EEG) Based Neurofeedback Training for Brain- Computer Interface (BCI)

Title Electroencephalography (EEG) based neurofeedback training for brain- computer interface (BCI)

Kyuwan Choi Rutgers University Psychology Department; Computational Biomedicine Imaging and Modeling, Computer Science

Psychology Department (Busch Campus)

152 Frelinghuysen Rd.

Piscataway, NJ 08854

Abstract Electroencephalography (EEG) has become a popular tool in basic brain research, but in recent years several practical limitations have been highlighted. Some of the drawbacks pertain to the offline analyses of the neural signal that prevent the subjects from engaging in real-time error correction during learning. Other limitations include the complex nature of the visual stimuli, often inducing fatigue and introducing considerable delays, possibly interfering with spontaneous performance. By replacing the complex external visual input with internally driven motor imagery we can overcome some delay problems, at the expense of losing the ability to precisely parameterize features of the input stimulus. To address these issues we here introduce a non-trivial modification to Brain Computer Interfaces (BCI). We combine the fast signal processing of motor imagery with the ability to parameterize external visual feedback in the context of a very simple control task: attempting to intentionally control the direction of an external cursor on command. By engaging the subject in motor imagery while providing real- time visual feedback on their instantaneous performance, we can take advantage of positive features present in both externally- and internally-driven learning. We further use a classifier that automatically selects the cortical activation features that most likely maximize the performance accuracy. Under this closed loop co-adaptation system we saw a progression of the cortical activation that started in sensory-motor areas, when at chance performance motor imagery was explicitly used, migrated to BA6 under deliberate control and ended in the more frontal regions of prefrontal cortex, when at maximal performance accuracy, the subjects reportedly developed spontaneous mental control of the instructed direction. We discuss our results in light of possible applications of this simple BCI paradigm to study various cognitive phenomena involving the deliberate control of a directional signal in decision making tasks performed with intent.

Index terms- EEG, Brain plasticity, Neurofeedback training, Motor imagery

Introduction Current experimental paradigms using electroencephalography (EEG) as a tool for basic research in cognitive and psychological sciences suffer from a speed and accuracy tradeoff (Stomrud et al. 2010;Liu et al. 2008). The use of external visual signals has several advantageous features. One of them is the ability to harness Evoked Related Potentials (ERP’s) in response to rich visual input of abstract cognitive content whereby the external input can be precisely parameterized to help infer correlations between the neural signal and the stimulus’ features. However visual signal processing is much too slow to allow for fast, spontaneous performance (Maunsell and Gibson 1992). The complex signal also induces eye strain, compounded with habituation effects. The latter are mostly due to long experimental sessions, necessary to average the noisy output over many repeats and extract the signal. An example of visually driven settings is the use of the P300 as steady state visually evoked potentials (SSVEPs). There, the ERP’s decrease over time and the quality of the signal to control an external device using BCI tends to degrade (Wolpaw et al. 2002; Serby et al. 2005; He et al. 2010). The habituation factor and the increase in the noise to signal ratios introduce considerable delays that reach timescales visual awareness. Such delays interfere with the subject’s ability to spontaneously adapt the neural-feedback signal and attain real time control of the external device. Experiments that use SSVEP have recognized that subjects also tend to feel eye fatigue and even admit to the risk of inducing epilepsy (Middendorf et al. 2000; Wang et al. 2006).

Experimental paradigms using internally driven motor imagery offer some advantages over externally driven ERP’s (Choi 2012). Motor imagery does not require visual search, so eye straining is not a problem. Users can engage in motor imagery for long periods of time and the response signal is immediate. In this context the subjects are unaware of explicit delays between the imagined command and the performance outcome, both attained under latencies that would enable them to attempt the mental control of an external device in real time. However, the motor imagery paradigm also has some drawbacks: (1) accuracy is not as high because a large part of the internal stimuli has to be inferred and cannot be parameterized, (2) it takes longer to train the participant, and (3) a large number of participants cannot control a BCI device through motor imagery alone, even upon training over an extended period of time.

Most BCI studies using motor imagery do offline analysis of the cortical activation with an eye for cortical regions that are well known for certain functions (Ramoser et al. 2000; Li et al. 2010). Because the analysis of the signal is off-line and pre-determined in phrenology-like fashion, the

3 participants have no way of knowing whether the motor imagery that they engaged into helped or hindered their performance. The experimenter can also potentially miss neural information from non-targeted cortical areas that may be of relevance for a given context. In general under such off-line and a priori determined settings subjects cannot engage in error-correction loops to try and adapt their motor imagery to eventually exert control over the external device.

