Objective

This research pursues the development of a real-time Digital Signal Processing system capable of detecting and classifying EEG patterns associated with the mere preparation of specific voluntary movements by a subject. The classifications achieved through this system will be used for the control of a computer cursor, effectively implementing a “Brain-Computer Interface”, or BCI.

Hypothesis

The preparation for execution of voluntary movements has been shown to influence the EEG measured at the scalp in at least two identifiable ways:

(1) The “Event Related Desynchronization” (ERD) or disruption of synchronized activity in the motor cortex that sets an idling low-frequency (8-12Hz) “mu-rhythm” over the motor cortex.
(2) the transient occurrence of characteristic spatio-temporal patterns in the EEG field on the side of the head, termed “Readiness Potentials” (RPs).

We propose that a Brain-Computer Interface for cursor control in a two-layer scheme could be developed once desynchronizations due to motion preparation can be detected and analyzed correctly.  We have identified two characterisitic EEG patterns from each side of the head that can be assigned to the four cursor directional commands: UP, DOWN, LEFT, and RIGHT, as a first step in developing the Brain-Computer interface.

Method

When an individual is not involved in planning a voluntary motion, the EEG measured at the side of the head (over the Supplementary Motor Area (SMA)) of the cortex seems to be dominated by an EEG signal that is of near sinusoidal rhythm at about 8-12Hz in frequency.  This rhythm, commonly termed the “mu-rhythm”, can be recorded with minimal phase difference   throughout the side of the head using electrodes.  It is also known that this mu-rhythm will collapse locally and patterns of activation will appear (Readiness Potentials) which are specific to neuronal populations involved in the preparation of certain movements.

Presently the approach taken in the development of the Brain-Computer Interface (BCI) system involves the ability to recognize two characteristic EEG patterns (finger vs. toes) from each side of the head, which will be assigned to the four  directional cusor movements (UP, DOWN, LEFT, and RIGHT).  The proposed BCI consists of two modules:

[1]  On-Line Pre-Selection (OLPS) Module:  This module will perform a continuous, real-time analysis of data, in charge of pre-selecting time intervals of the incoming EEG that are likely to correspond to one of the target forms of motion preparation.  It should be able to detect a mu-rhythm collapse and send as output the EEG data corresponding to the neuronal pattern.

[2]  Command Verification & Identification Module:  The pre-selected windows of data corresponding to that portion of the collapsed mu-rhythm will be subjected to more specific analysis in this module, to determine if one of the four cursor commands is present or if none can be verified in the segment.  Dynamic neural networks will be used, such as the Time Delay and Gamma neural networks, to classify the EEG patterns.

Preliminary Results: Mu detection

 

Electrode Placement:
Research_003_03.jpg (12073 bytes)

 

Recording

The following charts were obtained from an off-line analysis of a recording session in which the subject was asked to produce and suppress the Mu rhythm in 30 Sec time intervals.  The first 30 Sec time interval of each file, the subject was asked to produce the Mu rhythm.  The Electroencephalogram signals were amplified and digitized at a sampling frequency of 500 samples/sec.  The off-line analysis was performed using the MATLAB software.
 

 

I. Off-line Analysis:

Frequency analysis on the Electroencephalogram signals was  performed using two Goertzel algorithms.  The two frequencies investigated using the Goertzel algorithms were 6 Hz and 12 Hz.  The Goertzel algorithm provides a meaning result after N samples of the signal being studied have been processed. In our experiments we chose N = 500 such that within a given 30 Sec time interval,  a total of 24 windows can be taken.  Therefore, within a 30 Sec interval in which the subject was asked to produce the Mu rhythm, there should be 24 Mu rhythm detections.  A Mu rhythm detection was determined when the following was true: (Power in 12 Hz band)/( Power in 6 Hz band) > 1.

II. Interpreting the charts:

The blue bars corresponds to Mu rhythm detections made by the algorithm which implemented the two Goertzel algorithms.  The purple bars corresponds to Mu rhythm collapse detections made by the algorithm.  Therefore, in the 30 Sec time intervals in which the subject was asked to produce the Mu rhythm, there should be a total of 24 Mu rhythm detections (i.e., blue bar -> 24 and purple bar -> 0).  Similarly, in the 30 Sec time intervals in which the subject was asked to suppress the Mu rhythm, there should 24 Mu rhythm collapse detections (i.e., purple bar -> 24 and blue bar -> 0).

In the charts below, the first window corresponds to the subject being asked to produce the Mu rhythm, the second window to the subject asked to suppress the Mu rhythm, the third to the subject being asked to produce the Mu rhythm, and so forth in an alternating fashion (i.e., Mu, No Mu, Mu, No Mu, etc.).

Research_003_04.jpg (32135 bytes)
Research_003_05.jpg (34231 bytes)

The frequency analysis done on the EEG signals focused on monitoring two frequency bands: 6 Hz and 12 Hz.  The ratio between the frequency bands indicates the amount of “Mu power” in the signal.

 

Preliminary Results: Readiness Potentials Classification


I. Command Verification & Identification:

 

Research_003_06.jpg (66627 bytes)

 

II. Index Finger Single Trail:

Research_003_07.jpg (157391 bytes)


III. Toe Single Trail:

 

Research_003_08.jpg (151101 bytes)


IV. Dynamic Neural Networks:

Dynamic classifiers such as the Time Delay Neural Network (TDNN)  and the Gamma Neural Network have been used to classify toe and finger patterns.  Their performance was examined in terms of their Receiver Operating Characteristics (ROC) Curves.

Research_003_09.jpg (46677 bytes)


Two observations were made through our experiments:

1).  The longer the delay line in the network the better the results were.
2).  The Gamma neural network worked better for our purposes.