Computational
Sensorimotor
Systems Lab
Application of Independent Component Analysis (ICA) on
Magnetoencephalography (MEG)
Juanjuan Xiang (ECE, ISR), Jonathan Z. Simon (ECE, Biology, ISR)
Introduction
Magnetoencephalography (MEG) is a noninvasive technique by which the brain activity can be measured with
very good temporal resolution (1ms) and moderate
spatial resolution (1cm).
component is heart beat artifact. Figure 3 presents the
29th component, it is related to the auditory response.
ICA model
5
We use ICA to unmix the separate sources’ activity.
4
3
2
1
MEG signals are recorded in a magnetically shielded
room with a 160 channel whole-scalp Neuromag-160
neuromagnetometer, which is sensitive to small magnetic
field as 10fT.
0
0
2
0
4
0
6
0
I n
d
8
e
x
o
f
C
0
1
o
m
p
o
n
0
e
0
n
1
2
0
1
4
0
1
6
0
t
Figure 1 Spatial Contribution
6
5
5
4
5
4
3
2
4
3
1
0
-1
-2
3
-3
0
0.5
1
1 .5
2
2
2.5
T im e ( s )
2
• Assumption
Sources are independent
• Method
Estimate Weights so to maximize output entropy H(Y)
Ù Minimize mutual information I(Y)
• Goal
Learn the transform which unmix the recorded signal
so that the output is as independent as possible
• Example
1
1
0
0
-1
-1
-2
-2
-3
2
-3
0
5
10
T im e (s )
15
-4
20
0
10
10
5
5
5
10
T im e (s )
15
20
KIT-UMD whole head MEG system &
shielded room. 160 axial gradiometers,
To analyze MEG recordings, we need to extract the
essential features of the neuro-magnetic signals in the
presence of artifacts. A new method to separate brain
activity from artifacts is Independent Component
Analysis (ICA), which is based on the assumption that
the brain activity and the artifacts, e.g. heart beat or eye
movements are anatomically and physiologically
separate processes, and this separation is reflected in
the statistical independence between the magnetic
s i gna ls generated by those processes[1][2][3].
Experiment
Acoustic stimuli are sinusoidal amplitude-modulated
tones at 16, 32 and 48 Hz, with carrier frequency at 400
Hz. Each 20 second stimuli is presented 10 times in
pseudorandom order. 157 sensors are placed on that
subject’s scalp to record the magnetic field. 3 additional
magnetometer are placed away from the brain to record
the reference noise.
0
Power Spectrum Magnitude (dB)
Power Spectrum Magnitude (dB)
1
-5
-1 0
-1 5
-2 0
-2 5
0
10
20
30
40
F re q u e n c y
50
60
Figure 2 Heart Beat
Joint distribution of mixed
signals x1, x2
Joint distribution of unmixed
sources s1,s2
Results
-1 0
-1 5
-2 0
-2 5
-3 0
-3 5
0
-5
70
-3 0
0
10
20
30
40
F re q u e n c y
50
60
70
Figure 3 Component 29
Conclusion
ICA is a powerful way to enhance the MEG signal. In
addition to reducing artifacts, ICA can be used to
decompose auditory related responses, which enables
direct access to the underlying brain function.
After applying ICA on the MEG recording, we display the
unmixed compon ents according to their spatial Reference
distribution on the scalp. Figure 1 is the ordered spatial [1] Aapo Hyvärinen and Erkki Oja, Independent Component Analysis: Algorithms and
contribution for every component. Figure 2 presents the Applications, Neural Networks, 13(4-5): 411-430, 2000
first component’s waveform, power spectral density and [2] What is MEG, http://www.geocities.com/Tokyo/1158/meg.html
Anthony J. Bell and Terrence J. Seinowski, An Information-Maximization Approach to Blind
spatial contribution on the sensors. It shows this [3]
Separation and Blind Deconvolution, Volume 7, Issue 6 (November 1995) Pages: 1129–1159