J Comput Neurosci (2009) 27:357–368

DOI 10.1007/s10827-009-0148-4

Model analyses of visual biofeedback training for EEG-based

brain-computer interface
Chih-Wei Chen & Ming-Shaung Ju & Yun-Nien Sun &
Chou-Ching K. Lin

Received: 9 June 2008 / Revised: 20 January 2009 / Accepted: 16 March 2009 / Published online: 9 April 2009
# Springer Science + Business Media, LLC 2009

Abstract The primary goal of this study was to construct a experimental results. The serial changes of these parameters
simulation model of a biofeedback brain-computer interface over the ten sessions, reflecting the effects of biofeedback
(BCI) system to analyze the effect of biofeedback training on training, were analyzed. The model simulation could repro-
BCI users. A mathematical model of a man-machine visual- duce results similar to the experimental data. The group mean
biofeedback BCI system was constructed to simulate a success rate and information transfer rate improved signifi-
subject using a BCI system to control cursor movements. cantly after training (56.6 to 81.1% and 0.19 to 0.76 bits/trial,
The model consisted of a visual tracking system, a thalamo- respectively). All three model parameters displayed similar
cortical model for EEG generation, and a BCI system. The and statistically significant increasing trends with time.
BCI system in the model was realized for real experiments of Extensive simulation with systematic changes of these
visual biofeedback training. Ten sessions of visual biofeed- parameters also demonstrated that assigning larger values to
back training were performed in eight normal subjects during the parameters improved the BCI performance. We con-
a 3-week period. The task was to move a cursor horizontally structed a model of a biofeedback BCI system that could
across a screen, or to hold it at the screen’s center. simulate experimental data and the effect of training. The
Experimental conditions and EEG data obtained from real simulation results implied that the improvement was achieved
experiments were then simulated with the model. Three through a quicker adaptation rate in visual tracking gain
model parameters, representing the adaptation rate of gain in and a larger synaptic gain from the visual tracking system
the visual tracking system and the relative synaptic strength to the thalamic reticular cells. In addition to the purpose of
between the thalamic reticular and thalamo-cortical cells in this study, the constructed biofeedback BCI model can also
the Rolandic areas, were estimated by optimization techni- be used both to investigate the effects of different bio-
ques so that the performance of the model best fitted the feedback paradigms and to test, estimate, or predict the
performances of other newly developed BCI signal pro-
cessing algorithms.
Action Editor: Steven J. Schiff
C.-W. Chen : M.-S. Ju Keywords Brain-computer interface . Biofeedback .
Department of Mechanical Engineering,
Mu rhythm . Thalamo-cortical model
National Cheng Kung University,
Tainan, Taiwan
List of symbols
Y.-N. Sun τ Time constant of visual tracking system
Department of Computer Science and Information Engineering,
a1, a2 Constants in linearized thalamo-cortical
National Cheng Kung University,
Tainan, Taiwan model
D Logical operator for assigning m to F1(s) or
C.-C. K. Lin (*) F2(s)
Department of Neurology,
e Error (difference) between u and y, e=u-z
National Cheng Kung University Hospital,
Tainan, Taiwan F1(s), F2(s) Thalamo-cortical model for C3 and C4
e-mail: [email protected] positions, respectively.
358 J Comput Neurosci (2009) 27:357–368

f nC3 , f nC4
power spectrum of the nth windowed y1 and the brain-computer interface (BCI) (Pfurtscheller and
y2, respectively Berghold 1989; Fabiani et al. 2000; Curran and Stokes
g1, g 2 Gain magnifying m into thalamic reticular 2003). BCI is an artificial interface for extracting a human’s
model of C3 and C4 positions, respectively intention to control external devices, such as computers or
He, Hi Transfer function of excitatory and inhibitory prostheses, and is an assistive system for patients with
post-synaptic potential, respectively motor deficits due to diseases such as stroke and spinal cord
HRe, HRi Corresponding Transfer functions for reticu- injury (Birbaumer and Cohen 2007).
lar cells The magnitude of mu rhythm suppression was highly
HTe, HTi Corresponding Transfer functions for variable among subjects (Vaughan et al. 1996). Some
thalamo-cortical cells subjects did not show significant mu rhythm suppression,
J Cost function in optimization of estimating limiting the applicable range of BCI systems based on mu
model parameters rhythm detection. However, it has been found that subjects
k Proportional gain of visual tracking system without clear movement-induced mu rhythm suppression
m Output of visual tracking system could be trained to achieve increased mu rhythm suppres-
m1, m2 Modulating input to thalamo-cortical model sion and higher BCI success rates. Historically, most
of C3 and C4 positions, respectively studies have utilized the paradigm of biofeedback training
N Number of targets (i.e., the subjects adjusted their next response according to
q Stationary random sensory input to thalamo- their own previous responses). Biofeedback training using
cortical model many different physiological signals has been extensively
Pc3, Pc4 Power ratio of mu rhythm of y1 and y2, studied in the past. Miller (1969) found that, in animal
respectively subjects, visceral activities (such as heart beat and intestinal
Pc3*, Pc4* Power ratio of mu rhythm of y1* and y2*, contraction) could be modified by a reward contingent to
respectively the changes. Kimmel (1974) summarized the results of
QC3, QC4 Summed power in the 8–14 Hz band of y1 many studies indicating that control of autonomic functions
and y2, respectively could be learned in human subjects by instrumental
QT3, QT4 Summed power of y1 and y2, respectively conditioning. Wyrwicka and Sterman (1968) demonstrated
r Adaptation rate of tuning k. that cats learned to produce the sensorimotor rhythm by
RT, RR, RL, Success rate for total, moving rightward, operant conditioning. Fetz (1969) demonstrated in operant
RC moving leftward and holding central conditioning experiments of monkeys that the firing rate of
imaginary tasks, respectively. cortical neurons could be regulated. These earlier studies
S Setting upper and lower limits to BCI output have provided evidence that indicates the self-regulation of
u Desired target of biofeedback training visceral activities (reviewed by Brener 1981) and brain
V Visual tracking system waves (reviewed by Sterman 1981) could be achieved by
w A weighting constant in computing BCI proper forms of biofeedback training. In more recent
output studies, the biofeedback technique has been introduced
x, i, j, n Dummy indices into the training of BCI users. Subjects could be trained to
y1, y 2 Simulated EEG of the modified thalamo- spell words with letter-selection systems controlled by
cortical model at C3 and C4, respectively different BCI systems based on slow cortical potentials
y1*, y2* EEG at C3 and C4 obtained from real (SCP) (Birbaumer 1999) or P300 (Sellers and Donchin
experiments, respectively 2006). Pfurtscheller and Neuper (2001) found that the
z Output of BCI model success rates of classification were improved by biofeed-
z* Cursor movement recorded in experiments back training, and were finally increased to over 90%.
Most of the common evidence that revealed the
beneficial effects of training was in the form of empirical
indices (e.g., success rate, information transfer rate,
1 Introduction response time, EEG feature, or features from functional
MRI images) (Hinterberger et al. 2004; Julien et al. 2005;
The mu rhythm (8–14 Hz) recorded at the primary Meinicke et al. 2003). However, there is no solid theoretical
sensorimotor area is highly correlated with voluntary limb support about how biofeedback training can improve the
movements, and its magnitude is suppressed whether real performance of BCI systems. The lack of such support
or imaginary movements are attempted (Pfurtscheller and impedes the design of better biofeedback training programs
Lopes da Silva 1999). Because of this correlation, mu and signal processing algorithms for BCI systems. Inves-
rhythm has been widely investigated as a control source of tigation of the underlying neural substrate and mechanism
J Comput Neurosci (2009) 27:357–368 359

