Electrophysiology and Electroencephalography | A Paper Review
I read this research paper on Electroencephalography (EEG).
These are my repurposed notes and takeaways in article/blog format.
I’m practicing the Feynman Technique.
Enjoy.
Recording electrical signals is like reporting a soccer match.
Have you ever been to a soccer game? A football game? A basketball game?
If you have, you definitely know how loud those games can get. The crowd’s roars can be thunderous making it very difficult to hear the person right next to you.
Recording electrical activity using an EEG headset is analogous to being a reporter at a roaring soccer match.
For context, EEG is a test that aims to detect the brain’s generated electrical signals in order to measure brain waves. If you’d like to learn more, click here.
Let’s say you’re a sports reporter.
Recording the action potentials from a singular neuron is like reporting from the field and interviewing the coach. We’re able to intentionally select a very specific subject and gather data from it.
Recording the local field potentials (LFP) is like reporting from the press box. We can hear the roar of the audience and the commentary of nearby reporters. We can clearly get data from where we’re situated. But getting data from the match with a high resolution from areas further away from us, specifically the crowd, just isn’t possible.
Recording the extracellular field potential is like attempting to report a soccer match from right outside the stadium. We can only hear the fluctuations in the loudness of the stadium. No precise or high-resolution data can be gathered. This is EEG what recordings are like.
Different techniques to record signals of different quality.
Recording individual neurons requires the very precise placement of recording equipment, a very small yet very sensitive electrode, and a very high temporal resolution.
- Singular neurons measure ~10 microns
- Action Potentials can be 1 ms
- The amplitude of a neuron is a mere ~100 MV
We’d need to be very accurate to record individual neurons.
Fortunately, we can record the activity of individual neurons by using patch clamps.
Micro electrode arrays (MEAs) on the other hand, don’t record singular neurons.
Instead, they can be set to record multi-unit action potentials or lower frequency field potentials.
Multi-unit activity refers to the field potentials generated by a group of multiple neurons in a local area of the brain near to the electrodes. This involves recording groups of neurons that fire at a high frequency.
Lower frequency LFPs are the vector sum of action potentials generated by groups of neurons at areas of the brain further from the electrode.
These action potentials propagate through brain volume, demonstrating the effect of volume conduction, resulting in a field potential that’s recorded as having a lower frequency.
The lower recorded frequency in LFPs can be due to the lack of precision when recording the electrical activity.
Since we aren’t recording the signals at their precise origin, the propagation of the signals throughout brain volume induces a decrease in its frequency through signal attenuation and dispersion.
EEGs, completely contrast from MEAs and patch clamps when you consider the ability to measure precise electrical activity from precise groups of neurons.
They actually don’t have the ability to record precise electrical activity…
Instead they completely rely on extracellular field potentials.
Extracellular field potentials (EFP) are the vector sum of all electrical activity from a general brain region detected from outside it’s source.
An EEG system uses non-invasive electrodes placed on the scalp in order to measure the EFP. The electrical activity is recorded from larger areas of the brain rather than smaller and more specific areas.
Due to their application in recording EFPs, EEGs can very easily allow for the detection of brain states, determining how focused or relaxed you are.
For example, in the recorded EEG oscillations of larger brain regions, (the rhythmic, repetitive, and fluctuating patterns in electrical activity), synchronized oscillations can indicate a person is relaxed and calm.
Desynchronized oscillations, on the other hand, are a proxy of higher or lower electrical activity in certain regions of the brain than others.
This can indicate focus and concentration or relaxation and drowsiness depending on the frequency of the recorded electrical activity.
Though EEGs have easier access and usability when compared to MEAs and patch clamps, they suffer from the aforementioned volume conduction. This induces more noise and impedance to the recorded signal which makes it challenging to accurately record precise signals.
EEGs measure post-synaptic potentials
Action potentials represent the depolarization from resting neural activity to excitatory neural activity. It’s basically the difference in voltage between both states.
The resting potential of a neuron typically ranges from -70 mV to -55mV. Once a neuron reaches the threshold of -55mV, it generates an action potential further increasing its voltage by ~100mV to about ~50mV.
This depolarization of the resting neuron leads to the release of neurotransmitters (dopamine, serotonin, glutamate, etc). The neurotransmitters travel through the synapse and then binds to the post-synaptic neuron.
A post-synaptic neuron is just a neuron that receives a neurotransmitter from another neuron.
Once neurotransmitters bind to the post-synaptic neuron, an electrical signal is generated within that neuron. The signal is called the post-synaptic potential.
EEGs measure those post-synaptic potentials rather than action potentials.
