Article Preview
Top1. Introduction
Electroencephalography, or EEG for short, is an electrophysiological process to record the brain’s electrical activity using EEG electrodes placed on the scalp (Kosmyna and Lécuyer, 2019). EEG data are hard to deal with, as they can be very noisy. Neuroscientists, biomedical engineers, and clinicians often receive years of training to understand and extract meaningful EEG data information. Moreover, the raw recorded data must be processed before it can be viewed by professionals. Raw EEG data is just a discrete-time multivariate (i.e., with multiple dimensions) time-series. The number of EEG channels defines the dimension of each point in the time series. Each time point corresponds to an EEG sample obtained at the same time point. The number of points in the time series depends on the recorded time and the sampling rate (Perronnet et al. 2016). These raw signals are rarely used because they may include DC offsets and drifts, electromagnetic noise, and artefacts that must be filtered. Temporal and spatial filtering is frequently used to remove noise, filter out artefacts, or isolate an enhanced version of the signal of interest. Hereafter, frequency filters like band-pass filters or low-pass are applied on the clean EEG data to isolate the bands of interest and remove those frequencies of no interest as the human brain processes in the P300 evoked response that occurs in the Theta band (4-7 Hz).
Feature extraction is applied to extract important features from the cleaned EEG data (Dry et al. 2020). Before the popularity of deep learning, feature extraction relied on custom methods of the brain’s interest process ranging from handcrafted features to more complex technologies, such as linear and nonlinear spatial filtering. The following ranges from general methods such as independent component analysis and principal component analysis, to more specific EEG methods such as CSPs (Blankertz et al. 2008) and (Ang et al. 2008) for energy features and (Rivet et al. 2009) for the temporal ones. The features extracted are usually fitted to the preferences of a specific application, such as finding differences between experimental conditions, distinguishing between a set of predefined categories, predicting behavior, finding anomalies in relation to a normative database. The current state of the art technologies includes Riemannnet based classifiers, filter banks, and adaptive classifiers, used to deal with EEG data challenges with various levels of success (Lotte et al. 2018).
Once features are ready, the processed EEG data can then be inspected visually to recognize anomalies, alterations in mental status, or to examine average responses in a number of people. However, the visual inspection process is tedious, long, and costly. It does not scale accurately and cannot be ported to BCI applications. In contrast, Artificial Intelligence (AI), in particular, Machine Learning (ML), is an ideal approach to automating, expanding, and enhancing the analysis of EEG data. Automated machine learning approaches can be efficiently applied to solve the time series classification problems like EEG. A favorite type of machine learning is supervised learning, which uses a set of examples called training data to learn a model that can predict, classify, or identify EEG patterns based on the extracted features. A generous variety of techniques exist. The most well-known are classification techniques, which classify an EEG pattern into one of a set of predefined classes or regression techniques that transform the EEG pattern into a different signal (for instance, motion direction). Adopted approaches include simple linear techniques (Multiple Linear Regression and LDA for classification), SVM like kernel methods, K-Nearest Neighbors, neural networks, and many more.