purpose, two different approaches are basically considered in the modern rehabilitation practice: the substitutive and complementarystrategy.
Generally speaking, while BCIs for communication as well as BCIs including any kind of robotic device that aims to substitute limbs such as artificial arms, hands and wheelchairs implement a substitutive strategy, the last frontier of BCI is its usage as training tool for motor recovery of the patient's affected limb.
Since its origin in the late 70s, BCI was defined and implemented as any other communication or control system with an input, a signal processing unit and a translation algorithm that transforms the input in the output signals.
Figure 1. A typical BCI structure.
Therefore a typical BCI platform, as reported in Fig.1, is made by:
a signal acquisition unit;
a signal processor;
a feedback control unit;
a feedback provider.
Depending on the way in which the previous blocks are realized and on the kind of device used to acquire the cerebral activity, three main differentiations can be adopted to sum up the widely-spread and many sided research in this exciting field of the Information Technology applied to the Healthcare.Usually, indeed, BCIs are classified first of all on the basis of their synchronous or asynchronous mode of being switched on and off: if the user can manage this switch in a self-paced way, the device is named as synchronous; otherwise, in the most common case an external control unit activates the BCI system and allow the user to operate it. A second distinction deals with the so called dependent BCIs in respect to the independent ones. Dependence is related to the partial involvement of still existing sensory channels that produce a response that can be recorded from the brain: a typical example is represented by the Visual Evoked Potentials (VEPs) that are produced by a flashed letter on a matrix containing the whole alphabet. The generation of VEPs within the brain depends on the gaze direction and, consequently, on extra ocular muscles and cranial nerves that activate them [2]. On the contrary, when an Event Related Potential (ERP) is elicited by the subject during (purely cognitive) identification of the meaningful stimulus among many others, the information about the supporting cognitive processes is carried only in the brain and does not depend on any other communication pathway.
Finally, signals that carry information to drive BCIs can be collected in either an invasive or a non-invasive way: basically, devices like electrocorticogram (ECoG), microarray of electrodes or microelectrodes grids can record neural activity over the cerebral cortex or even inside brain, providing, respectively, data of the different parts of the gray matter or of the spikes and the local field potentials (LFPs) created by various neural populations.
Successful results were achieved in 2012 in Providence and in 2013 in Pittsburgh by means of such invasive systems: specifically, in 2012 Hochberg and his collaborators at the Donoughe's Laboratory in Providence provided a tetraplegic woman of a robotic arm that, after a period of training, was able to control by a neural interface made by an intracortical grid of 96 microelectrodes implanted on her motor cortex to control the BCI system that, in turn, controlled the robotic arm so that she was able to drink from a can placed on the plane in front of her wheelchair.
A year later, Collinger and colleagues in Pittsburgh gave another tetraplegic woman a robotic arm that she learned to properly command through her cerebral signals gathered by an analogous grid of microelectrodes implanted in brain as that of Providence.
If, on one hand, as confirmed by the aforementioned examples, the invasive solutions allow to collect information in a high signal to noise ratio (SNR) quality mode, implantation of these devices is crucial and highly invasive. A different approach consists in recording brain activity from non-invasive devices such as EEGs, functional magnetic resonance imaging scanners (fMRI), magnetoencephalograms (MEGs), functional near infrared spectroscopy (fNIRS) or others. Although each of them presents benefits as well as drawbacks in respect to each other, special attention in the BCI field is reserved to the EEG. In fact, in spite of its poor spatial resolution – of some centimeters – and low SNR, it can provide high temporal resolution data – in the range of milliseconds – at a relatively low expense along with a high portability. These benefits represent notable advantages over the other non-invasive devices and can bring EEG to become the gold standard for the future BCI home daily care. Large literature has already been written about EEG based BCIs and different EEG patterns and characteristics have been investigated to provide the best control for the entire system. To this end, several experimental protocols have been implemented to elicit endogenous neural behaviors, the previously mentioned ERPs – the most commonly used being the P300 - or to induce oscillations phase locked to externally presented events, named as Event Related Desynchronization (ERD) [10].
The optimal EEG pattern to use depends on the specific application, the output device to control and the reaction time required.
To this purpose next section highlights the main challenges that an EEG based BCI system is asked to address in order to properly control robotic arms, haptic device for rehabilitative training scopes, wheelchairs and other outputs that require an online control.
