Beruflich Dokumente
Kultur Dokumente
, Peter D. Lawrence, Senior Member, IEEE, and Gary E. Birch, Senior Member, IEEE
AbstractA novel approach is presented for using an eye
tracker-based reference instead of EOG for methods that require
an EOG reference to remove ocular artifacts (OA) from EEG. It
uses a high-speed eye tracker and a new online algorithm for ex-
tracting the time course of a blink from eye tracker images to
remove both eye movement and blink artifacts. It eliminates the
need for EOG electrodes attached to the face, which is critical
for practical daily applications. The ability of two adaptive lters
(RLS and H
algorithms
as implemented in [20]. Adaptive lters such as RLS and H
(4)
[ R(0) ]
1
= 100 I (5)
H(n) = H(n 1) +K(n)(s(n) r(n)
T
H(n 1)) (6)
H(0) = 0 (7)
where I is the identity matrix, and is a forgetting factor.
As in [7], = 0.9999, noting that for low values the algorithm
would not converge, and for 0.995 < < 1 the actual value did
not make a noticeable difference to the nal performance.
For H
(n) = r(n)r(n)
T
(9)
P(n) = [ P
1
(n)
2
g
r
(n) ]
1
(10)
P(n) = [ P
1
(n 1) + (1
2
g
)r
(n 1) ]
1
+ 10
5
I
(11)
P(0) = 0.005 I (12)
H(n) = H(n 1) +K(n 1)(s(n 1)
r(n 1)
T
H(n 1)) (13)
H(0) = 0. (14)
In this way, the proposed method of replacing EOG with an eye
tracker-based reference signal was evaluated using two existing
OA removal methods that require an EOG reference.
NOUREDDIN et al.: ONLINE REMOVAL OF EYE MOVEMENT AND BLINK EEG ARTIFACTS 2105
TABLE I
DATA COLLECTION TASKS
A. Data Collection
The proposed method was evaluated using real recorded
biosignals. Simultaneous EEG, EOG, and eye tracking data were
collected from four subjects (one female, three males) aged 19
63, sitting approximately 50 cm in front of a computer monitor.
The subject protocol was approved by the University of British
Columbia Behavioural Research Ethics Board (certicate no.
B03-0500). Each subject was instructed to perform four tasks
(see Table I), providing data during nine OA types: small left
(SL), small right (SR), small downward (SD), large left (LL),
large right (LR), large upward (LU), and large downward (SD)
saccade, and small (SB), and large blink (LB).
For each subject, 55 EEG channels, 2 linked mastoids, and
6 channels of EOG (see later) were collected during a single
3-h session. All signals were sampled by the same amplier at
200 Hz, with a 70-Hz low-pass lter and a 0.530 Hz digital
band-pass lter (BPF). Data were processed using a 2.67-GHz
Intel Core2 PC with 2-GB RAM. Horizontal (H
EOG
) and ver-
tical (V
EOG
) EOG were calculated as follows:
H
EOG
= H
R
H
L
(15)
V
EOG
=
(V
BR
V
TR
) + (V
BL
V
TL
)
2
(16)
where H
R
and H
L
were measured at the left and right outer
canthi, respectively and V
TR
, V
BR
, V
TL,
and V
BL
were measured
above (denoted by T) and below (denoted by B) the right
eye and above and below the left eye, respectively.
In addition, an eye tracker [12] was used to collect eye move-
ment and blink data at approximately 400 Hz simultaneously
with EEG and EOG. Before the start of each task, it was en-
sured that the subjects eye was in the eye tracker cameras eld
of view. For each frame, the eye tracker provided the following.
1) A grayscale image of the eye of the subject.
2) Atimestamp when the image was captured. This was com-
bined with hardware logic signals provided by the eye
tracker to accurately synchronize the eye images with the
EEG samples.
3) A ag indicating whether a pupil was found.
4) The parameters of the ellipse that best t the pupil.
5) A ag indicating whether the subject was performing a
xation at the time.
The previously mentioned was used to extract three signals
(see Fig. 2). The pupil centers x- and y-coordinates were ltered
using a 0.530 Hz BPF to produce H
ET
and V
ET
, corresponding
to horizontal and vertical eye movements, respectively. The third
signal B
ET
corresponds to predicted changes in EEG during
blinks. The rst time the pupil was found and was stable, base
Fig. 2. Overall operation of eye tracker for every image captured. The signals
H
ET
, V
ET,
and B
ET
are extracted as a reference for OA removal.
