These files contemplate my scientific initiation project in digital signal processing. In this project I developed a BCI using Ultracortex "Mark IV" EEG system from OpenBCI and I proposed a non-parametric analysis of data during stress tasks through Welch's power spectrum and coherence estimator.

The BCI is composed by acquiring, filtering, plotting data and interacting with user through Montreal stress task.

In order to communicate to EEG hardware (Ultracortex Mark IV) from OpenBCI, it was used OpenBCI library for python (acquisitio/open_bci_v3.py). Initializing the program accordingly to OpenBCI specifications, it was created a queue to receive data each sampling period (sampling frequency is 250Hz).

The data was filtered with a band-pass Hamming linear phase FIR with size 256. As the data arrives the samples are filtered and accumulated in another special list that allows a fast way to manage data, which is deque variable.

All filtered data is also saved in a .txt file with 9 collums and X rows where (X=t x 250) and t is the recording time. Each one of the first 8 collums is each of the 8 channels (electrodes) that are pinned to Ultracortex and the last collum is related to recording time (notice that it's created a time reference when the program is initialized). The files at raw_data/ follow this structure and each channel was recorded according to the following electrodes position (10-20 system):

• CH0 : FP2
• CH1 : FPz
• CH2 : F8
• CH3 : Cz
• CH4 : F4
• CH5 : P3
• CH6 : T6
• CH7 : T4

It's important to highlight that all EEG data was recorded with myself.

At raw_data/ there are recordings under normal activity and stress stimulated by Stroop and Montreal tasks. Montreal data was zipped because it was too large to updload to a git repository.

The plot data thread was developed based on pyqtgraph real time interface. It is initialized once in the beggining and then it occurs a loop that rotate the deque of filtered samples and plot them inasmuch as they are filtered. There is a waiting condition that happens if the plot thread runs faster than the receive and filter thread.

This other thread in responsible for displaying the Montreal stress task based on (Dedovic et al., 2005) while EEG data is captured, filtered and displayed. It was developed by my lab partner (it's referenced in the code) using tkinter libreary and appended to a thread in the main program.

The stroop task (Stroop, 1935) application was also developed by me and it can be found in stroopstress_class.py file. To run this application, some changes are necessary :

This one is the part of main code for Motreal task (montrealstress.py must be in the same folder as main code) :

from montrealstress import Question
# Montreal
levelqueue=Queue()
question  = Question(level=1, timer = 20,levelqueue = levelqueue)
# Montreal thread
procs.append(Thread(name="test", target=question.loop, args=()))

For Stroop, these previous code lines can be erased and the following should be added :

from stroopstress_class import Stroop
# Stroop
stroop  = Stroop()
# Stroop thread 
procs.append(Thread(name="test", target=stroop.loop, args=()))

The porpose of simulations (at simulations/) is to observe the behavior of processing tools that will be used to analyse EEG raw data in frequency domain. So, in these files some examples are performed to:

• Compare Welch's method with classic periodogram (noise and frequency resolution);
• Study coherence dynamics between pairs of sine signals with different amplitudes, frequencies and phases;
• Study the frequency domain effect of random inserted peaks (this simulate muscular contractions that add noise in EEG data);

Considering the recordings at raw_data/, each data channels was fitted into 50s windows and the channels power spectrum for each raw data were estimated through Welch's method. Average power bands were calculated for each channel window (inside 50 seconds) and displayed in radar plots.

For coherence estimation, it was used Welch's periodogram for both power and crossed spectra (with scipy.signal). Coherence is a frequency dependent function and, because of this, to estimate coherence in a band an average of coherences in each frequency of band was calculated. With these mean coherences, it was generated matrixes (similar to correlation matrixes) for each task and each band (theta, alpha and beta). From these matrixes, ratio matrixes comparing coherence in specific band between some stressor task and normal behavior were generated. All of this process is performed by coerencia_EEG.py, which generates all of the ratio matrixes.

Once ratio matrixes are ready, the visual way to analyse the increase or decrease rate of mean coherence was through correlogram, which is generated by correlogram.py .

Dedovic, K., Renwick, R., Mahani, N. K., Engert, V., Lupien, S. J., and Pruessner, J. C. (2005). The montreal imaging stress task: using functional imaging to investigate the effects of perceiving and processing psychosocial stress in the human brain. Journal of Psychiatry and Neuroscience, 30(5):319.

Stroop, J. R. (1935). Studies of interference in serial verbal reactions. Journal of experimental psychology, 18(6):643.