We propose in this study a simple paradigm that combines visual feedback from the subject’s real-time performance with motor imagery in closed loop so as to engage the subject in a co- adaptation process between the external performance and the internal cortical signal. By relying on the internally driven motor imagery we overcome the delays imposed by the visual processing of complex visual imagery. Motor imagery enables us to provide immediate on-line feedback of the performance based on 125ms sliding window of brain activation. Such a continuous, fast-processing window permits the spontaneous co-adaptation between the neural signal and the instantaneous visual feedback of the desired external cursor direction because the subject is unaware of processing and transmission delays, thus avoiding the types of interference that emerge in the abovementioned SSEVP setups. Lastly, although we cannot entirely parameterize the internal motor imagery, we can track the neural activation across all 64 channels and automatically select the features that maximize performance accuracy. In this way, instead of a priori harnessing the neural signal from pre-defined areas, we let a Bayesian sparse probit classifying algorithm automatically select the regions with highest activation to solve this specific intentional task. We discuss the results of employing this non-trivial modification of the BCI settings in the context of possible usage for future basic research in the cognitive and psychological sciences with translational value for clinical use.

Materials and Methods

Overview The EEG signal mainly reflects the superposition of the electrical activity created by the ionic charge oscillation due to postsynaptic potentials of neocortical pyramidal cells (Nunez 1995; Nunez and Silberstein 2000). Thus, a large population of neurons beneath an electrode is superimposed to the measureable EEG but relatively little spatial information can be derived from a single channel of EEG. To overcome these problems, studies estimating source currents from EEG channel signals have been steadily conducted in the BCI community. Congedo et al. (Congedo et al. 2004) constructed the first LORETA (low-resolution electromagnetic

4 tomography) tomographic neurofeedback system using source current estimation and conducted a study enhancing the low beta (16-20 Hz) and suppressing the low alpha (8-10 Hz) component of the anterior cingulate cognitive division (ACcd) which is related to attention process. Grave de Peralta Menendez et al. (Grave de Peralta Menendez et al. 2005) introduced a source reconstruction method into an existing motor imagery based BCI field. By using the ELECTRA inverse solution (Grave de Peralta Menendez et al. 2000), they demonstrated the utility of the source reconstruction. Qin et al. (Qin et al. 2004) used single-equivalent-dipole model, and Kamousi et al. (Kamousi et al. 2005) employed two-equivalent-dipole model to classify left and right hand motor imagery. They obtained about 80 % accuracy on it. Furthermore, the use of motor imagery proved useful when Noirhomme et al. (Noirhomme et al. 2008) reconstructed the sources of 400 dipoles and classified motor imagery using frequency power change and Bereitshaft potential as the features.

In the present study, source currents over 2,240 vertexes were estimated from EEG signals of 64 channels through a hierarchical Bayesian method introducing a hierarchical prior (Sato et al. 2004; Batenburg et al. 1994) that can effectively incorporate both structural and functional MRI data. In this hierarchical Bayesian method, the variance of the source current at each source location is considered an unknown parameter and estimated from the observed EEG data. Prior information is also incorporated using the hierarchical Bayesian method (Neal 1996). The fMRI information was imposed as prior information on the variance distribution rather than the variance itself so as to give a soft constraint on the variance. After estimating source currents from the EEG signals, a sparse probit classifier (Balakrishnan and Madigam 2008; Krishnapuran et al. 2005) classifies the left and right hand motor imagery. By using source currents estimated from the EEG signals as features, the number of features can be increased, and among the increased features, the sparse probit classifier automatically selects only useful features in classification. Therefore, it is very robust against the overfitting problem, and the accuracy of the classification can be improved. Furthermore, by using here source currents as features instead of response variance, the participants get an immediate feedback of their intentions rather than an estimate of a byproduct signal.

Participants Five healthy right-handed participants (aged: 26.40 (±6.80), four men and one woman) participated in the experiment. All participants submitted written consent before starting the experiment. The participants did not have any prior experience with BCI. Compensations were paid depending on the results of the experiment.

