| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
TOPICAL REVIEWS |
1 Istituto Scienze e Tecnologie della Cognizione - CNR, Rome, Italy
2 AFaR - Centre of Medical Statistics and IT, Fatebenefratelli Hospital, Rome, Italy
 |
Abstract |
|---|
 Top Abstract IntroductionReferences |
|---|
(Received 26 January 2007;
accepted after revision 20 February 2007;
first published online 1 March 2007)
Corresponding author F. Tecchio: ISTC – CNR, Unità MEG, Dipartimento di Neuroscienze – Ospedale ‘Fatebenefratelli’, Isola Tiberina, Rome, Italy. Email: franca.tecchio{at}istc.cnr.it
|
Introduction |
|---|
|
TopAbstractIntroductionReferences |
|---|
|
For BCI to be effective in real-world applications, many challenges must be addressed and overcome (Moore, 2003). The required cognitive load must be minimized through optimization of the cerebral activity extraction. In fact, while most BCI systems are tested in quiet laboratory environments, where users are able to concentrate on the task at hand with minimal distractions, BCI users in the real world have to deal with much more complex situations, including emotional responses, interactions with other people, and safety considerations. Every BCI has its own operational protocol which defines the procedures for switching on and off, the continuity or discontinuity of the communication, whether the relevant signal is generated consciously by the subject or in response to a stimulus triggered by the BCI (event-related), the exact sequence of interactions between the subject and the BCI, and the type of feedback provided to the subject. In real life the subject must be able to choose the message and to carry out the switching on and off procedures.
Feature extraction
The feature extraction stage processes the cerebral data and extracts relevant features to feed the translation algorithm (Fig. 1). The cerebral data are acquired by some monitoring system, using appropriate experimental protocol designs. Data are then AD converted, using the sampling rate suitable to the cerebral signal being processed, at least twice the expected highest relevant frequency component in the signal of interest. The digitized signal must be analysed further to extract information that is relevant to the mental task under investigation. Before the signals are used, however, often some unwanted artifacts need to be removed. Without exaggeration, thousands of different features have already been used in BCIs, being divided into three general domains: time domain, frequency domain, and joint time–frequency domain. The features are also likely to be location specific.
Mandatory for feature extraction success, the experimental protocol must be designed to suit the application and the environment in which the BCI will be used. This includes choice of mental task, stimulus parameters (e.g. visual scenery timing and constraints), a minimization of unwanted stimuli and distractions that may affect the properties of the signals to be monitored, at list in the first phase of initial algorithm training.
Feature extraction in patients
When developing BCI applications devoted to patients affected by different pathologies, it is mandatory to consider that the cerebral processing is hugely distorted by diseases. For instance, in the case of stroke patients, a relevant phenomenon is that different areas start to replace the function previously played by the damaged areas (plastic modifications characterized by unusual recruitments: Rossini et al. 1998, 2001; Oliviero et al. 2004; Tecchio et al. 2006d). In the case of peripheral damage, maladaptive (Flor et al. 2006) or adaptive (Tecchio et al. 2000a, 2002, 2006d) cerebral reorganizations can occur (Rossini et al. 1994b; Tecchio et al. 2005a, 2006c).
Translation algorithms
The relevant extracted features must be transformed into commands directed to the device which has to execute the subject's intentions. This algorithm can be based on linear (e.g. standard statistical methods) or non-linear (e.g. neural networks) techniques. Examples of artificial intelligence methods used are linear discriminant analysis, artificial neural networks, genetic algorithms, kernel-based learning methods (support vector machines, kernel Fisher discriminant), Bayesian networks, and hidden Markov models. In order to be really efficient, an algorithm should adapt to the human subject in at least three phases: (1) initial training of the algorithm, in which BCI adapts off-line to the physiological characteristics of the subject; (2) adaptation by periodic on-line adjustments; (3) mutual adaptation and reinforcement, where the subject's control of the physiological signals used to control the BCI, and the translation algorithm's ability to decode these signals reaches a stable state and an optimization of low error rate is reached.
Learning
Although BCIs provide alternate communication channels that bypass traditional neuromuscular channels, learning to operate a BCI successfully is similar to learning tasks that involve muscular control. Just as walking or speaking requires training and practice, the operation of a BCI is an acquired skill that involves many of the same learning mechanisms. Successful use of BCIs requires that the user maintains his/her capability to learn developing new abilities in controlling not the usual neuromuscular channels, but the EEG pattern that is recognized as relevant by the BCI. Throughout the repeated execution of any task, the brain undergoes plastic adaptation that may be relatively short lived or more persistent. From a system point of view, this plasticity is in essence a transfer function change on the human subject side that inevitably affects the overall behaviour of the BCI and requires adaptation on the machine side as well.
Feedback
The BCI requires the subject learning both during the relevant signal production and translation maintenance. The most important element of such learning is the presence of feedback. During the learning process, either while walking, speaking, or using a BCI, the subject makes adjustments based on the outcomes produced by their own efforts in order to eventually hone their skills appropriately. The incoming signal produced by the outcomes is known as feedback, and it is crucial to the learning process (Graimann et al. 2006). As mentioned above, most BCIs feedback to the subject the outputs on a computer monitor. The use of this visual information as feedback presents problems for fast applications. Treisman & Kanwisher (1998) concluded that it takes at least 100 ms after stimulus presentation for sensory perception and nearly another 100 ms for the information to become conscious. This relatively long delay to process conscious information complicates the use of visual feedback as the only one in real-time BCIs, e.g. a system devoted to move an arm in a dynamic and complex situation. Among a number of possible solutions that may be tried in the future to reduce the problems related to perceptual delays, the use of haptic feedback and electrotactile sensations seems to be the most compelling (Graimann et al. 2006).