is often impractical, especially in human subjects. Some model of a thalamo-cortical circuit in order to analyze the
early studies have investigated cortical activities and the effect of biofeedback training on BCI users at the neuronal
regulation of neural firing in animal experiments. In circuit level. We hypothesized that the effects of biofeed-
addition to Fetz (1969) mentioned in the above, Howe back training were achieved by modifying the transmission
and Sterman (1972) found that the sensorimotor rhythm strength between neuronal populations. Our approach was
originated from the ventrobasal relay nuclei of the to first observe the effects of biofeedback training on BCI
thalamus. Although later studies have indicated that performance by real experiments. Then, we optimized the
macroscopic EEG rhythms can originate from the cortex, simulation performance by adjusting the model parameters
the thalamus, or the interplay of the thalamo-cortical circuit to analyze and examine the trends and changes of
(Lopes da Silva et al. 1980), the effects of sensory afferents parameters. Finally, we investigated the effects of parameter
on mu rhythm are relayed through the thalamus in thalamo- modification by systematically changing the parameters and
cortical circuits. Biofeedback training could change EEG performing simulations. The validity of our model is
signals by modulating the activities of neural elements in supported by its performance and its reasonable trend of
thalamo-cortical circuits. Yet, it has not been definitely parameter changes.
established what modification of the thalamo-cortical
circuit is responsible for the effects of biofeedback training
on using the BCI system. 2 Methods
In addition to direct measurement of neural activities,
model simulation is another approach to investigate the This study consisted of two parts, model simulation and
contribution of neural components to the effects of real experiments. A mathematical model of a man-machine
biofeedback. Neuronal population (Wilson and Cowan visual-biofeedback BCI system was constructed to simulate
1972) and global (Wright et al. 2000) models have been a subject using a BCI system to control cursor movements.
proposed to describe and explain the generation of main The BCI system in the model was realized for real
EEG rhythms. While global models emphasize the spatial experiments of visual biofeedback training in eight sub-
relationship among different areas, neuronal population jects. Experimental conditions and EEG data obtained from
models focus more on the relationship among local the real experiments were then simulated with the model.
neuronal populations and their input-output responses. A The serial changes of model parameters over the ten
neural population model was developed by Lopes da Silva sessions, reflecting the effects of biofeedback training, were
et al. (1974) to explain the generation of alpha waves of 8– analyzed.
12 Hz. The authors indicated that the basic mechanism was
the reciprocal excitation of the thalamic rhythmic generator 2.1 Construction of the simulation model
and the cortical neurons. Because the neural circuit
responsible for the mu rhythm is still not exactly known, A block diagram of the proposed model, simulating a
it was hypothesized to have similar components to other subject using a biofeedback BCI system to control cursor
neocortical areas that exhibited rhythmic resting activities movements, is depicted in Fig. 1(a). The model consisted of
of 8–14 Hz (Lopes da Silva et al. 1980). An extended 3 components: 1) an adaptive visual-to-attempt conversion
model (Suffczynski et al. 1999) containing interconnected subsystem (V) translating the feedbacked visual signal to
inhibitory synapses was implemented to simulate spatially mental attempts; 2) a logical operator (D) and a thalamo-
inhomogeneous mu rhythm changes in two neighboring cortical model (TCM) simulating the generation of mu
cortical areas, called focal event-related desynchronization rhythm in the left and right central cortical areas (y1 and y2
(ERD) and surround event-related synchronization (ERS). respectively) and their suppression by movement imagina-
In other words, the authors demonstrated that hand move- tions (m); and 3) a BCI system (BCI) that detected and
ments induced desynchronization of mu rhythm in the processed EEG signals for visual feedback. From the view
cortical representation area of the hand, while they of the control system, the whole model is a position control
simultaneously resulted in the synchronization of mu system, requiring the processed EEG output to match the
rhythm in the foot cortical area. desired target. V is the controller, D and TCM together
Because experimental study of biofeedback effects in comprise the plant, and BCI is the sensor and feedback
BCI is always very time consuming and expensive, a element. In a physiological view, the three blocks (V, D,
simulation model of the biofeedback BCI system has the and TCM) together represent the human component and the
potential to facilitate the development and accelerate the BCI block is analogous to the computer component. The
testing of practical BCI systems. The primary purpose of model simulated the period that a subject was cued to
this study was to construct a simulation model of a perform imaginary movement. Both the desired target and
biofeedback BCI system based on a neuronal population output of BCI shown to the subject by a screen were
360 J Comput Neurosci (2009) 27:357–368