Experiments The study consisted of three main phases:

(1) The measurement of fMRI activities during actual movement of the fingers in an instructed direction (e.g. up or down) and the use of this information as priors in the estimation of cortical activity from the EEG signals.

(2) The performance of the motor imagery task without any feedback (session1).

(3) The performance of the motor imagery task while obtaining real-time feedback (sessions 2- 4).

1) fMRI experiment

[Figure 1] fMRI task. During 15 s, the participant moves the left or right index finger in the instructed direction every 1 s. The monitor then goes blank and the participant takes a break for 15 s while watching the monitor.

Figure 1 shows the fMRI task used to collect fMRI data as prior information to estimate cortical activity. One trial consisted of the execution task in which the participant moves the left or right index finger in the instructed direction (e.g. up or down) every 1 s depending on the direction of the arrow appearing on the monitor. This is followed by a resting period in which the participant takes a break for 15 s. Each participant conducted 24 trials of the fMRI task. The fMRI activity when participants take a rest (rest periods) was subtracted from the fMRI activity when participants moved their fingers (execution periods).

2) Experimental task without feedback

[Figure 2] Experimental task without feedback.

The EEG signals were measured in the state in which no feedback was present. The task of the experiment was as follows: a beep sounds to begin the trial in which a cross mark appears on the monitor (see Figure 2). There is fixation at the cross for two seconds, then, an arrow indicating left or right appears on the monitor. The participant performs motor imagery by imagining moving the left or right hand, depending on the direction of the arrow for 4 seconds. Then, the monitor goes blank, and the participant takes a break between 3 and 4 seconds. This is considered one trial. Thirty such trials are considered as one set. One participant conducted a total of 7 sets of experiment within one experimental session without feedback.

3) Experimental task with feedback

[Figure 3] Experimental task with feedback.

The EEG signals measured in the first session conducted without feedback are the baseline input data to an inverse filter to obtain the parameters of a classifier. Then, the EEG signals are measured while giving performance feedback visually. The visual feedback consists of the cursor motion according to the neural activity and the probability of selecting the instructed direction using the obtained inverse filter and classifier. The cursor changes color from white to blue according to the probability value of selecting the instructed direction.

The experimental task using feedback is similar in structure to the experimental task without feedback. The difference is that while the participant performs motor imagery, the probability of the classification result is plotted as a colored-bar on the monitor and refreshed every 125 ms. In this way the participant had access to a read-out from the most current state of his/her neural signal (see Figure 3). There is one session of training without feedback and 3 sessions of feedback based training.

Estimation of cortical activities from EEG signals To estimate cortical activities from EEG signals, an inverse filter in equation (1) was used. By multiplying real-time EEG signals to the obtained inverse filter (L) as in equation (1), it is possible to estimate cortical activity with high speed.

1  1  1  1  1 LGGGI()(),           M (1) 1 J()()() t L  E t

Here, E(t) represents measured EEG signals. J(t) denotes the estimated cortical activities. G represents the lead field matrix which was obtained by using the Sarvas’s equation [25] to relate source locations with EEG electrodes. The boundaries between brain, skull and scalp were generated by using the Curry 5 software (Compumedics, U.S.A). Here, the relative conductivities of the brain, skull and scalp are 1,0.0125 and 1. IM represents an identity matrix of M-by-M (M:number of sensors), β corresponds to the inverse of the noise variance of the

1 observed EEG signals.   denotes the source covariance matrix, and is calculated as

11 1    diag() . Here,  represents the source current variance, which is considered unknown parameters in this study and estimated from the measured EEG data by applying a hierarchical prior on current variance.

To obtain an inverse filter, EEG signals were measured at 256 Hz sampling rate from 64 channels in the first session, and a baseline correction (-1~0 s) was taken. Then, the EEG signals were band-pass filtered between 0.1 and 50 Hz using a fifth-order Butterworth filter. To remove any potential artifacts caused by eye movements, a two-step method was applied. The first step is achieved by allocating extra dipoles on the eyes. The cortical activities on the eyes were estimated using these dipoles. In the second step, the EEG signals that correlated with eye movements were estimated from electrooculography (EOG) signals by linear regression and subtracted from the measured whole EEG signals to minimize the effects of eye movements. After obtaining an inverse filter L using the EEG signals measured in the first session, cortical activities were estimated in real-time from the second session by multiplying the obtained inverse filter L to the EEG signals obtained by sliding windows with a length of one second and a moving interval of 125 ms. After averaging the estimated cortical activities for 1 s (see equation (2)), the averaged cortical activities were used as the input of a sparse probit classifier.