Throughput and latency
Two important numerical parameters that allow the selection of appropriate interfaces for driving specific applications are the throughput and latency. (1) Throughput (also called bitrate, bandwidth, or information transfer rate) is the rate at which a computer or network sends or receives data. It is a good measure of the channel capacity of a communication link – the throughput unit used in the connections to the internet is how many bits they pass per second (bit s–1). (2) Latency is a time delay between the moment something begins, and the moment its effect begins. Up to now, even the best average information transfer rates for experienced subjects and well-tuned BCI systems are relatively low, in the region of 24 bits min–1 (roughly three characters per minute, Wolpaw et al. 2002). This is too slow for natural interactive communication, so, to effectively use BCIs as an alternative to conventional interfaces, it is necessary to research ways of optimizing selection techniques and incorporating prediction mechanisms to speed up communication.
Signal-to-noise ratio
Like any other communication channel, the BCI system performance heavily depends on the signal-to-noise ratio (SNR). Feature extraction success depends on the SNR, specifically estimated as the ratio between the occurrence of the wanted (by the subject using the BCI) cerebral signal modulation and the ongoing variation of the same signal in a certain time period. This ratio will affect the translation algorithm efficiency, defined as the ratio between the desired and the produced output. For instance, in a BCI based on sensorimotor cortex mu rhythms, the biological noise includes the alpha activity from other brain areas (e.g. the visual cortex). It is worth noting the importance that BCI systems identify and eliminate signals generated outside the central nervous system, like the electromyography (EMG) from scalp and face muscles and electroculogram (EOG). A further and difficult step is the discrimination of signal from noise if they have similar topography, amplitude and frequency content. For instance, the EOG is more problematic than EMG for those BCI systems operating on the basis of slow cortical potentials (SCP; Birbaumer et al. 1999) because of their frequency overlap; similarly beta-dependent BCI systems are more sensitive to EMG artifacts (Goncharova et al. 2003).
Brain signal monitoring systems
EEG and MEG (Niedermeyer & Lopes da Silva, 1997; Del Gratta et al. 2001) are non-invasive techniques that detect electrophysiological signals with the temporal resolution of a millisecond or better. For both techniques, the origin of the signal is mainly the effect of the postsynaptic currents associated with synchronous neuronal firing in the brain. EEG detects the electrical potential difference measured from the scalp, i.e. it is a reference-dependent measure. It requires the contact between the recording electrodes and the scalp. MEG detects magnetic fields at the cranial surface, giving an absolute measure at each point. The MEG recording system is brought near the head and the sensors are not in contact with the scalp. EEG is equally sensitive to cerebral sources orientated both radially and tangentially to the head surface, whereas MEG is almost selectively sensitive to the latter. Moreover, EEG signal is distorted in space and time by the passage through the conductivity discontinuities of cerebrospinal fluid, meninxes, skull and scalp, while MEG signal is transparent to these discontinuities in the first approximation of spherical head. ECoG (Gastaut, 1952; Keene et al. 2000; Allison et al. 1991) is the invasive recording of cortical potentials from electrodes placed intracortically or in deep brain regions. With cortical implanted electrodes it allows the mapping of cortical functions; often this investigation is performed intraoperatively, with adjunctive difficulties due to anaesthetics. In humans, ECoG can be used only in pathological conditions requiring neuro-surgery or deep brain stimulation (DBS).
Functional magnetic resonance imaging (fMRI) detects changes in the concentration of deoxyhaemoglobin, dependent on a complex interplay among blood flow, blood volume and cerebral oxygen consumption (Heeger et al. 2000; Heeger & Ress, 2002). When neurons increase their activity with respect to a baseline level, a modulation of the deoxyhaemoglobin concentration is induced, generating the so-called blood oxygen level-dependent (BOLD) contrast (Rees et al. 2000). BOLD dynamics are characterized by an initial transient small decrease below baseline due to initial oxygen consumption (negative dip), followed by a large increase above baseline, due to an oversupply of oxygenated blood only partially compensated for by an increase in the deoxygenated venous blood volume. The BOLD signal could reflect both the firing of local neuronal assemblies and also the amount of their synchronized input, even if insufficient to evoke an action potential spike, as well as fluctuations in firing synchrony, which can increase or decrease without affecting the net firing rate (Rees et al. 2000; Heeger et al. 2000). A comparison among different techniques is schematized in Table 1.
|
Section 1. Functional source separation (FSS)
To ultimately discover and translate the subject's intentions, a BCI starts from cerebral activity as depicted by some brain signal monitoring system. In the case of EEG and MEG, the sensed signal is a linear mixture of source activities. This corresponds to the basic model of the blind source separation (BSS) technique. These algorithms estimate complete source time courses on the basis of the statistical properties of the generated signal, without taking into account the physical nature of the generating phenomenon. Thus, BSS procedures use only information contained in the waveform of original signals, a very convenient property for electrophysiological techniques (EEG and MEG), which provide the most informative time–frequency signal from the intact human brain.