Fig. 1 Block diagrams of the

BCI biofeedback model and the
experimental setup. (a) The
model consisted of the visual
tracking model (V), thalamo-
cortical model (TCM), and ex-
ternal BCI system (BCI) and the
corresponding components and
variables in the experimental
setup; (b) the diagram of TCM.
Symbols and abbreviations are
explained in the text

modeled by u and z respectively. u, equivalent to the visual response of the subject was quicker. The gain (k) was
cue shown to the subject in the real experiments, was a step adapted during sessions of biofeedback training by the
function with initial and step values of 0 and ±1. If u was 0, following rules,
the model simulated a cue appear in the middle of the
screen, if u was 1 or −1, the model simulated the cue of kði þ 1Þ ¼ kðiÞ þ r  kðiÞ  eðiÞ; ð2Þ
right or left arrow appear in the screen respectively. z
simulated the cursor position processed from the simulated where index i denoted the ith moving window, and r was the
EEG signals, y1 and y2. The negative feedback controller V adaptation rate of k. The inputs to TCM were generated by
modeled the visual tracking and visual-to-attempt conver- the logical operator D according to the following logical
sion subsystem, and represented the effort of the subject to rules:
minimize the difference between u and z. The input of the
controller (V) was the difference (e) between u and z (e=u– Dðm; eÞ¼ ½m1 ; m2 
z), and the output (m) of V represented the intention of ¼ ½0; m; if e > 0:1
imaginary movements. V could be written as ð3Þ
or ¼ ½m; 0; if e <
0:1
Lfmg Lfmg k
VðsÞ ¼ ¼ ¼ ; ð1Þ or ¼ ½0; 0; if jej  0:1;
Lfeg Lfu
zg ts þ 1
where L{} was Laplace transformation, τ was the time where m1 and m2 were the inputs to the neural circuits that
constant of normal visual tracking and was chosen as 0.2 s produced EEG signals at electrode positions C3 and C4,
(Poulton 1974), and k was an adjustable gain related to the respectively. D(m, e) modeled the subject’s judgment to
speed of visual tracking. A larger k meant that the tracking move the cursor to the right or left, and assigned the
J Comput Neurosci (2009) 27:357–368 361

modulating input (m) to left or right hemisphere (C3 or C4) EEG of y1 and y2, respectively, and the running index j
accordingly. represented the frequency of the spectrum. Pc3(n) and
The TCM presented in this study was based on the Pc4(n) were the ratios of power of mu rhythm to total power
lumped TCM published by Lopes da Silva et al. (1974) and of y1 and y2. QT3(n) and QT4(n) was the total power,
Suffczynski et al. (2001), and served as a transfer function QC3(n) and QC4(n) were the summed power in the 8–14 Hz
that transformed the intention of imaginary movements into band (the mu rhythm band) of the nth windowed EEG of y1
EEG signals at scalp positions C3 and C4. The TCM block and y2, respectively; w was a constant and was empirically
shown in Fig. 1(a) included two identical subsystems, F1(s) set as 1.25; and S set both upper and lower bounds to z(i) to
and F2(s), for representing the C3 and C4 areas, respec- prevent the cursor from moving out of the screen margins.
tively. The original model was proposed to simulate focal The necessary model parameters were adapted from the
ERD/surround ERS of mu rhythm in hand and foot original models and literature. Only three variables (g1, g2,
representation cortices (Suffczynski et al. 1999). In this and r) were undetermined. As described in the above, g1
study, the model was modified to simulate the interaction of and g2 were the relative weights of the modulating input to
C3/C4 areas in the suppression of mu rhythm while RE and TC cells, and r was the adaptation rate of the visual
performing imaginary movements. In Fig. 1(b), the thala- tracking system. As stated above, the primary purpose of
mocortical (TC) cells received sensory and modulating this study was to analyze the effects of biofeedback training
inputs from ascending sensory signals and the brainstem on these parameters.
through an excitatory synapse, and output of thalamic
reticular (RE) cells through an inhibitory synapse. In the 2.2 Setup of BCI biofeedback experiments
original model, the influences of modulating input (m) on
the TC and RE populations were equal. However, in our A prototype BCI system was built by realizing the BCI
preliminary analyses, we found that the level of ERD was system in the model presented in section 2.1 and included a
related to the relative influences of m on TC and RE clinical EEG machine (model 1A97, San-ei, Tokyo, Japan)
populations. Therefore, we hypothesized that the ratio was and a personal computer for digital signal processing. Two
modified during the biofeedback training, and we added a bipolar EEG signals (y1* from C3-Cz and y2* from C4-Cz)
parameter (g1 and g2, respectively) to represent the relative were amplified (10,000×) via the EEG machine (Fig. 2) and
influence of m on the TC and RE cell populations. The the EEG signals were sampled at a frequency of 256 Hz by
basic configurations of TC and RE cells were identical, a DSP card (DS1104, dSPACE Inc., www.dspaceinc.com)
both having functions of generating excitatory and inhibi- installed in a personal computer. The EEG signals were first
tory postsynaptic potentials (He(s) and Hi(s)) in the input filtered by a band-pass filter (an 8th order Butterworth filter
stage and a sigmoid function for generating output signals. with a pass band in the 5–30 Hz range), then processed by
The intermediate outputs of TCM, y1 and y2, were the the algorithm described in section 2.1 (Eq. (4)) via the DSP
simulated EEG signals at areas C3 and C4 and were card to control the movement of a cursor shown to the
processed by the algorithms of the BCI system and subject in real time as the visual feedback.
transformed into the simulated cursor position. The core The study protocol was approved by the NCKUH ethics
of the BCI system was the following signal processing committee on human subject study and the experiments
algorithm. Short-time Fourier transform with a 1 s rectan- were performed in an isolated room. Eight able-bodied
gular window and 1/8 s shift was performed successively to subjects, ranging from 22–25 years old and with no history
the simulated EEG signals (y1 and y2). The output (z) from of neurological diseases, participated in this study. Each
the BCI system was calculated as subject was asked to sit relaxed in a chair in front of a
" # screen with eyes open and looking at the screen naturally.
Xi
zð i Þ ¼ S w ðPC4 ðnÞ
PC3 ðnÞÞ ; ð4Þ The distance between the subject’s eyes and the screen was
n¼1 80–100 cm. A horizontal track was shown on the screen
where and the cursor moved along the track. Below the track, two
QC3 ðnÞ QC4 ðnÞ thick arrows at the track ends pointing to the right and left
PC3 ðnÞ ¼ QT3 ðnÞ ; PC4 ðnÞ ¼ QT4 ðnÞ ;
and a circle at the center served as candidate targets. The
P
14 P
14
subject was asked to move the cursor by performing
QC3 ðnÞ ¼ f nC3 ðjÞ; QC4 ðnÞ ¼ f nC4 ðjÞ;
j¼8 j¼8 imaginary movements. The subject was initially recom-
P
128 P
128 mended to perform imaginary movements of the upper limb
QT3 ðnÞ ¼ f nC3 ðjÞ; QT4 ðnÞ ¼ f nC4 ðjÞ; parts such as hands, wrists, or thumbs. However, the subject
j¼1 j¼1
determined freely the final tactics and the body part to
and index i represented the number of current window, f nC3 , imagine. In a trial, the cursor was initially located at the
f nC4 represented the power spectrums of the nth windowed center of the track, and was held fixed before a target was
362 J Comput Neurosci (2009) 27:357–368