1 Nsample Jii  J() t , (2) Nsample t1

In equation (2), Nsample is 256, and i indicates i-th current dipole.

Estimation of current variance

In this study, the current variance  1 was estimated by the Automatic relevance determination (ARD) hierarchical prior (Neal 1996),  P(()|,) J t  exp[  J () t  A  J ()] t 2

Pr(i ) (  i | 00 i , ) , (3) 1 P()   where  is the inverse noise variance of the observed EEG signals, A=diag( ), and is an I- by-1 vector whose component i is the inverse current variance corresponding to the i-th current dipole.  represents the Gamma distribution with mean 0i and degree of freedom r0 .

Intuitively, the hyper-parameter represents confidence of the hierarchical prior information. A

1 prior current variance v00ii  represents the prior information on current intensity. For large

9 and small v0i , estimated current Jti () tends to be large and small, respectively. These values were determined from the fMRI information:

ˆ 2 v00i v base ( m  1)  v base  ( t i ) , (4)

ˆ where ti is a normalized T-value on the i-th vertex. Normalized T-values are computed by dividing the original T-values by the maximum of those T-values (thus ranging from 0 to 1). vbase is a baseline of the current variance, which is estimated from the pre-movement interval (1.0 s ~ 0.5 s before the movement initiation) of the EEG data by a Bayesian minimum norm estimation. A variance magnification parameter m0 , which is the other hyper-parameter, specifies the scaling between the current variances in the baseline and task periods. =100 and r0 =10 were used.

Due to the hierarchical prior, the estimation problem becomes nonlinear and cannot be solved analytically. Therefore, the Variational Bayesian (VB) method (Attias 1999; Sato 2001) was employed. In the VB method, J(t),  and  are iterately updated until convergence.

Classification of left and right hand motor imagery from the estimated cortical activities If the whole cortical activities estimated over 2240 vertexes from 64 channels of EEG were used as the input of a classifier, it easily makes the over-fitting problem. Therefore, a Bayesian sparse probit classifier with the automatic relevance determination (ARD) prior (Shevade and Keerthi 2003) was used to avoid the over-fitting problem and increase generalization ability. To classify the left and right hand motor imagery from the cortical activities estimated over 2240 vertexes, equation (5) was used.

Nsource x()() t wii  J t  w0 , (5) i1

Here, Jti ()represents the average of the estimated cortical activity for one second, x(t) is the result of the left or right hand motor imagery, and wi denotes the weights of the classier which is decided by the sparse probit classifier. Nsource represents the number of useful cortical activities in classification selected by the sparse probit classifier. Since the sparse probit

10 classifier automatically selects only useful cortical activities in classifying the left and right hand motor imagery, it is a very strong tool against the over-fitting problem. As each session progresses, the inverse filter and weights of the classifier were updated using the EEG data measured in each session, and the participants were trained using the updated inverse filter and classifier at each session.

Results

A. fMRI Data Analysis Functional imaging data were analyzed using SPM2 (Wellcome Department of Cognitive Neurology, London, UK). The first four volumes of images in each session were discarded to allow for T1 equilibration and then spatially aligned the data to the first remaining volume. The data were spatially normalized to the Montreal Neurological Institute (MNI; Montreal, Quebec, Canada) reference brain and re-sliced to a 2-mm isotropic voxel size. Data were smoothed spatially with a Gaussian kernel of 6 mm full-width at half-maximum (FWHM). Voxel time series were high-pass filtered with a cutoff frequency of 0.002 Hz and low-pass filtered with a cutoff frequency of 0.25 Hz.

Statistical analyses were performed for each participant. Boxcar functions modeled execution periods and rest periods. They were convolved with the canonical hemodynamic response function in SPM2 to yield regressors in a general linear model. A parameter was estimated for each regressor by the least-squared method. T-statistics were used for comparison between the estimated parameters (execution – rest) to yield a t-value for each voxel. The yielded statistical parametric map was used as prior information in the estimation of EEG source currents.

[Figure 4] The fMRI information when the participant actually moves the index finger up and down every 1 s.