In the last decade the BSS techniques, in particular independent component analysis (ICA) algorithms, have been successfully applied to EEG and MEG data to estimate signal of interest (Makeig et al. 1996, 2002, 2004; Vigario et al. 1997; Tang et al. 2005; for a comprehensive review on BSS see: Hyvärinen et al. 2001; Cichocki & Amari, 2002). The aim of such techniques is to extract in a ‘blind’ fashion (i.e. without making specific assumptions) meaningful signals that have been mixed linearly, without knowing the original signals or the mixing coefficients. There appears to be something magical about BSS; we are estimating the original source signals without knowing the parameters of mixing and/or characteristics of the sources. In fact, without some a priori knowledge, it is not possible to uniquely estimate them. However, one can usually estimate them up to certain indeterminacies. In mathematical terms these indeterminacies and ambiguities can be expressed as arbitrary scaling, permutation and delay of estimated source signals. These indeterminacies preserve, however, the waveforms of the original sources. Although these indeterminacies seem to be rather severe limitations, in a great number of applications, like the electrophysiological ones, these limitations are not essential, since the most relevant information about the source signals is contained in their waveforms and not in their amplitudes or in the order in which they are arranged in the output of the system. In particular, the ICA assumption is that a set of statistically independent sources s have been mixed linearly in the recorded data x by means of a mixing matrix A. The aim is to recover both s and A starting from the observation of the linear mixture x = As without making any particular assumption other than statistical independence of the sources.
Based on the observation that when we deal with real-world signals we are never completely ‘blind’, in that we know (in a more or less detailed and quantitative way) some of their characteristic features, a new approach, called functional source separation (FSS), has been recently proposed by our group (Barbati et al. 2006; Porcaro et al. 2007). The aim of FSS is to enhance the separation of relevant signals by exploiting some a priori knowledge without renouncing the advantages of using only information contained in original signal waveforms. A modified (with respect to standard ICA) contrast function is defined: F = J +
R, where J is the statistical index normally used in ICA, while R accounts for the prior information of the sources. According to the weighting parameter it is possible to adjust the relative weight of these two aspects. Moreover, since prior information on the sources may also be described by a non-differentiable function, the new contrast function F is optimized by means of simulated annealing. This does not require the use of derivatives, and performs global optimization, while gradient-based algorithms usually employed in ICA only guarantee local optimization. To separate contributions representing different sources, the proposed procedure could be applied in two different ways: by using an orthogonal extraction scheme (as happens in the basic ICA model); after having estimated the first source, the second one is searched in the orthogonal space with respect to the first, and so on until the last source is estimated – with a stop rule that can be defined according to the data in hand. Since relevant components could not always be reasonably assumed independent/uncorrelated, in the FSS procedure the orthogonalization step could be also completely skipped, producing a non-orthogonal extraction scheme. In this condition, the order of extraction is not significant, because the procedure is applied each time to the original data. Different constraints are applied each time to produce different sources.
The provided sources are suitable to describe ongoing activity time courses, which allow, for example, single trial analysis, instead of describing the activations by averaging all sensors channels and only in specific instants, as is usually done in the standard procedures. Moreover, even if a source is extracted by exploiting a functional constraint related to a specific time portion of the experiment, the corresponding estimated signal could be studied all along the length of the whole session.
Functional constraints. The described optimization procedure exploits a wide variety of constraints, which express the functional properties, specific to the sources to be estimated (Fig. 2). The BCI applications we are interested in are devoted to the upper limb movements, in particular to the sensorimotor control of the hand. The MEG displays an optimal capability to identify cerebral regions devoted to the hand sensory representation, describing the physiological somatotopic organization (Hari et al. 1984; Tecchio et al. 1997; Wikstrom et al. 1997; Pizzella et al. 1999; Zappasodi et al. 2006) and its distortion as a consequence of central (Rossini et al. 2001; Oliviero et al. 2004; Tecchio et al. 2005a, 2006c,d) and peripheral damage (Tecchio et al. 2002). Moreover, MEG can provide a description of the peripheral–central connectivity of the neural network devoted to the hand (Tecchio et al. 2000b, 2005b).
|
We identified in the primary cortex devoted to the districts innervated by the median nerve two functional sources (FSs) related to the sensory flow induced by median nerve stimulation. The first one – named S1a – describes the activity related to the well-known marker of the stimulus arrival in the primary sensory cortex (Hari & Kaukoranta, 1985; Allison et al. 1991). This is known to be mainly generated by excitatory postsynaptic potentials impinging on broadman area (BA) 3b pyramidal cells. As it is maximally recruited at around 20 ms from the stimuli at the wrist, the functional constraint taking into account the ‘reactivity’ to the stimuli to identify S1a was defined as:
|
| (1) |
(
) is the time point corresponding to a field amplitude of 50% of the maximal power – by definition in t20
– before (after) t20; the baseline (no response) was computed in the time interval from 10 to 15 ms. As only one component is extracted at each time, it is possible to avoid the amplitude indeterminacy inherent to the general ICA method. Once the source which optimizes the contrast function F has been obtained, the estimated solution is multiplied by the Euclidean norm of its weight vector 
(
such as 
, with 
), allowing amplitude comparisons among sources in a fixed position.
We considered a second cerebral source named S1b, i.e. the source maximally activated in the sensory areas at around 30 ms from nerve stimulation. While invasive recordings using ECoG in human (Allison et al. 1991) showed BA 1 at the crown of the postcentral gyrus contributing to this wave of somatosensory-evoked potentials, this radial component is poorly detectable by MEG. The component around 30 ms as recorded by MEG is mainly generated by BA 3b inhibitory and BA 4 excitatory networks (Wikstrom et al. 1996; Kawamura et al. 1996; Huang et al. 2000; Tecchio et al. 2005a). The functional constraint to obtain S1b was defined as:
|
| (2) |
and 
were as in eqn (1), with t30 instead of t20; the baseline was defined as in eqn (1). Again, the estimated solution was multiplied by the Euclidean norm of its weight vector 
(
such as 
, with 
).