Fig. 3. Areas of eye tracker images used for extracting blink signal (not drawn
to scale). R is the pupil radius (average of lengths of ellipse axes).
intensities X
1
, X
2
, and X
3
were calculated
X
k
=
I
k
0
, k = 1, 2, 3 (17)
where I
1
, I
2
, and I
3
are image intensities corresponding to S
1
,
S
2
, and S
3
in the eye image (see Fig. 3), and
0
is the average
intensity of the entire eye image. For each subsequent image,
intensities Y
1
, Y
2
, and Y
3
were calculated
Y
k
=
I
k
, k = 1, 2, 3 (18)
where I
1
, I
2
, and I
3
are as aforementioned, and is the av-
erage intensity of the entire eye image. If the image was from
a xation, the current pupil position was used as P
0
in Fig. 3.
Otherwise, the pupil position from the last xation was used,
since during a blink the pupil position is likely inaccurate or not
2106 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 59, NO. 8, AUGUST 2012
Fig. 4. Eye tracking versus EOG during (a) simultaneous horizontal (left) and vertical (up) saccade, and (b) blink. Sample eye images are shown above and EOG
and eye tracking signals below, with arrows pointing to the time instant each image was captured. In (a), the x shows the current pupil position, and the dot shows
the pupil position at the start of the saccade. In (b), the ellipse shows the area where the intensity value is tracked during the blink.
available. For each image, Z was computed from X and Y
Z = 2
N
1
i=1
(Y
1
X
1
)
2
+
N
2
i=1
(Y
2
X
2
)
2
+
N
3
i=1
(Y
3
X
3
)
2
(19)
where N
1
, N
2
, and N
3
are the number of pixels in S
1
, S
2
, and
S
3
, respectively. The resulting Z was ltered using a 0.530 Hz
BPF to produce the blink signal B
ET
. Finally, the eye trackers
timestamps were used to synchronize H
ET
, V
ET
, and B
ET
with
EEG. Fig. 4 shows examples of EOG and eye tracker signals
corresponding to eye movements and blinks.
B. Reference Inputs
From the data collected, three different reference inputs [X
in (1)] were constructed for use by each of the algorithms (AF7,
AF8, and Fpz are signals from those frontal EEG channels, and
H
EOG
, V
EOG
, H
ET
, V
ET
, and B
ET
are as aforementioned)
1) X = [H
EOG
, V
EOG
]
T
(EOG reference).
2) X = [AF7, AF8, Fpz]
T
(frontal EEG or fEEG refer-
ence, as in [21]).
3) X = [H
ET
, V
ET
, B
ET
, AF7, AF8, Fpz]
T
(eye-tracker-
based reference or ET + fEEG). See later for the reason
for combining the eye tracker signal with EEG signals.
C. OA Removal Evaluation
EEG, EOG, and eye tracker measurements contain common
information about eye and eyelid movement, as well as infor-
mation not in common [22], [23], as shown in Table II. While an
eye tracker signal only measures eye movements, EOGcontains
other biosignals. Therefore, it would be misleading to compare,
TABLE II
CONTRIBUTION BY SPECIFIC SOURCES TO VARIOUS MEASURED SIGNALS
for example, EOG with an eye tracker signal: using EOG might
remove more of the signal, but it may also result in removing
parts of the EEG that are unrelated to the OA (Background
EEG or ERP).
To address this problem, the comparison was carried out on
the average of several trials of each OA type (see aforemen-
tioned). The average signal was calculated and analyzed sep-
arately for each subject. Thus, any components that were un-
related to the OA were minimized. Since the average signal is
mostly OA related, removing more signal could consistently
be considered a good measure of the performance of the OA
removal algorithm, and by extension the choice of reference
input.
The eye tracker does not contain information unrelated to OA,
thus minimizing EEG distortion after OA removal. However,
NOUREDDIN et al.: ONLINE REMOVAL OF EYE MOVEMENT AND BLINK EEG ARTIFACTS 2107
since there are OA-related components (e.g., OA-related EMG
or EEG) that are picked up by EOG and EEG electrodes but not
by the eye tracker, the eye tracker signals were combined with
frontal EEG. Including EEG channels helps remove sources
time-locked with OA, while including eye tracking signals min-
imizes the removal of nonOA related components. Initial at-
tempts to use just eye tracker signals H
ET
, V
ET,
and B
ET
re-
sulted in poor OA removal by both algorithms, thus conrming
the previous discussion.