Fig. 2 Schematic drawing of

the realized experimental BCI
system

given. A target was lit up 5–8 s after the trial began and the Both the ranges for optimization of g1 and g2 were from 0
cursor was released to the control of the subject at the same to 10, and the range for optimization of r was from 0.001 to
time. The variable latency (5–8 s) was utilized to minimize 0.2. The upper and lower bounds of g1, g2 were suggested
the effects of expectation. The subject was instructed to in Suffczynski et al. (1999), and those of r were determined
move the cursor to the target along the track as quickly as by trial and error. The cost function (J) of the optimization
possible. If the presented target was the circle, the subject was the mean-squared error between the simulated cursor
had to keep the cursor at the center of the track. The displacement (z) and the measured cursor position (z*),
displacement to the limits of the track was normalized as ±1 n
2
P
with the negative sign denoting the left end, and −1~−0.7, Zi
Zi
−0.15~0.15, and 0.7~1 were defined as the left, center and J ¼ i¼1 ; ð5Þ
right target zones, respectively. In a trial, the cursor was n
released to the subject’s control in 6 s after an instruction where i indexed the data point and n was the number of
cue was given. The trial was defined as successful if the total data points in a trial, and the parameters were
cursor was kept in the instructed target zone for greater than optimized by finding the minimum of J with steepest
3 s within a period of 19 s. On the contrary, if the chosen descent method. Because r represented the adaptation rate
target was incorrect or no choice was made within the time of visual tracking, and g1 and g2 were the weights of neural
period, the trial was defined as a failure. A trial was connections, these parameters were assumed to change
terminated when a target was chosen, or the time expired. slowly. Therefore, they were assumed to be constants in a
In total, 150 trials evenly distributed across the three trial, and were expected to change only trivially among
instructed targets were performed in random order in a trials in the same session. An optimal set of values for the
session. One practice session including explanation and three parameters were searched by minimizing J via
instruction was run before the main experiment. Three steepest descent algorithm. The average across all the trials
sessions were performed in a week, and 10 sessions in total in a session was used to represent the value of the
were performed for each subject. parameter in that session.

2.3 Estimation of model parameters and simulation 2.4 Validation between simulation and experimental
of experimental observation observation

As mentioned above, three parameters in the model were In order to test the model performance, validation was
undetermined. Their values were searched in the defined performed by estimating the model parameters with 50% of
ranges by optimization technique so that the simulated data randomly selected from each session of every subject,
results of the model best matched the experimental results. and comparing the simulation results with the remaining
J Comput Neurosci (2009) 27:357–368 363