Challenges for an EEG-based online control
As mentioned before, recently BCI community moved a lot of efforts from the alternative communication to the motor rehabilitative field where the real-time decoding of the intention to move of a patient and the subsequent execution of that action by means of any robot have to be performed.
Actually, this is only one side of the coin: BCIs, indeed, are systems where a mutual adaptation between the individual and the machine has to take place.
In order to realize this situation a dual mechanism of learning and adaptation to the human's cerebral activity expressing the movement has to be implemented by the machine, properly processed, and finally transferred to the output device.
At the same time, however, receiving the robotic feedback by such an output at each repetition of the movement the patient is subjected to his/her own neuroplasticity changes that would bring brain to modify its functionality (and, in longer time, even its structure) to take as much help as possible from the machine to accomplish the movement.
Therefore, one of the main challenges of properly providing such a feedback to the patient is to reliably decode his/her intention to move from the ongoing EEG.
Although literature has already reported and defined neural correlates of movement in such BCI applications [10], the large interindividuals and even intrasubject variability can cause identification algorithm to fail.
Specifically, the so-called desynchronization of the sensorimotor rhythms (SMR) can be considered as the sign of the intention to move and even of the planning phase before a movement and consists of a decrease in amplitude of the EEG components around 10 Hz, the world-wide known μ band.
From literature [10] it was observed to start in the cerebral hemisphere at the opposite side of the movement, i.e. contralateral, and gradually become bilaterally distributed.
The dual phenomenon of synchronization usually takes place at higher frequencies, the so-called β in the range (12, 30) Hz, during the recovery phase after movement has been accomplished.
Both the desynchronization and the synchronization phenomena experience, as referred above, a high level of variability that can cause significant changes in their frequency, temporal and spatial distribution.
The choice of the site, frequency and timing of EEG features extraction and their processing to prepare a suitable robotic feedback is then of crucial importance.
In next section, examples extrapolated from a particular EEG-based BCI application will provide a measure of such a challenging issue.
Furthermore, it has to be pointed out that one of the most important strategy to induce patients in (re)learning a lost motor function by means of a BCI system is the operant-conditioning that requires a contingent or slightly delayed rewarding (punishing) feedback to help (impede) patients in completing the movement whenever they (do not) produce the expected movement related desynchronization (MRD), i.e. the neural correlate of motion.
Therefore, a double issue has to cope with for providing an effective rehabilitative feedback: integrating the two challenges mentioned before, proper EEG features have to be identified as early as possible before to release the robotic feedback so that a suitable reaction to the subject's brain activity could be scheduled.
In the process towards the accomplishment of this computation, disturbing phenomena could distort the ongoing EEG signals destroying the reliability of a feedback planned on the features extraction phase from such kind of distorted traces. These kinds of disturbances are generally referred as artifacts. They can be classified as follows:
- External interferences. The main element of this set is the power line noise that usually corrupts the EEG recordings. For this reason, a notch filter around 50 Hz or 60 Hz is implemented to remove this considerable interference during EEG evaluations or experimental sessions.
- Physiological interferences. They can be further divided into two subclasses: muscular and neural noises. Eye-blinks, eye-gaze changes, chewing, gnashing, swallowing and head slight movements are muscles activations that can compromise the whole recordings. Skin sweat can be also a relevant phenomenon to cope with sometimes. Finally, distractions, habituation and other collateral cognitive phenomena can elicit neural populations of different cortical regions to spike and, at the scalp level, to show interfering waveforms. The latter are considered disturbance and are usually removed on the basis of their spatial and/or frequency occurrence.
Then artefacts can occur either accidentally or along the whole recording. For instance, artefactual activity due to mains is usually present along the entire registration while muscular contraction is in the most cases a very short phenomenon that can seriously corrupt a relatively short-lasting recording segment.
One of the most impacting causes of artefact is the previously cited electrode-pop: although quite rare, this kind of noise can be completely superimposed over the low amplitude useful signal and make the identification procedure of the EEG characteristics almost impossible. A typical example of its shape is captured by Fig.2 where the usual abrupt negative fall, overshoot and slow-oscillating return to baseline values are clearly visible.
Figure 2. An example of electrode-pop artefact.
It can be easily expected that such an artefactual activity compromises any kind of automatic features identification. Cautions in order to avoid this kind of artefacts can be taken during the recording preparation: clinical technicians are trained to pay attention on this type of occurrence.