Electroencephalogram (EEG) signals are records of the superficial electrical activity of the brain. These signals are the basis for brain-computer interface (BCI) systems where they are used to control a variety of external devices, and act as a communication channel between the human brain and the computer. Moreover, BCI systems translate the brain signals into commands used to control external devices in a variety of human performed tasks, and as such have a wide use in a number of biomedical engineering systems (e.g. [1, 2]). Furthermore, research in the use of AI systems for the analysis of EEG signals could be a very important direction in the detection and prediction of neurological diseases such as epilepsy, brain tumour and development abnormalities (e.g. ). The detection and prediction capabilities of AI systems are achieved through the implementation of various machine learning classification algorithms which tasks is to learn and recognise patterns from the labeled dataset extracted from the raw EEG signals.
The main steps involved in extracting meaningful information from EEG signals include signal processing, feature extraction and feature classification [4, 5, 6]. Any good implementation of such BCI or AI systems requires an efficient feature extraction scheme, capable of extracting relevant information from the EEG signals and resulting in satisfactory classification performance. Although optimal classification algorithms could be used to improve the classification accuracy, the selection of inadequate features could still lead to poor classification performance (according to the general machine learning principle “garbage in-garbage out”). Therefore, extracting relevant and meaningful features from EEG signals represents a crucial task which needs to be performed before the classification process.
In order to provide an adequate analysis of EEG data, it is very important to initially extract as many meaningful features as possible. However, EEG signals are time-varying which makes the feature extraction and classification of these signals a very challenging task.
A number of feature extraction methods have been developed and reported in literature. These include a variety of time domain, frequency domain, as well as time-frequency domain features (e.g. Stockwell transform, and the wavelet transform (WT) based features already discussed here). Although, the most commonly used methods for EEG feature extraction are based on frequency analysis, for example discrete Fourier transform (DFT) or power spectral density (PSD), WT-based features are still considered to be the most effective techniques for dealing with time-varying EEG signals. Examples of such features used in clinical trials are wavelet entropy, wavelet coefficients, and wavelet statistical parameters (e.g. mean, median, and standard deviations).
As mentioned in the previous blog post, two types of WT methods exist: continuous and discrete. The latter is more widely used in real-time applications. Furthermore, by using a base function (i.e. mother wavelet) such as the one in Figure 1, and a discrete set of scale and shift parameters, signals are decomposed into multi-resolution subsets in an iterative fashion . Such a decomposition allows for the separation of low and high frequency behaviours/components of the signal. The relationship between the scale and the behaviour of the system is as follows: low scales “compress” the wavelet and provide the fine details or high frequency components of the signal, while high scales “stretch” the wavelet and provide the low resolution approximations of the signal. The low and high frequency information of the signal obtained during the decomposition process is expressed in terms of approximation (cA) and details (cD) coefficients. As the signal frequency range of interest varies depending on the type of signal under examination (as well as the type of application the signal is being used for), different levels of WT decomposition should be applied to different signals.
Figure 1: An example of Morlet wavelet function.
In the next blog posts of this series, I will focus more on the other EEG features that can be extracted for classifying epileptic seizures. More so, the details of the most significant feature extraction methods will be discussed.
Thanks to Andrew Simmons and Nicola Pastorello for their help and useful suggestions/corrections.