Primary sensory finger areas.
To identify neural networks devoted to individual finger central representation, the ‘reactivity’ to the stimuli was taken into account in the whole period including the two early components. It was defined as follows: the evoked activity (EA) was computed separately for the two sensorial stimulations by averaging signal epochs centred on the corresponding stimulus (EAL, little finger; EAT, thumb). The reactivity coefficient (Rstim) was then computed as:
|
| (3) |
As already mentioned, the orthogonality constraint was removed and the two finger sources were searched for starting from the original data. In fact, in the specific and restricted cortical region of interest, neural networks are spatially highly interconnected and superimposed and temporal overlap of finger sources activation could be reasonably hypothesized.
Primary visual sustained induced activity.
In the visual cortex, to observe the cerebral activity induced by a sustained stimulus, the robust and temporally induced power increase of gamma activity was exploited, by the spectral power band relative variation during the whole period of sustained stimulus and the period of stimulus absence. The following ad hoc functional constraint R was:
|
| (4) |
(gamma, 20–70 Hz) frequency band and normalizing this difference with respect to the gamma activity level at No-stimulus (Barbati et al. 2007).
Primary motor hand area.
To identify the source in the primary motor area devoted to the control of the hand movements – named M1, the coupling of cortical and muscular rhythmic oscillations in the beta band was taken into account. In fact, it has been demonstrated that the component of the synchronized cortical activity, coupled to synchronous rhythmic motor-unit firing – assessed by surface EMG – within this band, characterizes aspects of cortical control on voluntary movement (Conway et al. 1995; Gross et al. 2000; Kilner et al. 2000), and is generated in the primary motor cortex (Brown et al. 1998; Brown, 2000; Gerloff et al. 2006a). A component of the synchronized cortical activity has been demonstrated coupled to synchronous rhythmic motor-unit firing assessed by EMG surface recordings both in animals and humans (Piper, 1907, 1912; Baker et al. 1997). In monkeys, for example, pyramidal neurons of the primary motor cortex show bursts of oscillations in the beta band during a precision grip task, coupled with the rectified EMG of the active muscles (Murthy & Fetz, 1992). MEG has provided the first direct extra-cephalic measurement of such cortical–muscular oscillation coupling (Conway et al. 1995). This technique has shown systematic cortico-muscular coherence related to the patterns of motor output and sensory input, both in healthy subjects (Gross et al. 2000; Brown & Marsden, 2001; Kristeva-Feige et al. 2002; Tecchio et al. 2006a) and in patients with motor disorders (Brown et al. 1998, 1999; Volkmann et al. 1996; Timmermann et al. 2003; Kristeva et al. 2004). The corresponding functional constraint to obtain the M1 source was:
|
| (5) |
is a function of frequency 
, obtained for each 
as the amplitude of the cross-spectrum between the M1 source signal and the rectified EMG, normalized by the root mean square of the power spectral densities of these two signals; 
(
) is the frequency point corresponding to a coherence amplitude of 50% of the maximal value between [13.5–33] Hz – called 
– before (after) 
. As for S1a and S1b, M1 was obtained by multiplying it by the Euclidean norm of its weight vector 
(
such as 
, with 
). The FSS algorithm is flexible allowing exploit of different constraints, also not differentiable as simulated annealing is the optimization procedure. The functional constraint could be defined suitably for the experimental conditions in use. In our examples, the static visual stimuli induce a power increase in the gamma band sustained along the stimuli duration, a response more evident than the visual evoked field. This latter point can be exploited in the functional constraint if flashing lights or pattern reversal are used as visual stimuli. For the primary motor area identification, the motor task we used was an isometric contraction for periods of about 20 s, inducing a significant synchronization between primary motor and electromyographic activities in the beta band and the constraint was chosen accordingly (maximization of MEG–EMG coherence). For this area identification the motor-related fields or potentials would be the suitable functional constraint if the motor task were the abrupt repetitions of a body district movement.
FS evaluation. The estimated FSs in all cases contained practically all the required signal features, demonstrating the ability to describe the dynamics of different primary cortical network responsiveness, primary visual sustained activity or primary motor synchronization phenomena (Fig. 3).
|
|
We give these examples to show the FS potential to be obtained exploiting a specific functional requirement in a specific time period and to be afterwards used to describe the behaviour of the neuronal pool identified with the specific functional property, in different experimental stages. In relationship with BCI, this property could be exploited in two main ways. The weight vector obtained in one session can be used in successive recordings. In this regard, our experience (authors' unpublished data) indicated that a suitable description is obtained if similar conditions occur (weight vectors extracted from similar experimental sessions in different days display correlations higher than 0.91); it would be of high interest to confirm these preliminary positive findings in FS components extracted from a controlled laboratory setting, later applied to a noisy real-world environment, where competing processes should be expected. The second opportunity is to use the weight vector to isolate the activity from a specific neuronal pool, in this way being more sensitive to volitional modulations of the activity from those areas. For example, the selection of the activity from primary visual areas by the above described functional constraint, could allow more sensitivity to volitional visual imagery.