OA removal methods are normally evaluated using simulated
data. Artifact-free EEG and an artifact signal are combined arti-
cially and processed using the algorithm (see Fig. 1). Metrics
such as the SNRof the output Y can be compared to the artifact-
free EEG V
B
. For real EEG, however, V
B
, V
N
, and V
A
are
unknown, so the performance on real data is reported subjec-
tively [3], [24], [25], often based on visual inspection of the
resulting waveforms. That is, it is not possible to measure the
SNR. As shown previously [26], the R metric [19] can be used
instead to measure the amount of OA removed
R =
N
k=1
(E(k) Y(k))
2
N
k=1
Y(k)
2
(20)
where N is the number of samples. For each subject, each algo-
rithm was applied to every trial of each type of OA at each
EEG channel using each of the three reference inputs (see
Section II-B). The ensemble averages of the original and OA-
removed trials were used for X and Y , respectively, in (20)
to calculate R in each case. While the actual value of R does
not hold meaning, its relative value can be used to compare re-
sults using different algorithms and reference signals at different
electrodes, as shown later.
III. RESULTS
Statistical tests were performed for each combination of:
1) 52 electrodes; 2) 2 algorithms (RLS, H
); and 3) 9 OA
types to compare R using the ET + fEEG reference versus the
EOG reference in each case. That is, for each set of variables
(electrode location, OA removal algorithm and OA type), the
signcance of the difference in R between the reference signals
used by the given OA removal algorithm was determined using
a pairwise t-test on the subject data. The difference between
using ET + fEEG versus fEEG was similarly tested. Out of the
resulting set of pairwise t-tests, the combinations that showed
a signicant difference between ET + fEEG and EOG and/or
fEEG are shown in Table III (all other cases were found not
to have signicant differences (p < 0.05) between using ET
+ fEEG and either EOG or fEEG). In each row, the last two
columns show the signicance of the comparison between R
values for ET + fEEG versus (a) EOG; and (b) fEEG.
Fig. 5 shows an example of the previously mentioned graph-
ically for each algorithm. Bar graphs showing mean R at each
electrode are shown, with a box indicating that the difference is
signicant. Overall, more OA is removed at frontal electrodes,
and H
at Fp2 dur-
ing SU saccades, using ET + fEEG is signicantly (p < 0.05)
better (R = 648 in the rst line of the bottom of Table III) than
EOG (R = 46.8), as shown in Fig. 6(b).
IV. DISCUSSION
For H
algorithms during small upward saccades. Electrodes where the difference between means is
signicant are highlighted with boxes. Amplitudes are scaled the same in both (a) and (b) to show the relative performance of the algorithms at each electrode.
Fig. 6. OA removal using (a) RLS at Fp1 during small right saccade and
(b) H
at Fp2 during small upward saccade. In each case, three sample trials and
the average of several trials are shown. The original EEG, and the corresponding
artifact-removed signal using each of the EOG and eye tracker combined with
frontal EEG (ET + fEEG) references are plotted.
removal than EOG. In all cases considered, using the eye tracker
resulted, on average, in as much OA being removed as using
frontal EEG.
Thus, for both algorithms, using the eye tracker was found
overall to be as good or better than using EOG or frontal
EEG at removing OA over the range of OA types and elec-
trodes investigated. Compared to EOG, the eye tracker re-
moved signicantly (p <0.05) more OA (almost 14 as much)
at Fp2 during SU saccades and less OA at P2, PO3, PO4,
PO8, Pz, and POz during SR saccades (EOG removed <3
as much).
Looking at RLS only, it may be argued that since the eye
tracker does not signicantly improve performance over frontal
EEG, there is no need for an eye tracker. However, H
has
been previously shown [26] to perform better than RLS for
OA removal. This nding was consistent with the results in
this study: H
.
Therefore, it can be concluded that the BSG algorithm gener-
ated as effective a reference as EOG for both algorithms during
blinks.
NOUREDDIN et al.: ONLINE REMOVAL OF EYE MOVEMENT AND BLINK EEG ARTIFACTS 2109
V. CONCLUSION
A novel approach was presented for using an eye tracker
to replace EOG for any OA removal method that requires an
EOG reference. The performance on real EEG of the RLS and
H