06-10-2012, 11:33 AM
EEG Signal Acquisition, De-Noising and Classification for Brain
Computer Interfaces
[attachment=35487]
Abstract
The focus of our B. Tech. Project is three-fold. The first module deals with the hardware implementation of a
single channel bipolar EEG acquisition system using MSI components followed by digitization of the signal using
a PCL-818HG data acquisition card. The second module addresses de-noising the EEG signal from biological/
electrode and/or power-line artifacts using Independent Component Analysis (ICA). The third module explores the
EEG signal as tool in biometric applications. It is motivated by a surmise that the EEG signal contains genetic
information and is thus unique across individuals as well as repeatable within. Independent progress has been made
on all fronts. Depending on our progress, we also propose to implement a simple real-time EEG-BCI application
which can serve as a prototype for a thought based authentication system.
Introduction and Literature Review
The Electroencephalogram (EEG) signal was first observed and recorded by Hans Berger in 1929. Since then, it
has been been applied extensively in medicine for detecting epileptic seizures and in sleep studies. Since the 1960s,
signal processing techniques have proliferated. Application of these techniques to the EEG signal has opened up new
research directions in the engineering domain. A significant advancement in this field has aimed to provide the brain
with external channels for communication and control. Recent studies suggest that electroencephalographic (EEG)
activity can provide the basis for such new channels [1, 2, 3, 4] and these have come to be known as Electroencephalographic
Brain Computer Interfaces (EEG-BCIs). The majority of BCI applications focus on enabling patients
suffering from neuromuscular disorders such as multiple sclerosis, cerebral palsy, spinal cord injury and severe paralysis
to communicate with the external world in the absence of methods to repair the nervous system. However, there is
emerging BCI research [5, 6, 7, 8, 9] that has explored signal processing techniques in order to quantify EEG patterns
as unique descriptors of human identity. The most common biometric method of identifying persons is through
fingerprint recognition [10]. In recent years, alternative biometric methods to replace or augment the fingerprint technology
have been researched. In this regard, biometrics like palmprint [11], hand geometry [14], iris [13], face [15],
and electrocardiogram [12] have been proposed. In this respect, using the EEG as a biometric is relatively new and is
motivated from the hypothesis that EEG signals contain genetic information as proposed by Vogel in 1970 [16].
Our work seeks to establish a one-to-one correspondence between identity of a healthy individual and certain
features of his/her EEG signal towards the further goal of developing a test for person identification based on features
extracted from the EEG. Potential applications include, among others, information encoding and decoding and access
to secure information [5].
Methods
Acquisition, Sampling and Quantization
The EEG signal amplitude is only a few (50− 100) microvolts and thus has to be passed through a very high gain
amplification stage (gain of the order of 10,000), before processing. Owing to its low amplitude, it is highly susceptible
to contamination from various sources including biological, electrode and power line artifacts. We have used a BJT
based Instrumentation amplifier (AD 521) which amplifies differential voltage between two electrodes while rejecting
the common voltage ( -110dB). Typical electrode materials, such as gold or steel, are polarizable which leads to
building up of large DC-offset owing to huge surface charge accumulation. A single stage amplifier with a gain of
10,000 will amplify a typical DC voltage of 200mV to a very large value, thus saturating the amplifier. Power line noise
can also be a significant contaminant if shielded electrodes are not used. In order to forestall saturation effects, the
amplifier is implemented in 2 stages. In between these stages, mains hum and DC frequencies are attenuated. First, a
high tunable Q notch filter at 50Hz is employed. Second, a high-pass filter of very sharp roll off (-80dB/decade) and
low pass-band frequency (0.1Hz) is used to reject DC. A Sallen-Key active resonant filter with a gain of 1.58 offers
a maximally flat response with a roll off of -40dB/decade [25]. Two such filters are cascaded. Finally, a Sallen Key
low pass filter with a cut-off frequency of 30Hz is used to prevent aliasing effects. A protection circuit before IA stage
helps in isolating the circuitry from electrostatic discharge (ESD), RF and protects the user from failing circuitry.
Schematic of our two stage Instrumentation based Bio-potential amplifier is shown below.
Mental Arithmetic, A Non-cognitive task
This dataset was obtained from an open source distribution available at the University of Colorado [30]. Subjects
were asked to perform mental arithmetic. 10 trials of 10 seconds each, for 4 subjects were conducted. EEG signals
were recorded from C3, C4, O1, O2, P3 and P4. The first 5 trials and the next 5 trials were conducted on separate
days. This data was divided into training and test sets in the ratio 1:1.
Feature Extraction and Classification
For both experiments, half overlapping windows of duration 1 second were used. From each window a training/test
pattern is formed. Features used are either FFT peaks or Linear Prediction co-efficients. Number of features used
were empirically determined and were found to be 70 (per channel).