ICA artifact removal. Muscle contraction on the cephalic district as well as eye movements produce electromagnetic signals which can be recorded from scalp electrodes with amplitudes that can easily overwhelm the brain generated signals, particularly on frontal, temporal and occipital regions (McFarland et al. 1997; Anderer et al. 1999; Croft & Barry, 2000; Goncharova et al. 2003). More importantly, the EMG signal from frontal muscles can imitate the frequency of the mu and beta rolandic rhythms and the electro-oculographic (EOG) signal and blinking can resemble the fronto-central theta rhythms. Although FSS could in principle be successful even in the presence of disturbance signals, if they are uncorrelated with the required functional characteristics expressed in the functional constraint, on some occasions, artifactual activities could be as higher as 10–100 times stronger than the signal of interest or be in phase with it (Goncharova et al. 2003). In these cases, it could be very helpful to previously identify and discard these artifacts. ICA was proven to be an efficient procedure to remove artifactual activity avoiding trails exclusion (Vigario et al. 1997; Ziehe et al. 2000, 2001; Cao et al. 2000; Delorme & Makeig, 2004). We introduced a suitable strengthening and simplification of ICA preprocessing data analyses, through an automatic detection system of artifactual components (ICs), based on statistical and spectral ICs characteristics (Barbati et al. 2004). Moreover, the procedure allows recovery of part of the non-artifactual signals possibly lost by the blind mechanism, via a control cycle on the difference between original data and those reconstructed using only ICs automatically retained. This step, after automatic pruning, seems to be a suitable way to render negligible the risk of loose non-artifactual activity when applying BSS methods to real data. An emblematic case of artifacts 50–100 times larger than the signal of interest comes from MEG fetal recordings, where the cerebral activity of the fetus is largely overwhelmed by the magnetic signal generated by mother heartbeat and partly by the fetal one. In this case, we developed an ad hoc functional selection procedure of ICs, removing the maternal and fetal cardiac activities, so disclosing the fetal auditory responses to the external sound stimulation (Porcaro et al. 2006; Fig. 5).
|
FSS comparison with other source identification methods. The main difference between FSS and the other source identification methods, ranging from inverse problem-solving algorithms (single and multiple dipoles: Scherg & Berg, 1991; Multiple Signal Classification (MUSIC): Mosher et al. 1992; recursively applied and projected-MUSIC (RAP-MUSIC): Mosher & Leahy, 1999; minimum norm estimates: Hämäläinen & Ilmoniemi, 1994; Low resolution brain electromagnetic tomography (LORETA): Pascual-Marqui et al. 1995) to spatial filtering like beam forming (for example synthetic aperture magnetometry, (SAM): Vrba & Robinson, 2001), is that no information about the physical relationship between cerebral source generators and generated field distribution is taken into account. Separated FSs provide the source activities in time and the spatial distribution of the field they generate, from which appropriate modelling can to be used to solve the inverse problem to know the source position. The solution of the inverse problem theoretically provides in one go both the source position and its time evolution. Unfortunately, on one side it is hill posed and adjunctive information is to be added, chosen properly time by time. On the other side the solution is based on the relationship, described by the Maxwell equations, between the current distribution and that of the generated field. This relation depends on physical properties, i.e. the shape of conductor volume (the head), the distribution and cytoarchitecture of the cerebral regions (place of the currents generating the field), the geometrical characteristics and conducibility values of the extra-cerebral tissues (cerebrospinal fluid, meninxs, skull, scalp) none of which are known with precision. As a consequence, the inverse problem solution is based on information less accurate provided by the electrophysiological techniques, while FSS algorithms do solve the source identification problem using the most accurate information, i.e. the statistical temporal frequency properties of the signal. As we said before, once the source has been identified, to know its position the proper inverse problem solution can be calculated. In many cases, the scientific interest is in the morphological and temporal characteristics of the signal and its modulation in the different experimental conditions, and the inverse problem solving step is not necessary. Whenever the source position is of specific interest, the advantage is in applying localization algorithms having isolated the field distribution generated by the specific source only.
Section 2. Hand sensorimotor organization and its changes in stroke patients
Many years of research in functional imaging has documented that the ultimate factor sustaining recovery is, in parallel with the ‘awakening’ of neurons in the perilesional zone of ischaemic penumbra (Heiss & Graf, 1994), the neural ability to change the properties of the activation induced by inputs from peripheral receptors and/or other brain areas. This ability is called cerebral plasticity (for review see Calautti et al. 2003; Rossini et al. 2003; Hummel & Cohen, 2005). When setting up BCI feature extraction, these cerebral reorganizations should be taken into account, especially in the electrophysiological counterpart, since EEG and MEG are the brain signal monitoring systems best suited for most realistic BCIs. In particular, MEG is especially suitable in poststroke studies, because the morbid tissue near the cerebral generators has minimal effects on the scalp distribution of the magnetic field (Huang et al. 1990; Maclin et al. 1994). BCI feature extraction is based on two main steps: spatial identification of the interesting signal sources and the time–frequency characterization of this signal. To guide BCI applications, it is expected that the activity from the areas positively contributing to the clinical recovery are the most appropriate and informative. In patients affected by a monolateral ischaemic lesion in the middle cerebral artery territory inducing a sensorimotor impairment of the upper limb, a diverse contribution to the clinical recovery by several area reorganizations in the affected (ipsi-lesional ILH) and the unaffected hemisphere (contra-lesional CLH) is going to be recognized. The activation of CLH non-primary