Characteristics of Pyramidal Neurons assists EEG
Pyramidal neurons located in the cortex are organized in a more orderly and uniform fashion.
This is pretty darn important for EEG measurements as it allows for systemic, predictable, increasingly reliable measurements.
ML models trained on EEG data from these neurons also have an easier time cleaning, filtering, and processing EEG data.
When an excitatory post synaptic potential is generated at an apical dendrite, it results in a positive electrical charge in the neuron, intracellularly.
This positive electrical charge results in a negative electrical charge in the extracellular region from the neuron.
The relationship and difference in the electrical charge between intracellular and extracellular region is called a current dipole.
Let’s break this down.
The intracellular region (inside the neuron) is what we call the current source while the extracellular region (surrounding the neuron) is what we call the current sink.
The current source generates the electrical signal while the current sink is what the signal propagates into.
To further understand this, let’s use a faucet and sink as an analogy.
The faucet is like the current source, it’s where the water comes from. The sink itself is like the current sink, receiving the water that comes out from the faucet.
The difference in signal strength between extracellular and intracellular regions is what we call the current dipole.
An EEG is able to measure electrical activity through these current dipoles to detect electrical frequencies, measure brain waves, and ultimately determine brain state, at least if it’s backed by robust signal processing systems.
EEG Signal Recordings and Technology
Non-invasive BCI methods, specifically EEGs require a combination of various technologies.
EEG electrodes are made to act as the medium by which electrical signals travel from the source (the brain) to the processing board.
They can come in the form of either wet electrodes, utilizing a type of aerogel, or dry electrodes that are made of silver, copper, or gold to increase conductivity in absence of an aerogel.
The importance of the materials in EEG electrodes brings up an opportunity for materials science to play an important role in the development of a reliable system.
Of course, once signals are captured through electrodes, the job isn’t done. Electrical signals need to be pre-processed in order to then provide for signal and noise filtering, signal amplification, signal transforms, and signal conversions.
To do this, EEGs make use of operational amplifiers.
Operational amplifiers are leveraged to pre-process a signal through:
- Signal amplification
- Signal filtering; Such as bandpass filtering
- Providing voltage differences; Useful for signal threshold detection
- Calculating the vector sum of electrical signals
- Mathematical operations on an electrical signals.
- Converting AC signals to DC
Commercial EEG systems approved for clinical use make use of those technologies but on a much higher quality. Clinical EEG systems can acquire data from up to 128 channels….with a sampling rate greater than 10,000 Hz! This means greater than 10,000 samples per second!
But wait, there’s a catch — medical grade EEG systems can cost in the range of tens of thousands of dollars. Even up to $60,000!
The Status Quo of EEG
Currently, EEG is primarily used for measuring how well our neurophysiologic pathways respond to auditory or visual stimuli through event related potentials (ERPs).
Through this EEG can analyze how our brain functions at different levels of consciousness and awareness. Including sleep!
It determines the precise temporal dynamics of our brain function. Meaning, it can determine the time measures of our brain activity.
When our brain fired and why it fired.
Aside from analyzing ERPs, EEGs can holistically characterize our neural networks and their strength in connectivity. While it may not serve as a precise and pinpoint measurement for individual neuronal functioning through spatial recordings, it’s ability to reliably record temporal data serves as a very important feature to characterize relationships between our neural networks.
Unfortunately, there are some concerns and limitations when it comes to employing EEG systems.
EEG headsets can incorrectly record electrical signals which then lead to the misinterpretation of brain activity. It’s important to optimize the EEG headset and avoid any possible artifacts that could interfere with the attainment of a quality signal.
The aforementioned volume conduction can play a heavy role in the misinterpretation of EEG signals.
It’s important to make sure we’re aware of this in order to mitigate the negative effects that can come of it.
EEG Signal Decomposition
Through signal processing techniques and the deployment of filters, transforms, amplifiers, and other processing tools, EEG data can be decomposed into waves allowing for the easy reading of brain activity.
The brain waves, through signal processing, are decomposed to sinuosoid waves in the following frequency bands
- Delta | .2–3.5 Hz
- Theta | ~4–8 Hz
- Alpha | 8–13 Hz
- Beta | 13–30 Hz
- Gamma 30–90 Hz
Those brainwave frequencies are temporally characterized by differing measures.
Amplitude or how tall the wave is at a specific time is expressed in voltage demonstrates it’s power in relation to it’s frequency — the power spectral density (PSD).
Phase or the timing of the wave at a specific point in time relative to a reference point can be valuable information when figuring out specific ERPs.
It’s basically the relation between the position of the wave and the time it took to get there.