For the stereopsis experiment, we have 60 seconds of data from 4 channels giving rise to 120 patterns per subject,
each of dimensionality 70 × 4 = 280. Out of these, 80 are used for training and 40 for testing. Similarly, for the
mental-arithmetic experiment, we have 200 patterns of dimensionality 70 × 6 = 420 per subject. 100 are used for
training and 100 for testing.
Two types of classifiers were used. The first technique is a one-to-one Support Vector Machine (SVM). Vapnik
introduced Support Vector Machines as powerful learning tools based on statistical learning theory. An SVM is a
binary classifier that makes its decision by constructing a linear decision boundary or hyperplane that optimally
separates data points of the two classes in Feature Hyper space [23] and also makes the margin maximized. Since the
SVM is a binary classifier, classes are first pairwise trained to build inter-class boundary functions. When applied
to the test pattern, each boundary function assignes a class to the pattern and a simple voting scheme is used to
determine the assigned class. From the aspect of implementation training a Support Vector Classifier is equivalent to
solving a Quadratic Programming Problem of linear restriction. We used the Sequential Minimization Optimization
(SMO) by John C. Platt [24] to do the same.
Conclusion and Future Work
A biopotential amplifier circuit was designed and built. Some experiments were conducted to study the application
of EEG as a biometric. Two feature extraction and three classification schemes were tested for three and four class
prblems. Relevant results are reported.
Despite having demonstrated individual stages of the acquisition circuit, much needs to be done before the dominant
50Hz hum is suppressed and a clean signal is obtained. FET based instrumentation amplifiers are low-noise compared
to the BJT based AD-521. A Driven Right Leg (DRL) circuit is a relatively new technique towards common mode
suppression in biopotential amplifiers [20]. It suppresses the 50Hz hum by cancelling it out. In our next semester, we
intend to explore the above de-noising options. If good results are obtained, the notch filter may be removed and the
signal saved from distortion. We are yet to apply ICA as a de-noising technique in software. We also intend to explore
this in greater detail the following semester.
Model based feature extraction techniques exploit certain characteristics of the signal known apriori. For example,
linear prediction analysis applied to speech signals is able to separate the exciation source from the vocal tract filter.
We intend to explore existing models for the EEG signal or build intuitive models and apply new feature extraction
techniques.
Computer Interfaces
[attachment=35487]
Abstract
The focus of our B. Tech. Project is three-fold. The first module deals with the hardware implementation of a
single channel bipolar EEG acquisition system using MSI components followed by digitization of the signal using
a PCL-818HG data acquisition card. The second module addresses de-noising the EEG signal from biological/
electrode and/or power-line artifacts using Independent Component Analysis (ICA). The third module explores the
EEG signal as tool in biometric applications. It is motivated by a surmise that the EEG signal contains genetic
information and is thus unique across individuals as well as repeatable within. Independent progress has been made
on all fronts. Depending on our progress, we also propose to implement a simple real-time EEG-BCI application
which can serve as a prototype for a thought based authentication system.
Introduction and Literature Review
The Electroencephalogram (EEG) signal was first observed and recorded by Hans Berger in 1929. Since then, it
has been been applied extensively in medicine for detecting epileptic seizures and in sleep studies. Since the 1960s,
signal processing techniques have proliferated. Application of these techniques to the EEG signal has opened up new
research directions in the engineering domain. A significant advancement in this field has aimed to provide the brain
with external channels for communication and control. Recent studies suggest that electroencephalographic (EEG)
activity can provide the basis for such new channels [1, 2, 3, 4] and these have come to be known as Electroencephalographic
Brain Computer Interfaces (EEG-BCIs). The majority of BCI applications focus on enabling patients
suffering from neuromuscular disorders such as multiple sclerosis, cerebral palsy, spinal cord injury and severe paralysis
to communicate with the external world in the absence of methods to repair the nervous system. However, there is
emerging BCI research [5, 6, 7, 8, 9] that has explored signal processing techniques in order to quantify EEG patterns
as unique descriptors of human identity. The most common biometric method of identifying persons is through
fingerprint recognition [10]. In recent years, alternative biometric methods to replace or augment the fingerprint technology
have been researched. In this regard, biometrics like palmprint [11], hand geometry [14], iris [13], face [15],
and electrocardiogram [12] have been proposed. In this respect, using the EEG as a biometric is relatively new and is
motivated from the hypothesis that EEG signals contain genetic information as proposed by Vogel in 1970 [16].