50% of experimental data. Mean-squared-error (MSE) of the mu rhythm in the C4 EEG signal remained relatively
between the experimental and simulated cursor trajectories stable, while the cursor was moved gradually to the right. In
was calculated. The MSE over the remaining 50% of trials the middle column of Fig. 3, the command was to move the
was averaged for each subject and the group mean and cursor to the left. It was not easy visually to determine from
standard deviation of averaged MSE were calculated. the EEG recordings which side of the mu rhythm was
suppressed. However, the difference in power ratios could
2.5 Quantification of results and statistics be observed after bandpass-filtering. The signal processing
algorithm of the BCI system correctly identified the
The success rate of a session was defined as the number of subject’s attempt and moved the cursor to the left. In the
successful trials divided by the total number of trials. The right column of Fig. 3, the subject was instructed to keep
success rate to the respective target (RL, RR and RC) and the the cursor at the center. The cursor deviated slightly to the
overall success rate (RT) were calculated. To verify whether right side in the middle part of the trial, but in general the
the changes in success rates after the 10-session training subject performed well.
were significant, the differences of RL, RR, RC, and RT Figure 4 summarized the success rates (R) of the first
between the first and last sessions were tested by Wilcoxon and last sessions of all subjects. In the first session, RT was
signed-test (Goulden 1959). In addition to the success rate, in the range of 40–70%. In the last session (10th session),
the information transfer rate (ITR) was also utilized to RT of 6 subjects were greater than 80%. The respective
quantify the effects of training (Shannon 1948), success rates of three targets of the first and last sessions
were also compared. It is interesting to note that, for most
1
Ri
ITRi ¼ log2 N þ Ri log2 Ri þ ð1
Ri Þ log2 ð6Þ of the subjects, the success rates of two directions (RR and
N
1
RL) were asymmetric and the improvement was larger for
where N was the number of targets, and subscript i indexed the initially lower-performing direction. Only subject S5
T, L, R, and C. demonstrated greater than 15% improvement for all three
The changes of parameters (r, g1, and g2) estimated from targets; other subjects achieved less than 10% improvement
the experimental data and model simulation before and after for at least one task (i.e., the improvements of RC for S2,
trainings were also analyzed. Because the inter-subject S3, S4, S6, and S8 were less than 10%), S1’s improvement
variations of the magnitudes of these parameters were large, of RL was approximately 9%, and the improvement of RR
the relative changes of each parameter with respect to the for S4 and S7 were less than 10%. The difference between
first session was defined as the RT of first (56.49%±11.25%) and last (81.11%±
13.52%) sessions were statistically significant (p=0.0039),
xi
x1
xi ¼ ð7Þ and the differences of RC, RL and RR between the first and
x1
last sessions were significant or marginally significant (p=
where xi was the value of parameter x of a subject in the ith 0.05, for holding at center and moving leftward, and p=
session, and the superscript * denoted the relative changes 0.06 for moving rightward, respectively) by Wilcoxon
of a parameter. signed test.
In general, the differences between the information
transfer rate (ITR) of first and last sessions were more
3 Results prominent than the counterpart comparisons of success
rates. After training, 5 of 8 subjects (S2, S5, S6, S7, and
3.1 Experimental results S8) had an ITR of more than 0.9 bits/trial. The ITR of
subjects S2, S5, and S8 were greater than 1 bits/trial.
Figure 3 shows an example of typical experimental results The ITRs of subjects S3 and S4 were much lower than
of the realized biofeedback BCI system. The first and the ITRs of the other 6 subjects. The differences of bit
second panels of Fig. 3 depicted the EEG recordings of C3- rates between the first (0.19 ±0.16 bits/trial) and last
Cz and C4-Cz, respectively. The solid and dashed curves (0.76 ±0.33 bits/trial) session were statistically significant
shown in the third panel were the corresponding power (p=0.004) by Wilcoxon signed test.
ratios (Pc3* and Pc4*) of mu rhythms shown in the first and
second panels, respectively. The solid lines in the fouth 3.2 Model simulations and analyses
panel were the actual position of the cursor, and the dashed
lines were the target for the cursor movement. In the left Figure 5(a) shows an example of EEG signals obtained
column of Fig. 3, the power ratio of the mu rhythm of the from model simulation and experimental measurement. The
C3 EEG signal decreased gradually after the instruction subject attempted to move the cursor to the right and there
was given to move the cursor to the right. The power ratio was a corresponding decrease in C3 EEG signal amplitude
364 J Comput Neurosci (2009) 27:357–368

Fig. 3 Sample experimental results. Typical imaginary task results of (8–14 Hz) EEG at C3 and C4, respectively. The solid and dashed lines
moving to the right, moving to the left, and keeping at the center, are in the fourth panel were the actual and target positions of the cursor,
shown in 3 respective columns from left to right. The first and second respectively. Thick vertical bars on the left lower panel mark the three
panels were experimental C3 and C4 EEG recordings, and the third target zones
panel presented the power ratios (Pc3* and Pc4*) of bandpass-filtered

starting at about 1 s. The trend was similar for EEG signals of the training. These results implied that our model
derived from model simulation and experimental measure- captured the experimental results well at the end of the
ments. Figure 5(b) depicts typical results of cursor position training.
derived from model simulation and experiments. It is clear Because the model behavior varied under different
that the simulated cursor position matched the experimental choices of parameters, systematic analyses of the effects
results well for all three targets. of changing the parameters were performed by simulation
The upper three panels of Fig. 6 depict the group (n=8) with combinations of incremental changes in parameters.
means and standard deviations of relative changes of The results (Fig. 7) indicated that the percentage of mu
parameters across the sessions. All three relative changes rhythm suppression and, consequently, the success rate,
of parameters presented increasing trends and the regres- were increased by increasing the magnitude of the
sion coefficients of g1, g2, and r were significantly larger parameters. Combination of parameters g1 and g2 set an
than zero, (p<0.0001, <0.0001, and =0.046, respectively). upper limit that increasing r would saturate. For larger g1
The increase in the standard deviations of g1 and g2 is and g2, the success rate saturated at a lower r value.
roughly proportional to the increase in the means. The
group mean MSE between the subjects’ experimental and
target cursor trajectories was also calculated (fourth panel 4 Discussion
of Fig. 6). Both the mean and the standard deviation of
MSE decreased with the training session. The results In this study, the experimental results demonstrated that the
demonstrated that the success rate of each subject became success rates of using BCI to drive a cursor by imagination
higher with progressive training sessions. The fifth panel of could be improved through biofeedback training, and the
Fig. 6 shows the group mean MSE between subjects’ model simulation results indicated that the improvement
simulated and experimental cursor trajectories. Both the was mediated through both an increase in the adaptation
mean and the standard deviation of MSE decreased with the rate of visual tracking and an increase in the synaptic gain
training session. The group mean MSE was 0.12 at the end from the visual tracking system to the RE cells in the
J Comput Neurosci (2009) 27:357–368 365