motor areas was reported to decrease proportionally with the improvement of motor performance (Calautti & Baron, 2003; Schaechter, 2004) and CLH SM1 activations were commonly considered as a marker of poor recovery (Loubinoux et al. 2003; Calautti et al. 2003). Conversely, increasing findings support a positive role of non-primary ILH areas in patients recovering incompletely (Johansen-Berg et al. 2002; Loubinoux et al. 2003; Fridman et al. 2004; Gerloff et al. 2006b). Moreover, a similar hypothesis was suggested for the recruitment of ILH areas excessively asymmetrical from homologous in the CLH (Pineiro et al. 2001; Calautti et al. 2003; Thickbroom et al. 2004). On the basis of the symmetrical organization of primary hand areas, demonstrated by different techniques in healthy subjects (Rossini et al. 1994b; Puce et al. 1995; Hallet, 1996; Tecchio et al. 1997; Wikstrom et al. 1997), the comparison of the ILH to the CLH in mono-hemispheric stroke patients allows a sensitive procedure to identify reorganization phenomena in the damaged cerebral regions. On this basis, an ad hoc procedure was previously introduced, for the identification of excessive interhemispheric asymmetries of primary sensory hand areas (Tecchio et al. 1997, 2000b, 2005b). In these studies, a galvanic median nerve stimulation was used, a protocol especially suitable in patients because it equals the input to the two hemispheres. It seems, for this reason, more appropriate in order to detect interhemispheric recruitment asymmetries of homologous areas, with respect to active motor tasks (Kotani et al. 2004) where the functional demand to control the paretic and non-paretic hand cannot be equalled. In patients in stabilized conditions (more than 1 year from symptoms onset), primary sensory cortical hand areas topographical modifications and responsiveness amplitudes were observed (Rossini et al. 1998, 27 patients; Rossini et al. 2001, 17 patients), with brain areas outside the normal boundaries and usually not reached by a dense sensory input from the opposite hand acting as somatosensory hand centres. These mechanisms were linked with the hand sensorimotor recovery (Rossini et al. 1998, 2001; Tecchio et al. 2006d; Altamura et al. 2007). A deeper knowledge of brain mechanisms sustaining recovery in stroke patients who do not achieve a normal neurological function by standard procedures, is of utmost importance, since this group could best benefit from adjunctive rehabilitative-stimulating interventions, like BCI-mediated procedures.
Once spatial properties of interesting signal sources have been identified, BCI will exploit their time–frequency characteristics. When studying rest activity in postacute and chronic phase, perilesional delta activity was commonly observed (Butz et al. 2004, 23 patients). Moreover, frequency-selective alterations related to specific dysfunctions were found: global clinical status was mostly impaired in patients with increased total and slow band activity powers, whereas hand functionality was mostly disrupted in patients with a reduction of high-frequency rhythms (Tecchio et al. 2006c, 56 patients; Tecchio et al. 2006c, 32 patients).
While all the above mentioned studies are most suitable to characterize recruitment changes in patients, the organization of the movement control remains the ability the BCI tools should help to recover. Motor imagery is a protocol design also suitable for patients with more severe movement impairment. It is clearly indicated by the literature that the brain activity subtending motor imagery is modified in patients (Sharma et al. 2006). In healthy subjects, movement imagery can focus specific facilitation on the prime-mover muscle for the mentally simulated movement (Rossi et al. 1998a) and mental simulation affects spinal motoneuronal excitability as well, although it has been proved that the main effect takes place at cortical level (Rossini et al. 1999). Reviews considering motor imagery in healthy subjects and in patients with stroke – which may disrupt the motor imagery network – suggested the encouraging effect of motor imagery training on motor recovery after stroke (Lotze & Cohen, 2006; Sharma et al. 2006). While in healthy volunteers, robust activation of the non-primary motor structures, but only weak and inconsistent activation of M1 occurs during motor imagery, in patients with stroke, the cortical activation patterns are proved to involve M1 in both the affected and unaffected hemispheres (Cicinelli et al. 2006). These results indicate that if an appropriate methodology is implemented, motor imagery may provide a valuable tool to access the motor network for BCI feature extraction in stroke patients.
As mentioned above, the computer-mediated visual feedback presents problems for fast applications, while the haptic feedback and electro-tactile sensations present more realistic timing. All studies devoted to the organization of the sensory and proprioceptive inflow underline that a successful motor control relies on the continuous and reciprocal exchange of information between activities of motor areas involved in the task program execution and those elaborating proprioceptive sensory information (Terao et al. 1999; Scott, 2004). As a matter of fact, subjects with complete deprivation of proprioceptive feedback (such as in some cases of peripheral neuropathy) can move their limbs using visual feedback (e.g. by looking at the limb), but their movements are typically sluggish, coarse, require substantial mental concentration and attention, and corrections are therefore delayed and often cause other mistakes (Sainburg et al. 1998; Gordon et al. 1995).
Short-term influences on motor cortical organization have been demonstrated by modification of sensory input both in animal experiments and in healthy humans by using several types of technology for functional brain imaging (Brazil-Neto et al. 1992; Rossini et al. 1994a; Sadato et al. 1995; Kristeva-Feige et al. 1996). Mechanisms of rapid anaesthesia-related perturbation with changes occurring at multiple levels of sensory system somatotopy were proven (Nicolelis et al. 1993), underlying the intimate relationship between primary sensory and motor regions. Experimental findings speak in favour of a significant role played by the tonic sensory flow from the skin receptors and from the phalangeal joint receptors in energizing the corticospinal tracts governing muscles (Johansson & Westling, 1987; Rossini et al. 1996a, b; Rossi et al. 1998b). For BCI applications, such evidence indicates the need to substitute with compensating feedback mechanisms the absence of the physiological one, to obtain a satisfactory movement control.
