Analysis of Different Algorithms for EEG Signal Feature Extraction

in BCI

Xinyi Jiang

School of Computer Science and Engineering, University of New South Wales, Sydney NSW, 2052, Australia

Keywords: Electroencephalography (EEG), Feature Extraction, Bidirectional Brain‑Computer Interface (BBCI).

Abstract: Brain-Computer interface (BCI) technology has made important breakthroughs in neuroscience and human-

computer interaction in recent years, allowing the brain to communicate directly with external devices. In

recent years, advances in feature extraction algorithms, signal processing methods, and deep learning models

have greatly improved the effectiveness of BCI in medical rehabilitation, cognitive enhancement, and

neuroprosthetics. However, bidirectional BCI (BBCI) is still in its infancy and research content is limited,

which limits its application in sports rehabilitation and cognitive intervention. In this paper, the algorithms

commonly used to extract EEG signal features in the field of BCI are discussed, and combined with the

experiments of several researchers, the key algorithms in time-frequency analysis, deep learning, and spatial

feature extraction are analysed, and their effects on BCI performance are analysed. The results show that

Short Time Fourier Transform (STFT), Tunable Q-Factor Wavelet Transform (TQWT), Long-Short-Term

Memory (LSTM), BiLSTM and Filter Bank Common Spatial Pattern (FBCSP) have significant accuracy

advantages. This paper also expected that BBCI would have promising applications in the fields of neural

rehabilitation, cognitive enhancement. Future research should focus on solving the individual differences of

EEG signals, optimization of denoising technology and real-time computing efficiency, to further improve

the practicability of BBCI. At the same time, the data privacy and neurosecurity of brain-computer interfaces

also need to receive more attention to ensure the safety and ethical compliance of BBCI technology.

1 INTRODUCTION

As artificial intelligence technologies become more

and more advanced and people insisting exploration

the brain of their body, in recent years, Brain-

Computer Interface (BCI) technology based on neural

technology and computer science, which directly

connects human brainwaves with computer systems,

has become a new promising research direction.

The technology of BCI is used in neural

experiments initially, aiming to enable patients who

are suffering from paralysis causing difficulty in

communicating with others to have communications

with others again by using a device that has some

cursors being put on relevant parts of their head,

reading active electronic signals in their brain,

decoding those data and express these patients’

thoughts. Gradually, this technique is used in

pragmatic applications, including medical,

entertainment fields. As a subpart of BCI,

Bidirectional BCI is also attracting a glut of research

attention, which could execute more complex

interactions and more accurate control, not only

reading brainwaves but also giving responses to

people’s brains.

There are several research directions of BCI. One

of those directions is collecting signals and improving

the technique of resolving. Conventional signal

collection method of electroencephalography (EEG)

was used widely because of the convenient and non-

invasive quality. But the spatial resolution was

relatively low and signal-to-noise ratio was limited,

so, researchers introduced advanced signal

processing algorithms to raise the accuracy and

efficiency of extracting features of brain signals,

including Filter Bank Common Spatial Pattern

(FBCSP) and deep learning models (Lin et al., 2023).

Another direction is applications on neural

rehabilitation, studies have shown that BCI-based

neurological recovery equipment and electrical

stimulation can apparently improve the rehabilitation

effect for patients of stroke, Parkinson's disease and

other diseases. For instance, if the patients of stroke

use the unilateral lower-limb exoskeleton robot to

Jiang, X.

Analysis of Different Algorithms for EEG Signal Feature Extraction in BCI.

DOI: 10.5220/0014399200004933

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Biomedical Engineering and Food Science (BEFS 2025), pages 58-63

ISBN: 978-989-758-789-4

train the ability to walk and use Neural Muscle

Electronic Stimulation (NMES) to help them, their

balance of walking would be improved. This research

revealed that the process of re-moulding neural

systems through electronic stimulations could

effectively help patients’ neural functions recover

(Huo et al., 2024).

Apart from more advanced algorithms and medical

applications, there are also some applications of

Bidirectional BCI. Unlike traditional one-way BCI,

Bidirectional BCI could not only read signals in

people’s brain, but also send signals to those subjects’

brain by electronic stimulations and other methods,

creating more interactions between people and

devices (Lee et al., 2021).

This essay will analyse research on BCI and give

critical thinking about the use of different advanced

technologies in different areas, including helping to

recover from several kinds of diseases or disability,

and the current situation of wearable devices, and

identification of medical images. Simultaneously,

this essay will analyse the theory and the promise of

Bidirectional BCI, combined with the current

development of this technique.

2 IMPORTANT RESEARCH

PART OF BCI

The important part of techniques used for BCI is how

to extract characteristics of EEG signals. The process

involves putting electrodes on subjects’ head, and

those electrodes would detect the electronic activities

of neuron system, and devices would be used for

collecting brainwave data, solving them and

extracting useful information. However, EEG signals

are relatively easily affected by noises, which would

have deleterious effects on the accuracy of

interactions between human and devices. Therefore,

there is a need to use advanced algorithms to solve

those collected data and improve the accuracy of

extracting data and make results reliable.

2.1 Time Frequency Algorithm of BCI

In the field of BCI, EEG signals are unstable signals,

and there are time frequency algorithms which could

overcome the unstable quality of the signals and

describe the relationship between signals and time

intervals. Short Time Fourier Transform (STFT),

Tunable Q-Factor Wavelet Transform (TQWT) are

typical time frequency algorithms. This kind of

algorithm could provide information of the frequency

distribution on different time intervals of EEG signals,

especially TQWT, which could not only provide high

frequency resolution in low frequency band, collecting

signals which change slowly and have various details

including recognition situation, but also high time

resolution in high frequency band, collecting signals

which change fast including muscle activities.