Figure 4 shows the fMRI information extracted by SPM2 (UCL, UK). When the participant moves the left index finger up and down, strong fMRI activity appeared in the right motor area. Additionally, when moving the right index finger up and down, the left motor area is activated, but the activation is weaker than that of the left finger movement, most probably due to the participant being right-handed. After combining the fMRI information of the left index finger movement with that of the right index finger movement, it was used as prior information to estimate cortical activity.

B. Participant 1’s 1st session results conducted without feedback

[Figure 5] Estimated cortical activity of the first session conducted without feedback (average of 105 trials for the left and right motor imagery [time interval: 0~4 seconds]).

Figure 5 shows the estimated cortical activity of Participant 1 in the first session conducted without feedback. In the left case, we found that Brodmann area (BA) 6R (premotor cortex and supplementary motor cortex) area is strongly activated. On the other hand, in the right case, BA 6L (premotor cortex and supplementary motor cortex) and BA 43L (subcentral area) are activated. By using the estimated cortical activities as the input of the sparse probit classifier, the EEG signals of left and right motor imagery were trained.

[Figure 6] Important features selected by the sparse probit classifier (1st session).

Figure 6 shows the important features selected by the sparse probit classifier. In this case, 29 (±3.25) features were selected. In Figure 6, we can see that the selected features are distributed over the entire area of the brain. In this study, seven sets of data were measured within each session. The classifier was trained and tested by the cross validation method. The accuracy of the first session was 53.73 % (±7.17 %). The performance was is almost at chance level.

C. Participant 1's 2nd session result conducted with feedback

[Figure 7] Estimated cortical activity of the second session conducted with feedback.

Figure 7 shows an average of 105 trials of the estimated cortical activities for the left and right cases. In the left case, strong cortical activity is estimated on the motor area and front part of the right side. In the right case, weak activity is estimated on the motor area of the left side, while strong activity is estimated on the right motor area. In this case, the participant showed a strong performance for the left motor imagery trials. However, his performance for the right motor imagery trials was very weak.

[Figure 8] Important features of session 2 selected by the sparse probit classifier.

After training the sparse probit classifier, a total of 30 (±3.30) features are selected (see Figure 8). The selected features are positioned over the entire area of the brain. The accuracy of session 2 was 43.03 % (±7.54 %). The performance accuracy of the session 2 was even lower than that of the first session conducted without feedback.

D. Participant 1's 3rd session results conducted with feedback

[Figure 9] Estimated cortical activity of session 3 conducted with feedback.

In session 3, in the left case, the strong activity located in the front part disappeared, and the motor area became strongly activated (see Figure 9). In the right case, the cortical activity on the left motor area became stronger compared with that of the session 2 and the activity of the right side turned weaker. Through the feedback training, the participant started to find a balance between left and right motor imagery tasks. The accuracy of session 3 was 84.82 % (±6.76 %), a significant improvement over that of session 2.

[Figure 10] Important features of session 3 selected by the sparse probit classifier.

Figure 10 shows the important features selected in session 3. Here, we show that the features with high weights selected by the sparse probit classifier are mainly located in the sensory cortex, thus implying that the participant gets the feel of motor imagery through the sensory cortex. In this session, the participant reported that he could feel the sensation of correct motor imagery on the fingertips.

E. Participant 1's 4th session result conducted with feedback

[Figure 11] Estimated cortical activity of session 4 conducted with feedback.

In session 4, strong activity emerged in the motor area (see Figure 11). This pattern is very similar to the fMRI information measured with actual finger movement. In session 4, the accuracy increased over 96.19 % (±2.30 %).

[Figure 12] Important features of session 4 selected by the sparse probit classifier.

In session 4, we see that important features selected by the sparse probit classifier are mainly located in BA 6 and the prefrontal cortex (see Figure 12).

[Figure 13] Change in cortical activities induced by left motor imagery over the time intervals.

Figure 13 shows changes in cortical activities over time intervals induced by left motor imagery. In the beginning, many areas are activated. However, as time passes, we observe that the right motor and visual areas are activated. One possible reason that the visual area is activated may

16 be that while conducting motor imagery, the participant received visual feedback on the results of his motor imagery. The activity on the right motor area fluctuated; sometimes becoming stronger, and sometimes weaker. Yet, we observed consistent activation of this area all throughout the session.