Fig. 4 Success rates (upper

panel) and ITR (lower panel) of
all subjects before and after
training

thalamo-cortical circuit. Model analyses with systematic describe the mu rhythm suppression arose from the
parameter combinations supported the hypothesis that interaction of thalamocortical and reticular cells and
increasing the parameters (i.e., adaptation rate and synaptic external inputs. The model was constructed in modules
gains mentioned above) enhanced mu rhythm suppression with anatomical-physiological correspondence. The advan-
and improved the success rate. tage was that each module could be replaced and updated
Existing computational neuronal models ranged from without altering the basic framework of the entire system.
single-neuron models, neural network models, and neuronal In addition to the application of interpreting experimental
population models to global models. The neuronal popula- observations, this model can also be used to test, estimate,
tion model was adopted in this study because it could or predict the performances of other newly developed

Fig. 5 Left: an example of exper-

imental and simulated EEG
signals. The time axis is shifted so
that 0 second represents the step
point of moving to the desired
target. Right: an example of
experimental and simulated cursor
displacements for moving to right,
moving to left, and holding at
center imagination. Solid and
dashed lines represent experimen-
tal and simulated results,
respectively
366 J Comput Neurosci (2009) 27:357–368

To analyze the TCM model (shown in Fig. 1(b))

mathematically, the sigmoid function in the neuron model
(Fig. 1(c)) is simplified with a linear function when the
input is limited within the unsaturated range. By this
simplification, the TCM model with inputs, q and m, and
output, y, can be expressed as

H e ðsÞ He ðsÞ þ a2  gj  Hi ðsÞ

yj ¼ q  þm ; ð8Þ
a1 a1
where subscript j indexes 1 and 2 for C3 and C4 EEG signals,
respectively, and a1 and a2 are lumped constants (which
included the simplified sigmoid function). The parameter gj
controls the balance of excitatory He(s) and inhibitory Hi(s)
effects of the modulating input m and, in turn, m amplifies
the weighted effects to the output yj. If gj is large, the
weighted effect is inhibitory and m amplified this effect to
yj, and vice versa. For a very small m (i.e., in the resting
state) the effect of gj becomes negligible and the behavior
of the whole system is dominated by He(s). He(s) is shown
by power spectral analysis to have a resonant frequency
around 12 Hz. From the above reasoning, it is clear that a
large gj makes the suppression effect more prominent when
the modulating input to the thalamo-cortical circuit is large.
During the training of movement imagination, the
Fig. 6 Time course of relative changes on parameters g1*, g2*, and r*
nervous system adapted to minimize the difference between
(panels 1–3), group mean mean-squared-error (MSE) between desired
targets and cursor displacements (panel 4), and group mean MSE the actual and the desired trajectories shown on the screen.
between model simulation and experiment results (panel 5) The results of model simulations indicated that the
increases in both mu rhythm suppression and success rate
signal processing algorithms by simply replacing the signal were accompanied by an increase in g1 and g2. Historically,
processing component in the BCI subsystem. Two simpli- a large body of experimental results on synaptic plasticity
fications were assumed in the model construction. First, the has been accumulated. Many of these experiments were
time constant in the visual-motor system was assumed to be inspired by Hebb’s assumption, which described how the
identical for all subjects (τ=0.2 sec). The subjects in this connection from a presynaptic neuron to a postsynaptic
study were all healthy and young. If the model is to be neuron was modified. When the axon of a presynaptic cell
applied to stroke or paralyzed patients, proper adjustment of was close enough to excite a postsynaptic cell and sent
τ would be required. Second, all experimental noise, pulses episodically or persistently, some growth process or
including the artifacts of EMG or environmental electro- metabolic change took place in one or both cells such that
magnetic waves, was ignored. the transmission efficiency of the synapse was increased

Fig. 7 Analyses of parameters

(g1, g2, and r) in the BCI
biofeedback model. Simulated
success rates as a function of g2
and r, when g1 was fixed at
8 (left panel), and when g1 was
fixed at 1 (right panel)
J Comput Neurosci (2009) 27:357–368 367

