Hardware
Our prototype utilizes OpenBCI materials in order to construct a low-cost, first generation electroencephalographic signal acquisition device. We elected to utilize the eight channel Cyton Biosensing board in order to optimize cost without sacrificing too much spatial resolution. Other materials include gold cup electrodes (wet), bluetooth to serial dongle, Ten20 conductive paste, and 3M micropore surgical tape.
Data Streaming Software
Our first major component of the software prototype is the data streaming layer. We needed to construct a program that would be able to read data from the OpenBCI board as long as necessary, and stream to a save file without dropping data points. This program needed to connect to the USB dongle through the computer serial port and send signals to the board to initiate transmission, regulate sampling timing and stop streams. Here, we used the Brainflow API. The raw data from the board is streamed directly to a file in comma-separated values (CSV) format. Our data streamer class has two methods: open() and close(), which govern the data stream from the board to the file using Brainflow.
Visual Stimulation Software
The next key component of the software is the stimulator. This program needs to generate stimuli and record the exact time that the stimuli is shown to users so that we can properly slice and extract relevant sections of the EEG data. Since we are utilizing event related potentials, the exact timing is crucial to identify the necessary signals. We used the Psychopy package to create a Python script that would generate stimuli and record the exact times that the stimulus fed to the user changes. Since Psychopy is able to call functions and record the time on frame changes, we are able to store the exact time that the stimulus changes. In the video on the left, the green dot is the target that we want our subject to recognize and the red dots are the distractors.
Data Processing
To process our data after acquisition, several steps were taken. A bandpass (5-50 Hz) and notch (60 Hz) filter were used for the removal of noise and to isolate relevant brainwaves. To minimize any problems that could arise from direct current voltage drift, data was both normalized and detrended. Then, signals were transferred from the time domain into the frequency domain using the fast Fourier transform (FFT), which benefits processing speed by decreasing the number of calculations needed for analysis. In order to represent the signals' power at any given frequency, power spectrum and power spectral density (PSD) plots are used. From this data, event-related potentials were extracted from the noise and averaged versus time (0-500ms) for all 8 channels. P300 was detected, which is a positive deflection that happens approximately 300ms following the presentation of the target stimulus. Finally, linear discriminant analysis (LDA) was utilized to perform dimensional reduction.
Classifier
In order to train a support vector machine (SVM) model, about 30 single subject samples were evaluated with 10 fold cross-validation. The collected brain waves were then classified as neglected or recognized stimuli. This data is used by the software to create a reconstructed map of the patient's field of awareness and quantify neglect severity by calculating the size of the neglected area.