To optimize knowledge on proprioceptive information properties in the region dedicated to hand control, we investigated the primary sensory and motor cortices interactions, obtaining a simultaneous assessment of sensory cortex activity modulation due to movement and of motor cortex activity modulation due to sensory stimulation (Tecchio et al. 2006b; Fig. 6). The introduced protocol is repeatable and suitable for patients, thus it is suitable to investigate patients eligible for BCI applications. In fact, primary sensory cortical excitability is indexed by the most repeatable and subject attention-independent components of brain response to the median nerve stimulation at the wrist (Hari & Kaukoranta, 1985; Allison et al. 1991). A relative high-frequency nerve stimulation (around 2 Hz) was used to reduce the recording time. The primary motor cortex contribution to movement control is indexed by cortico-muscular coherence (Brown et al. 1998) requiring a motor task both as simple and common as possible, thus most suitable in patients with upper limb impairments. In fact, the experimental contraction is preserved to extremely severe movement impairments, e.g. in diseases affecting the hand motor control, like stroke and Parkinson's disease or hand dystonia. As described in Section 1, the ad hoc FSS was developed to extract S1 and M1 source activities (Porcaro et al. 2007), allowing characterization of the spectral properties, the functional involvement of S1 and M1 and their coupling in the above described sensorimotor tasks (Fig. 6). It could be very useful to reveal alterations in pathological conditions and setting up feedback control systems for BCI use.
|
Our review presents innovative methods expected to enhance BCI feature extraction from EEG or MEG signals, for applications devoted to improve hand control. In particular, the provided experimental set-up and analysis tools could be used to ameliorate a BCI system. In fact, they increase the signal-to-noise ratio by removing non-cerebral artifacts even when they are many times (10–100) larger than the signal of interest. Moreover, they offer a procedure to extract exclusively the sources of interest, obtaining their time course along with the whole experimental session on the basis of appropriate functional requirements taking place in specific time periods. Since the FSS procedure is suitable for single trial analysis, it is promising for online developments – increasing BCI throughput and reducing latency. For applications in patients, information about the recruitments of unusual areas to control the hand were provided, useful for both protocol design definition and interesting signal extraction. Finally, information about primary sensory and motor area relationship was introduced, as a first step to use the haptic feedback and electrotactile sensations which have been indicated as the most compelling for BCI applications aiming at improving hand movement control.
|
References |
|---|
|
TopAbstractIntroductionReferences |
|---|
Allison T, McCarthy G, Wood CC et al. (1991). Potentials evoked in human and monkey cerebral cortex by stimulation of the median nerve. A review of scalp and intracranial recordings. Brain 114, 2465–2503.
Altamura C, Torquati K, Zappasodi F et al. (2007). fMRI-vs-MEG evaluation of post-stroke interhemispheric asymmetries in primary sensorimotor hand areas. Exp Neurol in press.
Anderer P, Roberts S, Schlögl A et al. (1999). Artifact processing in computerized analysis of sleep EEG – a review. Neuropsychobiology 40, 150–157.[CrossRef][Medline]
Baker SN, Olivier E & Lemon RN (1997). Coherent oscillations in monkey motor cortex and hand muscle EMG show task-dependent modulation. J Physiol 501, 225–241.[CrossRef][Medline]
Barbati G, Porcaro C, Hadjipapas A et al. (2007). Functional source separation applied to induced visual gamma activity. Hum Brain Mapp in press.
Barbati G, Porcaro C, Zappasodi F et al. (2004). Optimization of an independent component analysis approach for artifact identification and removal in magnetoencephalographic signals. Clin Neurophysiol 115, 1220–1232.[CrossRef][Medline]
Barbati G, Sigismondi R, Zappasodi F et al. (2006). Functional source separation from magnetoencephalographic signals. Hum Brain Mapp 27, 925–934.[CrossRef][Medline]
Birbaumer N, Ghanayim N, Hinterberger T et al. (1999). A spelling device for the paralysed. Nature 398, 297–298.[CrossRef][Medline]
Brasil-Neto JP, Cohen LG, Pascual-Leone A et al. (1992). Rapid reversible modulation of human motor outputs after transient deafferentation of the forearm: a study with transcranial magnetic stimulation. Neurology 42, 1302–1306.
Brown P (2000). Cortical drives to human muscle: the Piper and related rhythms. Prog Neurobiol 60, 97–108.[CrossRef][Medline]
Brown P, Farmer SF, Halliday DM et al. (1999). Coherent cortical and muscle discharge in cortical myoclonus. Brain 122, 461–472.
Brown P & Marsden JF (2001). Cortical network resonance and motor activity in humans. Neuroscientist 7, 518–527.[Abstract]
Brown P, Salenius S, Rothwell JC et al. (1998). Cortical correlate of the Piper rhythm in humans. J Neurophysiol 80, 2911–2917.
Butz M, Gross J, Timmermann L et al. (2004). Perilesional pathological oscillatory activity in the magnetoencephalogram of patients with cortical brain lesions. Neurosci Lett 355, 93–96.[CrossRef][Medline]
Calautti C & Baron JC (2003). Functional neuroimaging studies of motor recovery after stroke in adults: a review. Stroke 34, 1553–1566.
Calautti C, Leroy F, Guincestre JY et al. (2003). Displacement of primary sensorimotor cortex activation after subcortical stroke: a longitudinal PET study with clinical correlation. Neuroimage 19, 1650–1654.[CrossRef][Medline]
Cao J, Murata N, Amari S et al. (2000). Single-trial magnetoencephalographic data decomposition and localization based on independent component analysis approach. IEICE Transactions Fundamentals Electronics, Comms Computer Sci 9, 1757–1766.
