Browsing by Subject "Modeling, canonical correlation analysis (CCA)"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
Item Open Access New approach for designing cVEP BCI stimuli based on superposition of edge responses(2020-01) Yasinzai, Muhammad NabiElectroencephalography (EEG) based brain-computer interfaces (BCIs) are widely used in the field of neural engineering, due to their portability, noninvasive nature, and high temporal resolution properties. Among the different BCI modalities, code modulated visual evoked potentials (cVEP) are very popular due to their high classification speed and accuracy. Over the years, various cVEP stimulus sequences have been designed aiming to increase the classification speed, accuracy, and the number of supported targets. This study is carried out in order to present a novel cVEP stimulus sequence designing methodology, which is purely based on characteristics of the actual brain responses to visual stimuli. Seven male subjects participated in our study, and they were presented pulsetype visual stimulus sequences on a monitor with 60 Hz refresh rate (each bit of a stimulus sequence is presented for 16.67 ms). EEG was recorded using Brain Products V-Amp (16 channel) EEG Amplifier from O1, Oz, O2, P3, Pz, P4, P7, and P8 positions at the rate of 2000 sps and the recorded EEG was then bandpass filtered between 4 and 40 Hz. Electrode impedances were kept under 10 KOhms, and Canonical Correlation Analysis (CCA) was used to reduce the 8-channel data to a single signal. Matlab, along with psychtoolbox, was used for stimulus presentation on a PC with Ubuntu operating system. In the first part of this study, our aim was to reconstruct the EEG response to pulse-type stimulus patterns by superposing the EEG responses to simple stimulus patterns. It is observed that the EEG response is only sensitive to the changes in the stimulus sequence, that is, to positive (change from Black to White) and negative edges (change from White to Black). The edge responses have a delay of around 50 ms, and these responses can be observed up to 350 ms after the edge. Furthermore, the magnitude of the positive edge response is much larger than the negative edge response. Edge responses for every person are unique, and they are also repeatable. The 7 subjects of our experimental study have an overall average correlation of 84% between the positive edge responses obtained with two weeks of separation. It is also interesting to know that edge responses for all of the subjects have a similar overall pattern. A series of experiments are then carried out to determine how well the EEG responses can be predicted by superposition of the edge responses. The reconstructed and measured EEG responses are compared for different pulse widths, different pulse separations, and also for different pulse repetitions. It is observed that response to 1 and 2 bit wide pulses can be predicted accurately for all subjects with an average correlation of 70.3% and 68.1%, respectively. Further, for 1 bit wide pulses, if the separation between two pulses is 4 to 9 bits, the correlation between predicted and measured responses is above 51.5%. For 2 bit wide pulses, if the separation is between 3 and 9 bits, the correlation is above 53.4%. Furthermore, responses to repeating 2 and 3 bit wide pulses can be predicted with a correlation of up to 62.4% and 59.2% for 4 and 5 repetitions, respectively. In the third part of this study, we constructed 120 bit stimulus sequences based on the constraints explained above and compared them with two other types of stimulus sequences in the context of a BCI speller application. The proposed BCI speller consists of 36 targets that are presented as a 6x6 matrix on the monitor screen at the refresh rate of 60 Hz, and the experiments are performed on seven healthy subjects. The classification results of our BCI speller follow our expectations based on the second part of our study; in that, for our proposed BCI stimulus sequences, the accuracy and ITR are recorded to be 95.5% and 57.19 bits/min, respectively, whereas for the other two types of codes, the classification accuracies are 6.94% and 10.53% with information transfer rates (ITR) of 1.7 bits/min and 10.53 bits/min, respectively.