Emerging trends in EEG signal processing: a systematic review

Electroencephalography (EEG) is a diagnostic technique used for brain electrical activity. It is widely used for data analysis, evaluates time and frequency sequences by measuring voltage fluctuations generated by ionic currents in neurons.EEG captures changes in synaptic transmission; when an action potential reaches an axon terminal, neurotransmitters are released, leading to excitatory or inhibitory postsynaptic graded potentials in postsynaptic membrane cells. These potentials generate ionic currents that produce local field potentials. Synchronized activity between characteristic pyramidal neurons in the cerebral cortex can be detected as an electric field. The steady direction of these neuronal currents can be prevented from canceling out, resulting in a significantly stronger overall electric field. This summation process facilitates EEG measurements, which exhibit nonlinear and nonsmooth properties. For prolonged EEG recordings, visual inspection becomes time-consuming and can be inaccurate due to artifacts.

To overcome the challenges and obtain faster and more accurate results, computer-aided techniques are utilized to process and analyze brain EEG signals.EEG signal processing and analysis typically involves four basic steps:

1) The raw signal is pre-processed, in which techniques such as filtering are used to improve signal quality;

2) Extract key features from the preprocessed signal;

3) Select methods to optimize these features and enhance them;

4) Use machine learning models or statistical tests to diagnose or identify various functional brain states.

The main objective of this paper is to explore the large number of EEG-based research applications that exist in the existing literature. The following figure shows an overview of the contents of this paper.

EEG signal

EEG uses metal electrodes attached to the scalp to continuously capture electrical activity in the brain. Even during sleep or relaxation, neuron cells continue to produce electrical currents as a natural form of communication. EEG is analyzed and classified using different techniques.

The brain is anatomically divided into three main parts: the hindbrain, midbrain, and forebrain, each of which is further divided into two hemispheres. And the cerebral cortex includes the frontal, occipital, parietal, and temporal lobes, the four main lobes. The central nervous system integrates and processes sensory data and then passes it on to higher brain functions that are responsible for functions throughout the body. For example, the frontal lobe is responsible for personality, emotions, and complex cognitive functions; the temporal lobe processes auditory and other sensory information; the parietal lobe is primarily involved in language processing, attention, and sensory perception; and the occipital lobe plays a crucial role in the visual field.

Scalp recordings capture signals with frequencies below 40 Hz and amplitudes within 100 μV.EEG rhythms encompass a frequency range between 0.5 and 40 Hz, which can be further categorized into four major waves, as shown in the table below. The different frequencies of EEG signals have their own unique characteristics, which are described below:

1. Delta wave: This band is observed in both the temporal and parietal lobes of the brain, is characterized by a high amplitude, and is usually seen in adults during deep sleep or when the brain is deprived of oxygen. In addition, this band is the dominant band in the infant brain;

2. Theta wave: manifests itself mainly during adolescence and is related to the individual's emotional and mental state. The wave becomes especially pronounced when adults are in a negative mood;

3. Alpha wave: the highest-frequency wave in the EEG signal, usually observed in the posterior regions of the brain on both sides. This wave is usually seen when the brain is awake or relaxed with the eyes closed and is a fundamental part of the EEG signal;

4. Beta wave: This wave is mostly displayed on both sides of the brain and indicates increased excitability in the cerebral cortex. It usually indicates increased neuronal activity in the cerebral cortex and is associated with excessive mental stress;

5. Gamma wave: This band is observed in the sensory cortex and represents high-frequency brain waves associated with perceptual and cognitive processing.

EEG signals are widely used in a variety of disciplines, providing valuable insights into brain activity and cognitive processes.EEG is also an important diagnostic tool for localizing various neurological conditions and abnormalities in the body.Some of the main applications of EEG signaling include:

• EEG is used to monitor alertness, coma, brain death, stroke, and tumors, providing critical information about the functional state of a patient's brain;

• Cognitive engagement can be monitored by analyzing alpha rhythms to help assess a person's attention and mental focus;

• In anesthesia, EEG helps to control the depth of anesthesia and ensure optimal levels during medical procedures;

• It plays a vital role in investigating epilepsy and localizing the source of seizures, helping to diagnose and manage epilepsy;

• EEG can be used to test the effects of epilepsy medications and assess their impact on brain activity and seizure control;

• In assisting experimental cortical resection of epileptic foci, EEG helps to identify specific brain regions that cause seizures, guiding surgical intervention;

• EEG is valuable in monitoring brain development, especially in infants and children, to help understand neurodevelopmental maturity;

• EEG is used for the convulsive effects of drugs to help assess the potential effects of drugs on brain activation;

• EEG is an invaluable tool for studying sleep disorders, researching mental disorders, and understanding sleep physiology to shed light on a variety of sleep-related disorders;

• EEG can be integrated into hybrid data recording systems and combined with other imaging techniques to provide a comprehensive understanding of brain function in a variety of situations;

Data Acquisition Research

The data collection process can be categorized into two types: publicly accessible datasets and locally collected data.

Public datasets: Public datasets are collections of EEG signals that are accessible to researchers for research purposes. These datasets are freely available to the public and serve as a valuable resource for a variety of research applications;

Local data collection: researchers create their own EEG data independently of the public. These studies cover a variety of applications such as dementia research, depression analysis, motor imagery surveys, mood, and more.

EEG signal processing and analysis

EEG signal analysis involves four stages: preprocessing, feature extraction, post-processing and result analysis.

