EEG Signal Processing for Dummies
As promised in my previous post about Event-Related Potentials, I will explain the basics and standard steps commonly used in the analysis of EEG signals. There is a lot of literature and many concepts are involved in the field of EEG signal processing, and some of them can get very technical and difficult. That is why my aim in this post is to try to give a general overview of the different concepts without getting into too much detail.
Figure 1: Basic steps applied in EEG data analysis
As we can see from figure 1, the first thing we need is some raw EEG data to process. This data is usually not clean so some preprocessing steps are needed. These often include the application of filters, such as a high-pass filter to remove the DC components of the signals and also the drifts (usually a frequency cut-off of 1 Hz is enough). A low pass filter can also be applied to remove the high frequency components. In EEG we currently rarely study frequencies above 90 Hz which correspond to the Gamma range. There are many other preprocessing techniques such as Electrooculogram (EOG) artefact correction, which might be necessary to apply if the subject under recording is keeping his/her eyes open. The reason is that blinks and eye movements generate strong electrical fields that affect our EEG recordings. Once our signals are clean, i.e. preprocessed, it is quite common to cut them in epochs of a few seconds and then extract features out of each one of these. This allows us to have a large number of features from a single EEG recording, which is always good when performing statistics or when applying classifiers.
2. Feature Extraction
The next step could be considered the most important one: feature extraction. EEG signals are complex, making it very hard to extract information out of them using only the naked eye. Nowadays, thanks to computers, we can apply complex automatic processing algorithms that allow us to extract 'hidden' information from EEG signals. There are several techniques such as time domain features (mean, standard deviation, entropy, …), frequency domain features (Fourier transform, wavelets, …) and finally synchronisity features, which looks to the relationship between 2 or more EEG channels (coherence, correlation, mutual information, ….), just to mention a few.
There are other feature extraction methods that are worth mentioning, such as EEG tomography, that allows us to compute the regions inside the brain that are active (applying the so-called inverse-problem approach). This in turn usually needs a high number of electrodes (at least 32, although even better with 64 or more) in order to achieve a decent spatial resolution.
Figure 2: Example of a graph. Each node would represent an EEG electrode. If two nodes are connected, it means that the corresponding EEG signals are similar enough.
We can also apply more advanced methods such as converting our EEG recording into a graph in which each node represents an electrode and the connections of these nodes depend on the similarity of the EEG signals of each electrode. Once we have our graph we can then analyse its properties using standard complex network analysis techniques.
3. Feature selection
The next step, called feature selection, is optional and is used in the case in which we have a large number of features and we want to study the ones that are more relevant for our study. Imagine that we want to look for possible differences in the EEG in two different conditions: relaxed versus stressed. We can apply feature selection techniques to find out among our large number of features the ones that are more discriminative between these 2 conditions. We can apply statistical methods such as principal component analysis (PCA) or more complex techniques such as genetic algorithms.
Once we have our (selected) features we can plot them, or apply some statistics to them, close our research and write a potentially nice paper, or even go one step further applying the classification step. Using machine learning techniques, we can train a classifier to recognise from among our features which ones belong to one class (or condition, i.e. relaxed,) or to another (i.e. stressed condition). This is a very powerful technique and it is extensively used in EEG data analysis. For instance, all brain-computer interface systems follow this common scheme, in which the classification step is performed in order to decide what the user is thinking.
Well, that's all for now. I hope this post gives a helpful overview. You can dive deeper into EEG signal processing concepts by clicking on the hyperlinks provided in the text above, and you can also leave any comments or doubts below, and they will be answered as soon as possible.
Figure 2 credit: ilamont.com via photopin cc