Our work seeks to establish a one-to-one correspondence between identity of a healthy individual and certain
features of his/her EEG signal towards the further goal of developing a test for person identification based on features
extracted from the EEG. Potential applications include, among others, information encoding and decoding and access
to secure information [5].
Methods
Acquisition, Sampling and Quantization
The EEG signal amplitude is only a few (50− 100) microvolts and thus has to be passed through a very high gain
amplification stage (gain of the order of 10,000), before processing. Owing to its low amplitude, it is highly susceptible
to contamination from various sources including biological, electrode and power line artifacts. We have used a BJT
based Instrumentation amplifier (AD 521) which amplifies differential voltage between two electrodes while rejecting
the common voltage ( -110dB). Typical electrode materials, such as gold or steel, are polarizable which leads to
building up of large DC-offset owing to huge surface charge accumulation. A single stage amplifier with a gain of
10,000 will amplify a typical DC voltage of 200mV to a very large value, thus saturating the amplifier. Power line noise
can also be a significant contaminant if shielded electrodes are not used. In order to forestall saturation effects, the
amplifier is implemented in 2 stages. In between these stages, mains hum and DC frequencies are attenuated. First, a
high tunable Q notch filter at 50Hz is employed. Second, a high-pass filter of very sharp roll off (-80dB/decade) and
low pass-band frequency (0.1Hz) is used to reject DC. A Sallen-Key active resonant filter with a gain of 1.58 offers
a maximally flat response with a roll off of -40dB/decade [25]. Two such filters are cascaded. Finally, a Sallen Key
low pass filter with a cut-off frequency of 30Hz is used to prevent aliasing effects. A protection circuit before IA stage
helps in isolating the circuitry from electrostatic discharge (ESD), RF and protects the user from failing circuitry.
Schematic of our two stage Instrumentation based Bio-potential amplifier is shown below.
Mental Arithmetic, A Non-cognitive task
This dataset was obtained from an open source distribution available at the University of Colorado [30]. Subjects
were asked to perform mental arithmetic. 10 trials of 10 seconds each, for 4 subjects were conducted. EEG signals
were recorded from C3, C4, O1, O2, P3 and P4. The first 5 trials and the next 5 trials were conducted on separate
days. This data was divided into training and test sets in the ratio 1:1.
Feature Extraction and Classification
For both experiments, half overlapping windows of duration 1 second were used. From each window a training/test
pattern is formed. Features used are either FFT peaks or Linear Prediction co-efficients. Number of features used
were empirically determined and were found to be 70 (per channel).
For the stereopsis experiment, we have 60 seconds of data from 4 channels giving rise to 120 patterns per subject,
each of dimensionality 70 × 4 = 280. Out of these, 80 are used for training and 40 for testing. Similarly, for the
mental-arithmetic experiment, we have 200 patterns of dimensionality 70 × 6 = 420 per subject. 100 are used for
training and 100 for testing.
Two types of classifiers were used. The first technique is a one-to-one Support Vector Machine (SVM). Vapnik
introduced Support Vector Machines as powerful learning tools based on statistical learning theory. An SVM is a
binary classifier that makes its decision by constructing a linear decision boundary or hyperplane that optimally
separates data points of the two classes in Feature Hyper space [23] and also makes the margin maximized. Since the
SVM is a binary classifier, classes are first pairwise trained to build inter-class boundary functions. When applied
to the test pattern, each boundary function assignes a class to the pattern and a simple voting scheme is used to
determine the assigned class. From the aspect of implementation training a Support Vector Classifier is equivalent to
solving a Quadratic Programming Problem of linear restriction. We used the Sequential Minimization Optimization
(SMO) by John C. Platt [24] to do the same.
Conclusion and Future Work
A biopotential amplifier circuit was designed and built. Some experiments were conducted to study the application
of EEG as a biometric. Two feature extraction and three classification schemes were tested for three and four class
prblems. Relevant results are reported.
Despite having demonstrated individual stages of the acquisition circuit, much needs to be done before the dominant
50Hz hum is suppressed and a clean signal is obtained. FET based instrumentation amplifiers are low-noise compared
to the BJT based AD-521. A Driven Right Leg (DRL) circuit is a relatively new technique towards common mode
suppression in biopotential amplifiers [20]. It suppresses the 50Hz hum by cancelling it out. In our next semester, we
intend to explore the above de-noising options. If good results are obtained, the notch filter may be removed and the
signal saved from distortion. We are yet to apply ICA as a de-noising technique in software. We also intend to explore
this in greater detail the following semester.
Model based feature extraction techniques exploit certain characteristics of the signal known apriori. For example,
linear prediction analysis applied to speech signals is able to separate the exciation source from the vocal tract filter.
We intend to explore existing models for the EEG signal or build intuitive models and apply new feature extraction
techniques.