[Figure 14] Change in cortical activities of Participant 1 in the session 4 over time intervals induced by right motor imagery.

Figure 14 shows changes in cortical activities induced by right motor imagery. In the beginning, the activity on the left motor area was weak. However, we quantified stronger activation as time passed.

F. Participant 2’s result after feedback training

[Figure 15] Estimated cortical activity of Participant 2 in the session 4 conducted with feedback and the selected features by the sparse probit classifier after finishing feedback training.

Figure 15 shows the experimental result of another Participant 2 after finishing the 4th session. When conducting left hand motor imagery, the right side motor area was strongly activated. And, when conducting right hand motor imagery, the left side motor area and prefrontal cortex were activated. In Participant 2’s case, the features automatically selected by the sparse probit classifier were focused on the prefrontal cortex.

[Figure 16] Accuracy of each session of the all five participants.

Figure 16 shows the accuracy of all of the participants participated in the experiment. In the first session, which was conducted without any feedback, the accuracy was about 50%. This is at chance level. For the second session, the participants conducted experimental task while receiving continuous visual feedback. However, there is no significant improvement in the second session. This suggests that in the second session, the participants were not accustomed to the feedback yet. They needed time to adapt to the experimental task. In the third session, the accuracy began to rise and in the fourth session, all participants showed almost perfect performance. One-way repeated measures ANOVA was conducted for each session. The results are F(3,18)=403.42, p<0.01 for subject 1, F(3,18)=308.14, p<0.01 for subject 2, F(3,18)=315.53, p<0.01 for subject 3, F(3,18)=269.55, p<0.01 for subject 4, and F(3,18)=386.48, p<0.01 for subject 5. There was a significant improvement in accuracy for all subjects.

Discussion

A. Which brain area do participants use to conduct hand motor imagery as feedback training progresses?

[Figure 17] The relationship between the averaged accuracy of all of the participants and the averaged number of features selected by the sparse probit classifier.

Figure 17 shows the relationship between the averaged accuracy of 5 participants and the averaged number of features selected by the sparse probit classifier. We see that as accuracy increases, the number of features selected by the sparse probit classifier gradually decreases (One-way repeated measures ANOVA, F(3,19)=388.73, p<0.01 for accuracy, F(3,19)=198.50, p<0.01 for the number of features). When the participants exhibited low accuracy, they used many more brain areas to perform motor imagery than what they utilized when exhibiting high accuracy. This result suggests that in the beginning, they did not know how to correctly perform motor imagery. However, as they get the feeling of how to correctly perform motor imagery

20 through feedback training, the participants used a smaller number of brain areas to perform motor imagery.

[Figure 18] Selected important features after completion of feedback training.

Figure 18 shows the features selected by the sparse probit classifier after the participants completed the fourth session. We see that after feedback training, features located mainly in the prefrontal cortex and BA 6 were selected.

The processing of motor-related information in the brain is thought to follow a progression: first, the urge to move the arm is reflected in the premotor cortex; then the signal shifts to the primary motor cortex via the supplementary motor area. The primary motor cortex, assumed to be the final output portion of the motor-related signals in the brain, send the signal to the muscles through alpha motor neurons of the spinal cord thus producing the physical movement of the arm (Fried et al. 1991). In our study however, once the subjects were performing well above chance level, the prefrontal cortex was automatically recruited. The prefrontal cortex is involved in many aspects of planning and cognitive control. The present results implies that a more abstract representation of the motor imagery signal exists possibly indicating a neural correlate of intentionality in this region. Figure 18 shows BA6 of the prefrontal cortex engaged in the directional control task and poses the question of whether this signal precedes the actual motor command.

[Figure 19] Important features selected by the sparse probit classifier over the time interval while Participant 3 performs motor imagery task.

Figure 19 shows the important features selected by the sparse probit classifier over the time interval while Participant 3 performed motor imagery. Between 0 and 0.2 s, the features in the prefrontal cortex were selected, and from 0.2 s to 4 s, the features in BA 6 were chosen. This result suggests that the participants may use the prefrontal cortex as part of a planning stage for impending motor decision prior to engaging BA4 in the output execution stage.

B. What will the result be of training the participant for a long period of time?

[Figure 20] Long term training for Participant 1 over the 10th session.

