Part V P300-Based Brain Computer Interfaces
Part V P300-Based Brain Computer Interfaces
Part V P300-Based Brain Computer Interfaces
Outline of Presentation
BCI Companies
References
BCI methods
BCI can be divided into 2 subsections :
1) Dependent BCI: Doesnt use brains normal output pathways
other measuring methodologies are not practical due to their size and nonportability, MEG, fMRI etc.
There are two basic methods for brain activity measurement.
i) Invasive methods i.e, require a surgical operation such as
Invasive Methods
Electrocorticogram and Cortical Microelectrodes
Cortical Microelectrodes
Similar to ECoG but placed inside the cortex.
Electrodes developed with VLSI technology.
The signal quality is improved by integrated analog circuits
design.
Possible to detect the activity of a single neuron with high spatial
resolution and excellent signal-to-noise ratio (SNR)
Non-invasive methods
1) Electroencephalography (EEG)
2) Magnetoencephalography (MEG)
3) Functional Magnetic Resonance Imaging (fMRI)
1)Electroencephalography(EEG)
Electroencephalography (EEG) is the recording of electrical
EEG used to be a first-line method for the diagnosis of tumors, stroke and
other focal brain disorders, but this use has decreased with the advent of
anatomical imaging techniques with high (<1 mm) spatial resolution such
as MRI and CT. Despite limited spatial resolution, EEG continues to be a
valuable tool for research and diagnosis, especially when millisecond-range
temporal resolution (not possible with CT or MRI) is required.
preparing the area by light abrasion i.e., corrosion and remove dead
skins to reduce impedance. Cap is used for when high density array of
electrodes are needed. [26]
2) Each electrode is connected to one input of a differential amplifier, a
common system reference electrode is connected to the other input of
each differential amplifier. These amplifiers amplify the voltage between
the active electrode and the reference (typically 1,000100,000 times,
or 60100 dB of voltage gain) because a typical adult human EEG signal
is about 10V to 100 V in amplitude when measured from the scalp
and is about 1020 mV when measured from subdural electrodes. [26]
Fig. 12 The 10-20 international system is the standard naming and positioning
scheme for EEG applications [27]
Channel Selection
10 channels
Fz, Cz, C3, C4, Pz, P3, P4, PO7, PO8, Oz
Meinicke Kaper
Guger tech.
1)
2)
3)
4)
5)
2) Magnetoencephalography(MEG)
Magnetoencephalography (MEG) is a technique for
3) Functional Magnetic
Resonance Imaging (fMRI)
fMRI is a method to measure the amount of oxygen in
4) Evoked Potentials
EEG.
These potential shifts occur over 0.510.0 s and are called slow
cortical potentials (SCPs). Negative SCPs are typically associated
with movement and other functions involving cortical activation,
while positive SCPs are usually associated with reduced cortical
activation .
They can be generalized as anticipation tasks [1]
2) SSVEPs
1) http://www.bsp.brain.riken.jp/~hova/ssvep_data_Bakardjian_LABSP.html
2) Show videos
engagement)
The amplitude of the signals may change during different brain
activities such as concentrating, voluntary muscle movement.
Movement or preparation for movement is typically
accompanied by a decrease in mu and beta rhythms and
increase in alpha rhythm called as event-related desynchronization or ERD Its opposite, rhythm increase, or
event-related synchronization (ERS) occurs after movement
and with relaxation [1]
ERD and ERS do not require actual
movement, they occur also with motor imagery. [1]
SMR&ERD&ERS
Wave Characteristics
P300 Speller
The idea in this paradigm is to detect the P300 responses elicited
Spelling Paradigm
2 target 12
nontarget visual
stimulations
Counting
the
target
intensifications
will elicit the so
called
P300
responses
P300 Speller
Target (P300) responses
Positive signal pattern peaking nearly 300 ms after the presentation of the target
stimulation
Have latency of 300 - 400ms
50
200
600
700
0 0
-50
-50
100
200
100
300
400
500
600
200 time
300(ms)400 500
time
Channel
Cz (ms)
Channel Pz
Fig. 18 Target vs non-target amplitude
50 in P300 [5]
alue)
100
50 50
50
700
600
700
-50
Channel
Channel
OZ FZ
alue)
50
Channel FZ
C value)
C value)
Reality (VR)
Controlling robotic or prosthetic
devices.