It can be divided into three steps: downsampling, artifact removal and feature scaling. Among them, artifact removal, mainly removes artifacts due to experimental settings, environmental noise, physiological signals, etc. As for feature scaling, one of the most commonly used methods is normalization, the objective function of some machine learning algorithms will not work without normalization because the range of values in the raw data fluctuates greatly.

Feature extraction involves extracting features from the primary signal to obtain a consistent classification, which is useful for classifying neurological disorders and other monitoring applications that use EEG signals. The most common methods such as Wavelet Transform (WT), Power Spectral Density (PSD), Statistical, Short Time Fourier Transform (STFT), Wavelet Entropy (WE), Differential Entropy (DE), Empirical Modal Decomposition, etc. The following table gives the different feature extraction techniques for EEG

Time domain analysis of EEG signals involves examining changes in signal characteristics over time. This includes understanding the amplitude, duration, and shape of various EEG waveforms.

• Amplitude analysis: measures the strength of the EEG signal. High amplitude may indicate strong neural activity, while low amplitude may indicate that the brain is in a relaxed or inactive state;

• Temporal patterns: include recurring patterns in the EEG signal, which may indicate regular brain activity or abnormal patterns such as seizures;

• Statistical measures: variability, mean, standard deviation, skewness and kurtosis are statistical measures applied to EEG to understand its distribution and characteristics.

Aims to reveal potential frequency components in EEG signals. Different brain activities are expressed in different frequency bands.

• Power Spectral Density (PSD): describes how power is distributed over various frequencies, which is calculated using techniques such as the Fast Fourier Transform (FFT) and provides insight into the primary frequency band;

• Band Power: Analyzing power within the EEG band can provide specific information about the state of the brain.

The EEG signal is dynamic, showing changes in frequency over time, highlighting the importance of time-frequency analysis.

• Short Time Fourier Transform (STFT): the STFT allows us to observe changes in the frequency content of an EEG signal over a short period of time. By using windows of different sizes, it provides time localized frequency analysis, making it suitable for non-stationary signals;

• Continuous Wavelet Transform (CWT): analyzes signals on different scales and provides high resolution in the time and frequency domains. It is particularly useful for capturing transient events and frequency changes over different time intervals;

• Spectrogram: illustrates how the spectrum of a signal changes over time; darker areas on a 3D spectrogram indicate higher power or amplitude in a specific frequency range at a specific time.

Feature selection and dimensionality reduction techniques are types of post-processing. To construct a model, a subset of relevant features must be selected using feature selection techniques. Feature selection eliminates irrelevant variables or noise, reduces computational complexity, reduces overfitting, and improves the generalization of the model. Dimensionality reduction (sometimes called feature extraction) combines features to extract a new set of features. Reducing the dimensionality or features of the EEG signal is required to interpret the data and obtain more accurate results as well as to improve classification accuracy. Using feature extraction, high-dimensional data is converted to lower dimensions.

EEG signal analysis can be performed using various machine learning algorithms such as supervised, unsupervised, deep learning neural networks and GSP (Graph Signal Processing) techniques. These techniques combined with EEG signals can diagnose neurological disorders such as seizures. Monitoring other applications such as mood monitoring, sleep stage classification etc.

EEG signal classification methods

The convergence of deep learning and machine learning has significantly improved our understanding of brain activity, especially in the classification of EEG signals. Informative features such as time domain statistics, power spectral density and spectrograms are extracted and classified using machine learning models such as support vector machines and random forests.

In addition, the introduction of deep learning models, especially CNN and LSTM, revolutionized EEG signal classification.CNN is good at automatically learning hierarchical spatial patterns in EEG data and utilizing multiple convolutions for feature extraction and recognition.RNN is good at capturing temporal dependencies in EEG sequences, making it suitable for temporal dynamics which is only an important task. The integration of machine learning and deep learning greatly improves EEG signal classification. By combining preprocessing, feature extraction and classification phases with traditional algorithms and deep learning models, accuracy is improved for better understanding of complex brain activities.The most popular classification algorithms in EEG signal analysis include KNN, SVM, naive Bayes(NB), logistic regression(LR), gradient boosting (GB), LDA, RF, and DT.

Research gaps and future directions

A fundamental problem is how to effectively represent, process, analyze, and visualize large-scale structured data, especially from web and graphically complex domains, which is one of the major challenges facing existing machine learning techniques.GSP, a developing branch of signal processing models and methods aimed at estimating data based on graphs, provides a new line of research on graphs to address this problem.

Challenges such as noise, artifacts, irregular outcome data, large datasets, limited spatial resolution, and individual variability require innovative approaches.The gaps in EEG signal processing arise from inherent limitations in the data, such as noise and resolution constraints, as well as the complexity and variability of the human brain. Addressing these gaps requires innovative approaches, including advanced signal processing techniques such as GSP, to improve the reliability and utility of EEG-based research and clinical applications.

Image signal processing of EEG signals

Graphical Signal Processing (GSP) is widely used for brain image processing because of its effectiveness in analyzing irregular data and graph networks. In neuroscience, complex brain activities are associated with intricate networks of functional and structural connections. Traditional deep learning methods have limitations in dealing with irregular data, and trying to use CNNs on graph structures or irregular data would be beyond the scope of CNN analysis. Considering the advantages of GSP, a classifier GCNN (Graph Convolutional Neural Network) was developed. The following table lists some graph neural network classifiers based on graph signal processing.

Sharma Ramnivas,Meena Hemant Kumar. Emerging Trends in EEG Signal Processing: a Systematic Review[J]. SN Computer Science,2024,5(4).