This work presents the experimental design for recording Electroencephalography (EEG) signals in 20 test subjects submitted to Steady-state visually evoked potential (SSVEP). The stimuli were performed with frequencies of 7, 9, 11 and 13 Hz. Furthermore, the implementation of a classification system based on SSVEP-EEG signals from the occipital region of the brain obtained with the Emotiv EPOC device is presented. These data were used to train algorithms based on artificial intelligence in a Raspberry Pi 4 Model B. Finally, this work demonstrates the possibility of classifying with times of up to 1.8 ms in embedded systems with low computational capacity.
⭐ For more information visit our blog:
https://vasanza.blogspot.com/
Influencing policy (training slides from Fast Track Impact)
⭐⭐⭐⭐⭐ SSVEP-EEG Signal Classification based on Emotiv EPOC BCI and Raspberry Pi
1. SSVEP-EEG Signal Classification based on
Emotiv EPOC BCI and Raspberry Pi
Karla Avilés-Mendoza , Víctor Asanza , Hector Trivino-Gonzalez, Félix Rosales-Uribe,
Jamil Torres-Brunes, Francis R. Loayza, Enrique Peláez, Ricardo Cajo and Raquel Tinoco-
Egas
Escuela Superior Politécnica del Litoral, ESPOL, Guayaquil, Ecuador
Facultad de Ingeniería en Electricidad y Computación, FIEC
4. 1 billion
According to OMS over
people live with some form of disability, this is about 15%
of the world’s population.
5. ● First approach for a low cost
device capable of processing EEG
signals in real-time to control an
external actuator.
● More affordability for people with
low economic respurces.
● Improve response time.
● Visual stimuli comming from a
mobile device, eg: to control a
wheelchair, a robotic arm, among
others.
● To achieve a high accuracy with
low computational cost.
● Accesible and easy to use.
● To achieve a aceptable accuracy
with a faster pre-processing and
classification time.
● The light of a mobile device is not
that intense in comparison to
light bulbs.
Motivation Challenges
6. ● Zhang et al., 2015 proposed to analize beta wave for a range of [5-20] Hz in the
front parietal and occipital regions. This requires more computational capacity
for the data processing device.
● Khosla et al., 2020 proposed feature selection methods in order to choose
relevant features that contribute to a successful classification of the user’s
intentions, resulting in an increase in the quality of the later results.
● Han et al., 2018 proposed to focus on the activity of the occipital and parietal
regions of the subject, in order to obtain a high classification rate.
Related work
7. AI techniques
• Reduce the
complexity of noisy
data.
• Increase data
classification accuracy.
• Feature extraction
using temporary
windows.
• Reduce costs and
improves chances of
usage.
BCI based on EEG
• Brain waves are use to
communicate the user’s
intentions to external
actuators.
• Non-invasive
electroencephalography
(EEG) techniques.
• Low – cost devices, more
accesible
8. Challenges
• High error rate.
• Multiclass EEG data
classification techniques.
• Data acquisition is highly
noise susceptible: a blink,
a small movement of the
scalp, hair, adipose
tisssue, among others
can cause noise.
SSVEP
• Evoked potential
produced by a visual
stimuli flashing at a
specific frecuency.
• Between 6-75 Hz
9. Dataset
02
Dataset Citation:
Raquel Tinoco-Egas, Karla Aviles, Jamil
Torres-Brunes, Hector Trivino-Gonzalez,
Víctor Asanza, Félix Rosales-Uribe, Francis
R. Loayza, Enrique Peláez, April 27, 2021,
"SSVEP-EEG data collection using Emotiv
EPOC", IEEE Dataport, doi:
https://dx.doi.org/10.21227/0j42-qd38.
10. Experiment organization
● 20 adult subjects were recruited between the age of 20 – 35.
● Staff avoided wearing brightly colored clothes that could distract the subjects.
● Staff respected COVID biosecurity measures.
● Temperature: 25 degrees Celsius.
● Noise: 30 – 55 decibels (air conditioning and car passing through)
● Participants signed an informed consent.
11. Acquisition device
● Emotiv EPOCx
● Sampling frecuency: 128 Hz
● 14 electrodes (2 ground references) - International 10-10 eeg system
● Conductive gel to reduce impedance between the electrodes and the scalp
13. ● Tasks duration: 3.5 seconds
● Frecuency tasks were shown 40
times each.
● When a frecuency task was being
shown, the other squares turned
opaque by 80%.
Experimental methodology
16. Data pre-processing
● Occipital region – electrodes: O1 and O2.
● The single output files were divided into
several files containing a temporary
window.
● Files were splitted into folders
representing their respective frecuency.
● A Butterworth filter of order 20 was
applied. Frecuency limit: 5 Hz – 30 Hz.
● Outliers were extracted.
17. Data pre-processing – Data augmentation
● Data augmentation was used by applying White noise of different amplitudes to the data.
● White noise: randomly generated array of values added to the signal, so it doesn’t lose the general
behavior but the values do change.
Small amplitude white noise Big amplitude white noise
18. Feature extraction
● Extracted 21 features.
• Mean
• Mean - weight I
• Mean - weight II
• Log Detector
• Median
• Variance
• Mean absolute difference
• Mean frecuency
• Peak frecuency
• Variance central frecuency
• Maximum PSD
• Amplitude Histogram (10 ranges)
19. Data normalization
● Sklearn MinMaxScaler.
1. Normalize training data and save minimum and maximum values
2. Normalize validation and test data using minimum and maximum values from the training data.
25. ● Without data augmentation RF and MLP whose accuracy values were 54% and
52%, respectively.
● With data augmentation RF and XGBoost whose accuracy values ware 58%
and 57%, respectively.
● Over-adjustment in the classification algorithms, due to the limited number of
examples.
● Real-time responses, shorter classification time were MLP and XGBoost with
times of 1.8 and 7.12 milliseconds, respectively.
● Adequate experimental design, because cleaner data will help to improve the
classification.
● Recruit more subjects to eliminate over-adjustment in the algorithms.
● Deep Learning techniques (DL) and spectral images.