Feedback training was continuously done for Participant 1 over the 4th session, in which he showed almost perfect performance (see Figure 20). Participant 1 was trained over an extended period of time, completing 10 sessions. The accuracy of the 10th session is almost the same as that found in the 4th session. There was no apparent improvement in accuracy. However, in the 4th session, the participant mainly used BA 6 to control the bar’s direction. Long term training decreased BA6 activation with a recruitment the prefrontal cortex. Subjects reported that as they mastered the mental control of the cursor they thought less of the direction and felt a spontaneous, rather than a deliberate performance. Unanimously the subjects reported that they moved the bar left or right by intention without explicit thoughts of direction. It is possible that this sensation is accompanied by the recruitment of the prefrontal cortex that we quantified here. The lack of recruitment of BA 4 (motor cortex) suggests a more abstract representation of motion may exist in the prefrontal cortex, initiated here by BA6 with a later shift towards more frontal regions upon further training.

Ojakangas et al. (Ojakangas et al. 2006) conducted a study that revealed that human prefrontal/premotor cortex neurons can provide information about movement planning and decision-making sufficient to decode the planned direction of movement. Our results are congruent with this previous research. In addition to that, we here also tested the blindfolded participants providing auditory commands and found comparable levels of performance accuracy (see video clip in supplementary materials). This suggests that a true abstract amodal signal exists in the prefrontal cortex and that this signal is useful to eventually, spontaneously direct our thoughts.

Conclusion We have shown that a non-trivial modification of the BCI paradigm can automatically recruit specific cortical areas as subjects use motor imagery to master the mental control of an explicit direction. Upon neuro-feedback based training the strength of the signal shifts to BA6 and eventually to more frontal regions of the prefrontal cortex. The transition from deliberate to spontaneous control reported by the subjects coincided with the decrease of activation in BA6 and the shift to maximal strength of the signal to the prefrontal cortex.

The proposed modification to the BCI could be extended to other experiments of basic research in the psychological and cognitive sciences involving deliberate decision making with more complex visual stimuli. For example, in the context of match to sample tasks participants could mentally perform the choice and the transition from deliberate to automatic decisions be precisely quantified in this region. Furthermore, because the task involves executive control, it could also be used in studies involving clinical populations suffering from dysexecutive function (e.g. in patients with dementia (Edin et al. 2009), schizophrenia and ADHD (Lim et al. 2010). In summary we have presented here a non-trivial modification of the BCI task that invites multiple applications of the task to basic science and clinical research.

Acknowledgments

I would like to thank Dr. Elizabeth B. Torres for editing the manuscript.

References Attias H (1999) Inferring parameters and structure of latent variables models by variation Bayes. Proceedings of 15th conference on uncertainty in artificial intelligence, pp. 21-30 Balakrishnan S, Madigam D (2008) Algorithms for sparse linear classifiers in the massive data setting. Journal of machine learning research 9, 313-337 Batenburg P, O’hagan A, Veenstra R (1994) Bayesian discovery sampling in financial auditing: a hierarchical prior model for substantive test sample sizes. The Statistician 43, No. 1, pp99-110

Choi K (2012) Control of a vehicle with EEG signals in real-time and system evaluation. European journal of applied physiology, Feb;112(2):755-66

Congedo M, Lubar J, Joffe D (2004) Low-resolution electromagnetic tomography neurofeedback. IEEE Transactions on neural systems and rehabilitation engineering, vol. 12, No. 4

Edin F, Klingberg T, Johansson P, McNab F, Tegner J, Compte A (2009) Mechanism for top-down control of working memory capacity. PNAS 106, 6802-6807

Fried I, Katz A, McCarthy G, Sass K, Williamson P, Spencer S, Spencer D (1991) Functional organization of human supplementary motor cortex studied by electrical stimulation. J. Neurosci. 11, 3656-3666

Grave de Peralta Menendez R, Andino S, S. Morand, C.M. Michel, T. Landis (2000) Imaging the electrical activity of the brain: ELECTRA. Human brain mapping 9:1-12

Grave de Peralta Menendez R, Andino S, Perez L, Ferrez P, Millan J (2005) Non-invasive estimation of local field potentials for neuroprosthesis control. Cogn. Process, vol. 6, pp. 59-64