While the frequency bands are based on empirical evidence, they may not be fully representative of a person’s neurological or physiological state. Brain activity is very complex and saying that one frequency band corresponds to a singular neurological state isn’t accurate and can lead to misinterpretation.
Brain activity isn’t always smooth and sinusoidal. In fact at times they can be arrhythmic, meaning unpredictable and without rhythm or regularity.
So, EEGs and traditional frequency models may not be fully accurate for assessing precise ERPs.
We’d need to take a more specific look at the waves, referencing P300, P50, and other event related potentials along with the association of different waveforms such as Lambda, Positive Occipital, and Phantom waves to brain activity.
Yet, due to their ability to record with a high temporal resolution, they can be useful for determining the ERPs from the EFPs. When a broad view of brain activity is necessary, EEGs can then be reliable.
A reference problem is present as well. There isn’t a uniform state within the brain. It’s constantly moving, so accurately attempting to find a measure for ‘baseline’ brain activity can be difficult. In addition, a ‘baseline’ state can be different for each and every individual person.
When creating an EEG system that can reliably measure a specific metric of brain activity across large populations, it’s important to overcome this reference problem. Especially when aiming for consumer EEG use.
The use of machine learning models can fortunately mitigate this problem and lead to more reliable EEG systems.
Neurofeedback and Signal Processing
Neurofeedback is a type of biofeedback that uses brain activity as input in order to generate feedback based on that brain activity. This allows users to modulate and guide their brain state to the ‘ideal state’.
To accurately and reliably do so, signal processing plays a crucial role.
EEG signal processing consists of multiple steps, all undertaken in order to mitigate the effect of possible noise from contaminating the signal.
The noise that contaminates the needed signal is referred to as artifacts.
The first step of the process is to collect brain data. This is done through specific protocols which measure the brain activity of a user that correlates with a specific action or thought process.
Electrodes are attached to the scalp and as the subject performs the prescribed protocols, the brain activity is picked up through those electrodes.
This is referred as signal acquisition.
The next step of the process is to pre-process the data which involves the conversion of AC signals to DC, signal amplification and signal filtering.
At this stage, we attempt to de-noise the signals and remove the artifacts.
Afterward, we perform feature extraction which involves the classification of signals into specific categories based on specific features such as frequency, power, temporal characteristics, or a combination of them all.
Machine learning models and other algorithms such as support vector machines (SVMs), linear discriminant analysis (LDA), neural networks, and gaussian classifiers, should be typically used at the stage of feature extraction and classification.
Once the useful signals are extracted, they’re utilized as neurofeedback, where the user can dynamically modulate their brain activity in real-time.
Neurofeedback can be communicated to the user in the form of visual or auditory signals allowing for the potential gamification of this technology.
Mapping the Brain with EEG
The location of intracranial sources of electrical activity can be estimated as a function of time based on the electrical recordings at the scalp using EEG
What’s cool is that simulations and empirical findings have indicated that the prediction of these intracranial sources only have errors on localization less than 1 cm.
So what if we combine this with computational methods…
Computational Methods + A high grade EEG headset = Reliable Brain Map?
In conjunction with advanced computational neuroscience, EEG technology can allow us to map out the brain on the computer and gain a deeper understanding of brain functions.
Computational neuroscience has allowed for the modeling and prediction of neural firing and modeling of neural networks. They prove to be extremely helpful in identifying the ERPs from external movements.
Mapping the Brain with EEG
Single trial analyses are being considered to further understand data variability in smaller groups and individuals.
Single Trial Analyses refers to to study methods that consider the individual change between subjects.
Without single trial analyses, brain signals and ERPs are extracted from the brain by averaging a signal under repeated presentation of stimuli.
But this doesn’t consider the change that can occur within a subject during signal acquisition trials.
A subjects brain state can change depending on their levels of awareness, restfulness, and stress. Therefore, taking the average of a signal out of multiple trials, especially if they’re spaced out, can be redundant.
This is where single trial analyses, analyzing trials individually considering the variability of the subject, comes into high importance.
Machine Learning
Developments in advanced machine learning techniques can more accurately and reliably help detect the subject physiological state. Allowing for advanced signal processing during EEG trials will help denoise the signal and remove artifacts.
ML models do this through discriminant functions applied onto the ML model. Varying models and methods such as K-NNs, SVMs, LDA, Decision Trees, and more can all be applied to identify key features.
Key features can be performance based, meaning the identification of the key features in a signal is based on their level accuracy or alignmnet to the current dataset.
They can also be measured through the neurobiological significance. The identification of the features are based on the significance to underlying brain processes. This is done through the prior knowledge of what differeing brain functions represent.