Cichocki A & Amari SI (2002). Adaptive Blind Signal and Image Processing. John Wiley & Sons, Chichester, UK.
Cicinelli P, Marconi B, Zaccagnini M et al. (2006). Imagery-induced cortical excitability changes in stroke: a transcranial magnetic stimulation study. Cereb Cortex 16, 247–253.
Conway BA, Halliday DM, Farmer SF et al. (1995). Synchronization between motor cortex and spinal motoneuronal pool during the performance of a maintained motor task in man. J Physiol 489, 917–924.[Medline]
Craelius W (2002). The bionic man: restoring mobility. Science 295, 1018–1021.
Croft RJ & Barry RJ (2000). Removal of ocular artifact from the EEG: a review. Neurophysiol Clin 30, 5–19.[CrossRef][Medline]
Culpepper BJ & Keller RM (2003). Enabling computer decisions based on EEG input. IEEE Trans Neural Syst Rehabil Eng 11, 354–360.[CrossRef][Medline]
Del Gratta C, Pizzella V, Tecchio F & Romani G-L (2001). Magnetoencephalography – a non invasive brain imaging method with 1 ms time resolution. Rep Prog Phys 64, 1759–1814.[CrossRef]
Delorme A & Makeig S (2004). EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. J Neurosci Meth 134, 9–21.[CrossRef][Medline]
Flor H, Nikolajsen L & Staehelin Jensen T (2006). Phantom limb pain: a case of maladaptive CNS plasticity? Nat Rev Neurosci 7, 873–881.[Medline]
Fridman EA, Hanakawa T, Chung M, Hummel F, Leiguarda RC & Cohen LG (2004). Reorganization of human premotor cortex after stroke recovery. Brain 127, 747–758.
Gastaut H (1952). Etude electrocorticographique de la reactivite des rythmes rolandiques. Rev Neurol 87, 176–182.[Medline]
Gerloff C, Braun C, Staudt M et al. (2006a). Coherent corticomuscular oscillations originate from primary motor cortex: evidence from patients with early brain lesions. Hum Brain Mapp 27, 789–798.[CrossRef][Medline]
Gerloff C, Bushara K, Sailer A et al. (2006b). Multimodal imaging of brain reorganization in motor areas of the contralesional hemisphere of well recovered patients after capsular stroke. Brain 129, 791–808.
Goncharova II, McFarland DJ, Vaughan TM et al. (2003). EMG contamination of EEG: spectral and topographical characteristics. Clin Neurophysiol 114, 1580–1593.[CrossRef][Medline]
Gordon J, Ghilardi MF & Ghez C (1995). Impairments of reaching movements in patients without proprioception. J Neurophysiol 73, 347–360.
Graimann B, Townsend G & Pfurtscheller G (2006). Brain-computer communication – A brief introduction (Technical Report). BCI-Info, BCI Laboratory, Technische Universität Graz, Graz, Austria.
Gross J, Tass PA, Salenius S et al. (2000). Cortico-muscular synchronization during isometric muscle contraction in humans as revealed by magnetoencephalography. J Physiol 527, 623–631.
Hallet M (1996). Transcranial magnetic stimulation: a tool for mapping the central nervous system. Electroencephalogr Clin Neurophysiol Suppl 46, 43–51.[Medline]
Hämäläinen MS & Ilmoniemi RJ (1994). Interpreting magnetic fields of the brain: minimum norm estimates. Med Biol Eng Comput 32, 35–42.[Medline]
Hari R & Kaukoranta E (1985). Neuromagnetic studies of somatosensory system: principles and examples. Prog Neurobiol 24, 233–256.[CrossRef][Medline]
Hari R, Reinikainen K, Kaukoranta E et al. (1984). Somatosensory evoked magnetic fields from SI and SII in man. Electroencephalogr Clin Neurophyisol 57, 254–263.[CrossRef]
Heeger DJ, Huk AC, Geisler WS et al. (2000). Spikes versus BOLD: what does neuroimaging tell us about neuronal activity? Nat Neurosci 3, 631–633.[CrossRef][Medline]
Heeger DJ & Ress D (2002). What does fMRI tell us about neuronal activity? Nat Rev Neurosci 3, 142–151.[CrossRef][Medline]
Heiss WD & Graf R (1994). The ischemic penumbra. Curr Opin Neurol 7, 11–19.
top, 
, bottom) where the functional constraint, i.e. the FS responsiveness, is maximized. Note that both S1a and S1b well explain the generated field at their respective latencies. B, PSD in the Stimulus\No-stimulus condition. Left: PSD of one representative MEG sensor signal in the occipital region displayed in the frequency window [0, 80] Hz. Centre: retro-projected channel with the estimated FS (MEG_rec). Right: PSD of the original MEG data minus MEG_rec channel. The grey area indicates the frequency interval [20, 70] Hz) where the functional constraint, i.e. the band power difference between Stimulus and No-stimulus conditions, is maximized. C, cortico-muscular coherence during voluntary contraction. The two channels most coherent with electromyographic activity are chosen. Their coherences with the rectified EMG in the frequency window [0, 45] Hz. The confidence limit is indicated (0.015, horizontal dashed line). Left: original data. Centre: retro-projected channels with only M1 (MEG_recM1). The two channels display the same coherence with the EMG signal, as all the channels obtained by retro-projecting only one FS display the same time evolution, unless a multiplicative factor and the coherence is independent from the signals amplitude. Right: original MEG data minus MEG_recM1 channels. The grey area indicates the frequency interval (
) where the FS–muscular coherence (dimensionless) is calculated.