(Brown et al. 1989). The increase in parameters g1 and g2 In the beginning of the training session, the subjects were
can be interpreted as an increase in synaptic plasticity. asked to perform motor imagination in their minds (e.g., to
In this study, we used the ability to remain in the correct stretch one of their hands into the screen and drag the cursor).
zone for more than 3 s as the criteria of success. In the initial Unexpectedly, for these non-trained and able-bodied subjects,
stage of our study, we found that the cursor could drift to perform pure imagination without any real movement was
regardless of active control by the subject. In order to verify the not intuitive and straightforward. Although the subjects were
cursor movement and target selection were under subject’s initially suggested to perform imagination of hand move-
control, the duration that the cursor remained in a zone without ments, this approach may not be suitable for all subjects. If the
active control was investigated. In a pilot study, with the same recommended mental task did not work satisfactorily, they
experimental setup, 4 subjects were asked to be relaxed and not had to find another mental task for themselves. Subject S1 was
to actively control the position of the cursor. Sixty trials were comfortable with repeatedly imagining opening and closing
performed. The duration that the cursor remained continuously one of his hands. Subject S2 adopted imagination of waving
in a zone was 1.32–2.25 s. Based on this result, we set 3 s as the one of his arms, and he said the faster the waving he imagined,
success criteria to define active control. the faster was the cursor moved. Subject S5 preferred to
One unique design of training in this study was that there imagine the feeling of sweeping a rough surface (like a brick
were two imaginary movements (leftward and rightward) wall) with the palm. Subject S7 tended to imagine holding a
but three target zones. We thought that, in addition to joystick and swinging it leftward or rightward, or clicking a
moving a cursor, holding the cursor at a certain position button with left or right thumb quickly and repeatedly.
was an important capability. Obermaier et al. (2003) Interestingly, these four subjects who finally developed their
developed a letter spelling system called Virtual Keyboard own best-fitted mental tasks all had more than 20%
that was operated by performing one of two imaginary improvement in the total success rates (RT) after training.
tasks to choose a desired (or assigned) letter from 32 letters On the other hand, subject S4, who showed minimal
with a bit rate of about 1.02–0.67 bit/min. In order to improvement after training and poor RT, stated that he could
successfully operate letter selection using Virtual Keyboard, not determine a suitable mental task to control the cursor
on the average, five consecutive successful mental imagi- even at the end of training. According to our signal
nations were necessary. Obermaier et al. (2001) investigat- processing algorithm Eq. (5), the cursor movement depended
ed the success rate and bit rate under different numbers of on both the difference in relative powers of mu rhythm in C3
imaginary movements and targets (n = 2 to n = 5) by and C4 positions, and the duration of maintaining the state of
classification with the hidden Markov model. The results difference. Six subjects felt that it was easier to keep the
indicated that the best bit rate (0.8 bits/trial) was achieved cursor near the center than to move it toward the ends,
when the target number was 3, while the highest success because moving the cursor toward the ends required great
rate was obtained when the target number was 2. One of the concentration to maintain imagination for several seconds.
main differences in the design among the experiment of After 4 to 6 sessions, 5 subjects (S1, S2, S5, S6, and S7) had
Obermaier et al. (2001, 2003) and our experiment was that learned to maintain the cursor around the center better than
in the design of Obermaier et al. (2001), the number of they did at the beginning of training.
mental tasks was the same as the target number, while in From the results of this study, we suggest that there are
Obermaier et al. (2003) and our designs, the number of three possible modifications of the biofeedback training
targets was greater than that of the mental tasks. paradigm that may improve training efficiency. First,
According to Eq. (6), increasing selectable targets can because our results demonstrated that most subjects were
increase the bit rate. However, increasing the target less successful in one target, the training program can be
number usually makes a successful selection more difficult designed to distribute the number of training trials according
and decreases the success rate. Therefore, the bit rate is a to the relative success rates of three targets (i.e., more
compromise between the target number and accuracy. training trials would be allocated to the less successful
Julien et al. (2005) discussed the bit rates of several target). Second, we observed that the primary cause for
published BCI systems under Nykopp’s (2001) and more than 50% of the trial failures was speed. The subject
Wolpaw’s definitions (Shannon 1948). The results showed moved the cursor in the correct direction but could not reach
that the bit rates calculated by these two definitions were the target zones within the pre-set time limit. Therefore,
similar, and the highest bit rate (by Wolpaw’s definition) training to increase the moving speed of the cursor, or to
in the summarized results was 3.42 bits/trial (Meinicke et adjust the target zones individually, may increase the
al. 2003), while others were between 0.1 and 0.95 bits/ success rate. Third, for those subjects who fail to achieve
trial. Though our final mean bit rate (0.76 bits/trial) was more balanced success rates in three target zones after
within the same range, our trial time was substantially prolonged training, an asymmetric target zone definition
longer. may also increase the success rate.
368 J Comput Neurosci (2009) 27:357–368