He Z, Cichocki A, Xie S, Choi K (2010) Detecting the number of clusters in n-way probabilistic clustering. IEEE Transactions on pattern analysis and machine intelligence, 32(11): 2006-2021 Kamousi B, Liu Z, He B (2005) Classification of motor imagery tasks for brain-computer interface applications by means of two equivalent dipoles analysis. IEEE Transactions on neural systems and rehabilitation engineering, vol. 13, No. 2

Krishnapuran B, Carin L, Figueir redo M, Hartemink A (2005) Sparse multinominal logistic regression: fast algorithms and generalization bounds. IEEE Transactions on pattern analysis and machine intelligence, vol. 27, No.6 Li Y, Kambara, H.,Koike Y., Sugiyama, M. (2010) Application of covariate shift adaptation techniques in brain-computer interfaces. IEEE Transaction on biomedical engineering, 57 issue: 6, 1318-1324 Lim C, Lee T, Guan C, Fung D, Cheung Y, Teng S, Zhang H, Krishnan K (2010) Effectiveness of a brain-computer interface based programme for the treatment of ADHD: a pilot study. Psychopharmacology bulletin, vol. 43, No. 1 Liu T, Shi J, Zhao D, Yang J (2008) The relationship between EEG band power, cognitive processing and intelligence in school-age children. Psychology science quarterly, Volume 50, pp. 259-268 Maunsell J, Gibson J (1992) Visual response latencies in striate cortex of the macaque monkey. Journal of neurophysiology, vol. 68, No. 4 Middendorf M, McMillan G, Galhoun G, Jones K (2000) Brain-computer interfaces based on the steady-state visual evoked response. IEEE Transactions on Neural systems and rehabilitation engineering, Vol. 8, No. 2, pp. 211-214 Neal R (1996) Bayesian learning for neural networks. Springer, New York Noirhomme Q, Kitnew R, Macq B (2008) Single-trial EEG source reconstruction for brain computer interface. IEEE Transactions on biomedical engineering, vol. 55, No. 5

Nunez P (1995) Neocortical Dynamics and Human EEG Rhythms. Oxford, U.K.: Oxford Univ. Press

Nunez P, Silberstein R (2000) On the relationship of synaptic activity to macroscopic measurements: Does co-registration of EEG with fMRI make sense? Brain Topogr., vol. 13, no. 2, pp. 79-96

Ojakangas C, Shaikhouni A, Friehs G, Caplan A, Serruya M, Saleh M, Morris D, Donoghue J (2006) Decoding movement intent from human premotor cortex neurons for neural prosthetic applications. Journal of clinical neurophysiology, vol. 23, No. 6 Qin L, Ding L, He B (2004) Motor imagery classification by means of source analysis for brain- computer interface applications. Journal of neural engineering, 1,135-141

Ramoser H, Muller-Gerking J, Pfurtscheller G (2000) Optimal spatial filtering of single trial EEG during imagined hand movement. IEEE Transactions on rehabilitation engineering, vol. 8, No. 4 Sarvas J (1987) Basic mathematical and electromagnetic concepts of the biomagnetic inverse problem. Phys. Med. Biol., 32(1), 11-22 Sato M (2001) Online model selection based on the variational Bayes. Neural computation 13, 1649-1681 Sato M, Yoshioka T, Kajihara S, Toyama K, Goda N, Doya K, Kawato M (2004) Hierarchical Bayesian estimation for MEG inverse problem. NeuroImage, 23 806-826

Serby H, Yom-Tov E, Inbar G (2005) An improved P300-based brain-computer interface. IEEE Transactions on rehabilitation engineering, volume 13, 89-98

Shevade S, Keerthi S (2003) A simple and efficient algorithm for gene selection using sparse logistic regression. Bioinformatics 19(17) 2246-2253

Stomrud E, Hansseon O, Minthon L, Blennow K, Rosen I (2010) Slowing of EEG correlates with CSF biomarkers and reduced cognitive speed in elderly with normal cognition over 4 years. Neurobiology of aging, vol. 31, Issue 2, pages 215-223 Wang Y, Wang R, Gao X, Hong B, Gao S (2006) A practical VEP-based brain-computer interface. Neural systems and rehabilitation engineering, IEEE Transactions on, Vol. 14, Issue 2 Wolpaw J, Birmbaumer N, McFaland D, Pfurtscheller G, Vaughan T (2002) Brain computer interfaces for communication and control. Clinical Neurophysiology, 113, 767-791