P300 Speller
Problems :
The noise in EEG recordings
Factors in the cognitive process (fatigue, being unable to focus)
Repeating the intensification procedure for the focused
character
Reducing the effect of noise by ensemble averaging of the
observations.
Main Problem:
Decreasing prediction time. [5]
1) Signal Acquisition
Signal Enhancement(2/2)
Commonly used signal enhancement techniques are
1) Surface Laplacian (SL)
2) Common Average Referencing (CAR)
1) Surface Laplacian
The common average reference spatial filter calculates the mean of all channels,
and subtracts this value from the output channel of interest. [16]
If electrodes are equally spaced result is zero mean spatial distribution. [6]
3) Small Laplacian?
4) Large Laplacian?
Results (1/2)
Fig. 23 Average voltage spectra for top targets (solid lines) and bottom targets (dashed lines) and
average spectra of r^2 for the top/bottom difference for all sessions of all subjects for the
locations that controlled cursor movement online.[6]
Results (2/2)
1) Feature Extraction
Fig 27. Feature Extraction methods in BCI designs based on sensorimotor activity,
VEP, P300, SCP, response to mental tasks, activity of neural cells and multiple
neuromechanisms. [2]
Feature Extraction
Commonly used methodologies for BCI are:
1. Time and/or frequency methods.
Time methods have great temporal resolution whereas frequency methods
are preferred due to simplicity in use and fast computation.
2.
temporal resolution
Spectral resolution
providing
time-varying
spectral
representation of a signal which corresponds to the
power spectrum w.r.t time.
There are 2 methods:
Short Time Fourier Transform(STFT)
Morlet Wavelet Transform
STFT- briefly
STFT is fundamental for analyzing the slowly time
varying signal.
In contrast to FT, it can give information on the time
resolution of the spectrum by analyzing the frequency
response at different time instant.
Most popular one is Fast Fourier Transform(FFT)
based on STFT.
Feature Selection
Algorithms are used to find the most informative
which are :
1) Principal Component Analysis (PCA)
2) Genetic Algorithms (GA)
3) Learning Vector Quantization (LVQ)
4) Common Spatial Pattern (CSP)
Principal Component
Analysis(PCA)
PCA is linear transformation that reduces
Genetic Algorithms
Heuristic (depends on exploring) search techniques.
[11]
[11]
[11]
As can be seen, the results using only the selected features are far better than those
using all features. This shows how important feature selection is in the context of
EEG classification where a lot of channels only partially contain information about
the studied phenomenon.[11]
Classification
Translating brain signals into device commands is
high dimensionality
time information
non-stationary
Classifier Taxonomy
In order to choose the most appropriate classifier, the
feature vector.
Unfortunately this cannot be applied in all BCI systems
due to training data set size.
The Bias-Variance trade-off
Classification error can be described under 3 major
possible sources
2) Neural Networks
3) Non-linear Bayesian classifiers
4) Nearest Neighbor Classifiers
5) Combination of Classifiers
1) Linear Classifiers
Discriminant algorithms to distinguish the classes.
Fig 32: A hyperplane which separates two classes: the circles and
the crosses [12]
Cons
1) Provides poor results on complex non-linear EEG
data. [12]
SVM
H3 (green) doesn't
separate the two classes.
H1 (blue) does, with a
small margin and H2
(red) with the
maximum margin. [13]
in BCI research.
NN is an assembly of artificial neurons.
NNs can be clustered under two categories:
1) Multilayer Perceptron (MLP)
2) Other Neural Network architectures
boundaries.
Their generative characteristics enables them to
perform more efficient rejection of uncertain samples
than discriminative classifiers.
They are not widespread as linear classifiers or Neural
Networks in BCI applications because they are not fast
enough for real time BCI applications. [12].
5) Combinations of Classifiers
Recent trend is to combine different classifiers.
Strategies are:
2) Dynamic classifiers
3) Combination of classifiers
Properties of Classifications
SVM
Classification Translation
Algorithm
1)
teknoloji/4/15/odtululer-beyin-sinyallerini-hareketecevirdi
http://www.youtube.com/watch?v=gnWSah4RD2E
http://www.youtube.com/watch?v=ppILwXwsMng&f
eature=fvwrel
2) Emotiv Epoc
Gtec Projects:
1) ALIAS: Adaptable Ambient Living Assistant -
Emotiv - Epoc
Show video!