5 Conclusion processes in the generation of alpha rhythms, revealed by partial

coherence analysis. Electroencephalography and Clinical Neuro-
physiology, 50, 449–456. doi:10.1016/0013-4694(80)90011-5.
In this study, we proposed a mathematical model of a man- Meinicke, P., Kaper, M., Hoppe, F., Heumann, M., & Ritter, H.
machine visual-biofeedback BCI system that was capable (2003). Improving transfer rates in brain computer interface: a
of simulating EEG signals and cursor movement in the real case study. Advances in Neural Information Processing Systems,
15, 1107–1114.
experiments. The model successfully manifested the effects
Miller, E. N. (1969). Learning of visceral and glandular responses.
of biofeedback training by tuning three essential model Science, 163, 434–445. doi:10.1126/science.163.3866.434.
parameters. Nykopp, T. (2001). Statistical modeling issues for the adaptive brain
interface. Master thesis, Helsinki University of Technology,
Acknowledgements This research was partially supported by the Department of Electrical and Communication Engineering.
ROC Department of Economics via a contract 96-EC-17-A-19-S1- Obermaier, B., Neuper, C., Guger, C., & Pfurtscheller, G. (2001).
053. We thank Dr. P. Suffczynski for generously offering the code of Information transfer rate in a five-classes brain–computer
his thalamo-cortical model. interface. IEEE Transactions on Neural Systems and Rehabilita-
tion Engineering, 9(3), 50–55. doi:10.1109/7333.948456.
Obermaier, B., Müller, G. R., & Pfurtscheller, G. (2003). “Virtual
References
Keyboard” controlled by spontaneous EEG activity. IEEE
Transactions on Neural Systems and Rehabilitation Engineering,
Birbaumer, N. (1999). Slow cortical potentials: Plasticity, operant 11(4), 422–426. doi:10.1109/TNSRE.2003.816866.
control, and behavioral effects. The Neuroscientist, 5, 74–78. Pfurtscheller, G., & Berghold, A. (1989). Patterns of cortical
doi:10.1177/107385849900500211. activation during planning of voluntary movement. Electroen-
Birbaumer, N., & Cohen, L. G. (2007). Brain-computer interfaces: cephalography and Clinical Neurophysiology, 72, 250–258.
Communication and restoration of movement in paralysis. The doi:10.1016/0013-4694(89)90250-2.
Journal of Physiology, 579, 621–636. doi:10.1113/jphysiol. Pfurtscheller, G., & Lopes da Silva, F. H. (1999). Event-related EEG
2006.125633. EMG synchronization and desynchronization: Basic principles.
Brener, J. (1981). Control of internal activities. British Medical Clinical Neurophysiology, 110, 1842–1857. doi:10.1016/S1388-
Bulletin, 37, 169–174. 2457(99)00141-8.
Brown, T. H., Ganong, A. H., Kairiss, E. W., Keenan, C. L., & Kelso, S. R. Pfurtscheller, G., & Neuper, C. (2001). Motor imagery and direct
(1989). Long-term potentiation in two synaptic systems of the brain-computer communication. Proceedings of the IEEE, 89,
hippocampal brain slice. In J. H. Byrne & W. O. Berry (Eds.), Neural 1123–1134. doi:10.1109/5.939829.
models of plasticity, pp. 266–306. San Diego: Academic Press. Poulton, E. C. (1974). Sine wave tracks. In Tracking skill and manual
Curran, E. A., & Stokes, M. J. (2003). Learning to Control Brain control, pp. 107–139. New York: Academic Press.
Activity: A review of the production and control of EEG Sellers, E. W., & Donchin, E. A. (2006). A P300 based brain computer
components for driving brain-computer interface (BCI) systems. interface: Initial tests by ALS patients. Clinical Neurophysiology,
Brain and Cognition, 51, 326–336. doi:10.1016/S0278-2626(03) 117, 538–548. doi:10.1016/j.clinph.2005.06.027.
00036-8. Shannon, C. E. (1948). A mathematical theory of communication. The
Fabiani, G. E., McFarland, D. J., Wolpaw, J. R., & Pfurtscheller, G. Bell System Technical Journal, 27, 379–423.
(2000). Conversion of EEG activity into cursor movement by a Sterman, M. B. (1981). EEG biofeedback: Physiological behavior
brain-bomputer interface (BCI). IEEE Transactions on Neural modification. Neuroscience and Biobehavioral Reviews, 5, 405–
Systems and Rehabilitation Engineering, 12(3), 331–338. 412. doi:10.1016/0149-7634(81)90036-1.
doi:10.1109/TNSRE.2004.834627. Suffczynski, P., Pijn, J. P. M., Pfurtscheller, G., & Lopes da Silva, F. H.
Fetz, E. E. (1969). Operant conditioning of cortical unit activity. (1999). Event related dynamics of alpha band rhythms: A neuronal
Science, 163(870), 955–958. doi:10.1126/science.163.3870.955. network model of focal ERD/surround ERS. In G. Pfurtscheller & F.
Goulden, C. H. (1959). Methods of Statistical Analysis (2nd ed.). New H. Lopes da Silva (Eds.), Event-Related Desynchronization. (Volume
York: Wiley. 6, Handbook of Electroencephalography and Clinical Neurophysi-
Hinterberger, T., Weiskopf, N., Veit, R., Wilhelm, B., Betta, E., & ology - Revised Series), pp. 67–85. New York: Elsevier.
Birbaumer, N. (2004). An EEG-driven brain-computer interface Suffczynski, P., Kalitzin, S., Pfurtscheller, G., & Lopes da Silva, F. H.
combined with functional magnetic resonance imaging (fMRI). (2001). Computational model of thalamo-cortical networks:
IEEE Transactions on Bio-Medical Engineering, 51, 971–974. Dynamical control of alpha rhythms in relation to focal attention.
doi:10.1109/TBME.2004.827069. International Journal of Psychophysiology, 43, 25–40.
Howe, R. C., & Sterman, M. B. (1972). Cortical-subcortical EEG doi:10.1016/S0167-8760(01)00177-5.
correlates of suppressed motor behavior during sleep and waking Vaughan, T. M., Wolpaw, J. R., & Donchin, E. (1996). EEG-based
in the cat. Electroencephalography and Clinical Neurophysiolo- communication: Prospects and problems. IEEE Transactions on
gy, 32, 681–695. doi:10.1016/0013-4694(72)90104-6. Rehabilitation Engineering, 4, 425–430. doi:10.1109/86.547945.
Julien, K., Svyatoslav, V., & Thierry, P.(2005). Analysis of bit-rate Wilson, H. R., & Cowan, J. D. (1972). Excitatory and inhibitory
definitions for brain-computer interfaces. International conference on interactions in localized populations of model neurons. Biophys-
Human-Computer Interaction (HCI’05), Las Vegas, Nevada, USA. ical Journal, 12, 1–25. doi:10.1016/S0006-3495(72)86068-5.
Kimmel, H. D. (1974). Instrumental conditioning of autonomically Wright, J. J., Bourke, P. D., & Chapman, C. L. (2000). Synchronous
mediated responses in human beings. The American Psycholo- oscillation in the cerebral cortex and object coherence: Simula-
gist, 29(5), 325–335. doi:10.1037/h0037621. tion of basic electrophysiological findings. Biological Cybernet-
Lopes da Silva, F. H., Hoeks, A., Smits, A., & Zetterberg, L. H. ics, 83, 341–353. doi:10.1007/s004220000155.
(1974). Model of brain rhythmic activity: The alpha-rhythm of Wyrwicka, W., & Sterman, M. B. (1968). Instrumental conditioning
the thalamus. Kybernetik, 15, 23–37. doi:10.1007/BF00270757. of sensorimotor cortex EEG spindles in the waking cat.
Lopes da Silva, F. H., Vos, J. E., Mooibroek, J., & Van Rotterdam, A. Physiology & Behavior, 3, 703–707. doi:10.1016/0031-9384
(1980). Relative contributions of intracortical and thalamo-cortical (68)90139-X.