|M.Sc Student||Kaidar Ruthy|
|Subject||EEG Based Brain-Computer Interface for Movement Control|
|Department||Department of Mechanical Engineering||Supervisors||PROFESSOR EMERITUS Gideon Inbar (Deceased)|
|PROFESSOR EMERITUS Moshe Shoham|
A Brain Computer Interface (BCI) is a direct communication channel between the human brain and a computer or another external device. A BCI is intended for people affected by severe neuromuscular disorders, which leave them partially or completely paralyzed, often without the ability to communicate with their surroundings, remaining in a locked-in state. A BCI uses electrical brain activity to detect the user's intent to perform a movement, to spell words or to activate external neuroprosthetic devices.
Gamma-band, in the range of 28-40 Hz, is known to be associated both with attention and with voluntary motor tasks, serving as an ideal candidate for a driving input into a BCI device.
In the present study, gamma-band power and time-domain predictive features serve as two separate dimensions of input into a BCI device. Time-domain features predict intent to initiate movement and gamma-band power features determine laterality of the desired movement.
A 22-channel scalp EEG system is used for recording the electrical activity from subjects. Subjects are presented with two types of auditory selective-attention tasks, containing one or two target stimuli, and requiring a motor response.
Time-domain data analysis showed a difference between movement and non-movement segments in the 360 ms prior to motor response, suggesting that the time-course preceding the actual motor response can be used to predict intent to initiate movement and thus as a predictive tool in a BCI device.
Frequency-domain data analysis was based on the spectral power of the signals and estimated using two spectral estimation methods: Welch's method and the Multitaper method. Estimation using both methods showed a significant difference in spectral power in the gamma-band frequency range, between left and right movement tasks. Multitaper-estimated data exhibited smaller variance than Welch-estimated data, leading to better left-right classification results using Multitaper-estimated features. No subject training is required. Thus, gamma-band activity, exclusively, is sufficient to determine the laterality of the desired movement.
Single-trial classification was performed using Support Vector Machines. Single-trial classification between movement and non-movement segments, using time-domain features, resulted in 76% average accuracy. Single-trial classification between left and right movements, using Multitaper-estimated frequency-domain features, resulted in 77.9% average accuracy, compared to 68.8% average classification accuracy using Welch-estimated features.
Comparing the two types of auditory selective-attention tasks showed that gamma-band activity was maximized when subjects attended to a single target, as opposed to multi-targeted attention. Hence, gamma-band activity is state-dependent (attended side) rather than response-dependent (movement side).