Thank you!!
References
[1] Wolpaw J.R., Birbaumer N., McFarland D.J., Pfurtscheller G., Vaughan T.M., Brain
Computer Interfaces for Communication and Control, Clinical Neurophysiology, 113:
767-791, March 2002
[3] http://en.wikipedia.org/wiki/P300_(neuroscience)#P3a_and_P3b
[4] Akinci,B. Realization of a cue based motor imagery brain computer interface with its
potential application to a wheel chair. ,METU Library, 2010
[5] Erdoan H. B., A Design and Implementation of P300 Based Brain- Computer Interface,
Metu Library, 2009.
[6] McFarland D.J., McCane L.M., David S.V., Wolpaw J.R., Spatial filter selection for EEG-based
communication, Electroencephalogr. Clin. Neurophysiol,Vol. 103, pp. 386-394.
References
[7] Chapin J.K., Nicolelis M.A. L., "Principle component analysis of neuronal ensemble activity
reveals multidimensional somatosensory representations", J. Neurosci. Meth, Vol. 94, pp. 121-140,
1999.
[8] Bayliss J.D., Ballard D.H., Recognizing Evoked Potentials in a Virtual Environment, NIPS, pp.
3-9, 1999. 137
[9] Ramoser H., Muller-Gerking J., Pfurtscheller G., Optimal spatial filtering of single trial EEG
during imagined hand movement, Rehabilitation Engineering, IEEE Transactions on Neural
Systems and Rehabilitation, Vol. 8, No. 4, pp. 441-446, Dec. 2000.
[10] http://en.wikipedia.org/wiki/Short-time_Fourier_transform
[11] Pregenzer M., Pfurtscheller G., "Frequency component selection for an EEG-based brain
computer interface (BCI)", IEEE Trans. Rehab Eng. Vol. 7, No. 3, Sep. 1999.
[12] Cover T. M., Hart P. E., Nearest neighbor pattern classification, IEEE Transactions Information
Theory, Vol. No. 13, pp. 21-27, 1967.
[13] http://www.engadget.com/2009/05/04/mind-controlled-wheelchair-prototype-is-trulyinsanely-awesome/
[14] https://wiki.engr.illinois.edu/display/BIOE414/ECoG
References
[15] http://ajatubar.feld.cvut.cz/bisig/research
[16]http://www.bci2000.org/wiki/index.php/User_Reference:SpatialFilter#CA
R
[17] http://www.idiap.ch/~marcel/labs/faceverif/face-verif-fordummies/facefeature-to-dct.png
[18] http://phucopierservice.com/nashuatec_micro/image-features-extractionmatlab-i0.jpg
[19]http://micro.magnet.fsu.edu/primer/java/digitalimaging/processing/spatialr
esolution/
[20] http://en.wikipedia.org/wiki/Artificial_neural_network
[21] http://www.aafnh.org/wpcontent/uploads/2012/07/17089nlm_nih_gov.jpg
[22]http://en.wikipedia.org/wiki/File:MS_Demyelinisation_CD68_10xv2.jpg
[23]http://upload.wikimedia.org/wikipedia/commons/thumb/7/75/MuscularD
ystrophy.png/230px-MuscularDystrophy.png
[
References
[24]http://www.humanillnesses.com/original/images/hdc_0001_0002_0_i
mg0181.jpg
[25] http://en.wikipedia.org/wiki/Amyotrophic_lateral_sclerosis
[26] http://en.wikipedia.org/wiki/Electroencephalography
[27]https://wiki.engr.illinois.edu/download/attachments/44733162/ecog.jp
g?version=1&modificationDate=1292382236000
[28] http://www.gtec.at/Research/Projects/ALIAS
[29]http://www.bci2000.org/wiki/index.php/User_Tutorial:EEG_Measure
ment_Setup
[30]http://www.theredmenmovie.com/2009/11/magnetoencephalographymeg-scanner.html
[31]http://blogs.oem.indiana.edu/scholarships/index.php/2009/10/26/neu
rons-and-electrodes/fmri_groot/
[32] Hoffman, U. Bayesian Machine Learning Applied in a Brain-Computer
Interface for Disabled Users, EPFL, 2007