A small story behind it. I am a third-year student and have always been interested in how the brain functions. In November I was lucky to have an opportunity to work with the Emotiv Epoc+ neuroheadset in one of emotiv labs. One can consider this article as a shortlist of what to pay attention to when working with EEG and particularly Emotiv Epoc+.
What is an electroencephalogram?
Electroencephalography is a method of monitoring (performance metrics) the brain based on recording its electrical activities. There are millions of neurons in our brains that are constantly exchanging information.
Thus, when a large group is jointly activated, the potential difference becomes so significant that the device can read it. Rhythmic activity generators (alpha, theta, gamma…) also contribute to the EEG (EEG data) by activating and suppressing systems.
A high-frequency rhythm generator is working when you are trying to solve a problem using your brain.
If you are relaxed, then the alpha rhythms generator is working, etc.
Activating and suppressing systems allow you to switch between rhythms. For example, they make you feel sleepy in the evening and fall asleep, and wake you up in the morning.
To understand what the sensor signal refers to, you need to roughly understand the functional differences in brain zones.
If you go from the forehead to the back of the head and enumerate the zones one after another, you’ll find out that the forehead area is the center responsible for decision-making and program selection with help of an emotive headset..
After it goes the motor cortex, which provides voluntary movements, and the somatosensory cortex, which is responsible for the body perception.
Further, there is the associative cortex, which is responsible for our “world map”, i.e. the meaning of words, abstractions, the names of things. It connects sound and visual perception.
The visual cortex is placed in the occipital area, and the auditory cortex is just above the ear.
The headset has at least a couple of sensors in each of the main zones, which we are okay with.
If you want to read/listen more about the work of the brain, then use an open-source course on the “Public Education” website on the Neurophysiology of Behavior. There is also a good course on Computational Neurobiology developed by HSE jointly with Jet Brains. Both courses provide detailed information about the work of the brain.
Emotiv Epoc+ neuro headset
Accuracy
Should we be satisfied with the accuracy of a neuro-headset, which costs a lot less than lab devices, doesn’t have the largest number of sensors (14), and not the highest operating frequency (128Hz)?
Here you will find a comparison of the Neuroscan encephalography with the Emotiv Epoc+. It’s for those who are too lazy to read; the quality is not that worse, although since it’s not updated so frequently, there are slight differences in the EEG.
Free software
You can use the free CyKit library to receive a signal from the neural interface. It requires Python 3.7, but it’s only available for Windows so far.
In the repository, you can find scripts with examples. When a headset is connected, vectors of 14 numbers are printed 128 times per second. This is how the data looks like:
Issues that may arise when recording data
Artifacts from facial muscle movements, blinking, physical movements
An example of artifacts from blinking. Rhythmic (without strong splashes).
The following diagram shows examples of facial muscle movement artifacts. For comparison, while recording, I raised my eyebrows, lowered my eyebrows, and just sat relaxed. The graph shows the result in green, red, and blue, respectively. As you can see, the movement of the facial muscles has a fairly strong effect on the encephalogram.
You can use the preprocessing – independent components analysis (ICA) module mne library to remove artifacts. But some convolutional neural networks may learn these signs well separately and process them correctly.
Other issues
Dried sensors, shifted electrodes, interference from operating devices: all this affects the EEG recording.
We should have a separate talk about correlation.
Judging by the difference between the T7 and T8 channels, the T8 sensor does not fit well or is dry.
Data correlation from parallel sensors and from sensors located in adjacent zones can be observed. Although for some methods it decreases the prediction accuracy, a deeper analysis will reveal some regularities.
Generally, the signal from the sensors can be considered as the degree of activity of the brain zones. The zones communicate with each other. For example, zone A gives a command to zone B, zone B does some work and responds to zone A. For a short time, you will notice a correlation between the two zones, but with a slight difference in time.
Experiment setup
What time interval can be used for classification
To classify emotions, it is convenient to use a time interval of 1 – 10 minutes, depending on the task. This is the first example. Then the same example, but with additional technical details.
An interval from 100 ms to 1-2 seconds can be used to classify tension/relaxation. Example (in Russian), example 2 (in English, Malaysia)
Presentation of incentives
There are the following types of evoked potentials, depending on the modality of the presented incentives.
- visual
- auditory
- somatosensory
- tactile
- olfactory
- taste
- vestibular
- kinesthetic
The evoked response time to the first four stimuli is very similar in healthy people (100-200 ms). The response time to the last four stimuli is strongly individual. It should take this into account when setting up an experiment.
Search keywords P100, N100, P200, N200
If you are interested in the research of cognitive processes associated with the presented incentives, then in this case the greatest burst of activity occurs 300 ms after the presented incentive.
What signs can be used
There are many ways to extract signs to work with EEG. The most common are:
- Standard deviation (numpy.std)
- Quantiles
- Frequency characteristics
- Coefficients of wavelet expansions (pywt.wavedec)
- Fast Fourier transform (numpy.fft)
- Short-time Fourier Transform
- Shapelets method
- Integral estimation of the spectral density of the signal energy (Welch’s method)
- Separation of signals by rhythms (alpha, etc. …) and their separate analysis (scipy.signal.butter)
From my own experience, with records of half a second or more, one of the best indicators is the standard deviation. In combination with division by rhythms, the quality increases a lot (simple experiments on standard sklearn classifiers show the quality of about 98%).
For wavelets, you can select levels 4-7. Actually, the combination of these three signs showed the best result.
I hope something came in handy. The article will most likely be supplemented and corrected.