2.1.1 STFT

STFT can divide signals into many time intervals and

apply Fourier Transform on each interval. In the

formula of this algorithm, s(t) represents signals

which should be solved and h(t) means window

function which usually regard 0 as the centre. The

formula indicates that by changing the time window

function h(t), the frequency distribution on various

time intervals could be determined. In addition to

that, STFT is suitable for analysing EEG signals

(Zhang et al., 2022).

𝑆𝑇𝐹𝑇

(

𝑡, 𝑓

)





𝑠

(

𝜏

)

ℎ(𝜏 − 𝑡)



𝑒



𝑑𝜏





(1)

In fields of processing voice signals and

recognizing emotion, STFT could effectively transfer

voice signals to the representation of time frequency.

For instance, there is research that used techniques

based on STFT, named Mel Spectrogram with Short-

Time Fourier Transform (Mel-STFT). The method

extracted features about voice by using the formula of

STFT, transferring signal amplitude distribution into

the Mel scale and researchers also introduced

Improved Multiscale Vision Transformers (MViTv2)

as their classifier (Ong et al., 2023). Compared with

the previous versions of classifiers, this classifier

enhanced the ability of space-time interaction

modelling and reduced the loss of information during

pooling. When the team of these researchers was

testing the method, they used some speech emotion

datasets, and all of them covered various kinds of

emotions and over 1000 audio samples, the result

showed that the accuracy datasets applied with Mel-

STFT had the highest accuracy among all datasets,

around 90.57%, 81.75% and 63.49%, indicating that

the method using techniques based on STFT is an

effective approach which could be used for extracting

and analysing features of information of human-

computer interaction.

Analysis of Different Algorithms for EEG Signal Feature Extraction in BCI

2.1.2 Limitations of STFT and Mel-STFT

While the method used by the researchers obtained an

outstanding effect, there are still some limitations of

this kind of approach. The limitation of STFT is that

if the window function was fixed, then the

adaptability to different signal bands would not meet

the expectation. Different parts of the voice signal

may need different relevant shapes and lengths of the

window functions. As for Mel-STFT, the

characteristics would be influenced by background

noises and neglect some information about the high

frequency. Using Wavelet Transform could help

solve the problem because it can retain more

information of high frequency and improve the ability

to solve unstable signals.

2.1.3 TQWT

TQWT could decompose complex unstable signals

into multiple scale sub-bands, and the Q-factor means

the quality factor which could be adjusted, with

different values of Q representing different types of

signals. Zhang, et al. used the datasets of epileptic

patients and other healthy subjects, applying TQWT

to their EEG signals and extracting features of sub-

bands on different frequencies (Zhang et al., 2022).

They adjusted Q-factor to adapt to the characteristics

of signals, and decomposed signals from high

frequency to low frequency to capture some

information of a certain frequency band when

epilepsy started seizure. In the experiment, the

combination of TQWT and deep learning, TQWT

with deep residual shrinkage network (TQWT-

DRSN), had better performance than only time-

frequency TQWT methods, with the accuracy of

99.92%, 95.20% and 90.46% on different

classifications. The result shows that the combination

could be used in multiple fields such as detection for

epilepsy and Parkinson, because of its effective

extraction from useful frequency bands and automatic

extraction from higher level through deep learning.

2.1.4 Limitations of TQWT

In the experiment of the research, as the difficulty of

classification becomes higher, the accuracy becomes

lower, and under this situation, TQWT should be

combined with the deep learning algorithm to raise

the accuracy. Moreover, TQWT is sensitive to the

noise in signals, which should be solved by

preprocessing to reduce noise.

2.2 Deep Learning Algorithms Applied

to EEG Signals

In deep learning algorithms, there are some typical

algorithms such as Long-Short-Term Memory

(LSTM). Unlike traditional algorithms, deep learning

algorithms could automatically learn features from

raw data of EEG signals, reducing the workload of the

extraction. Besides, in the classification of EEG

signals, this algorithm has higher accuracy and has

the ability to recognise patterns, which could be

applied to the domain of emotion recognition,

sleeping-level classification.

2.2.1 LSTM

LSTM is a type of deep learning algorithm that is used

to process constant time series signals and could

automatically extract features of EEG signals about

reliability in terms of time, reflecting the continuous

characteristics of human brain activities. This

algorithm is suitable for analysing long-time signal

patterns.

In research of automatically detecting EEG signals

in Parkinson's disease (Göker et al., 2023), an

extension version of LSTM, BiLSTM was used for

extracting features of time series in EEG signals and

making the classification for those features. BiLSTM

can calculate forward and backward simultaneously,

which is more effective than the normal kind LSTM.

This research used datasets from Parkinson's patients

and other healthy subjects and calculated the power

spectral density (PSD) of EEG signals between 1Hz

to 49Hz and used 4 classification algorithms to

extract features of PSD. Compared to other

algorithms, BiLSTM algorithm had the highest

performance in their experiments when it is combined

with Welch method, with the sensitivity of 99.4%, a

specificity of 96.5%, a precision of 96.4% and an

accuracy of 97.92%, which means that algorithms

based on deep learning could have an extraordinary

performance on extracting features of EEG signals

and offer an important and significant access to

diagnose neural diseases.

2.2.2 Limitations of LSTM

Although the performance of results of BiLSTM is

more accurate and precise than others, the cost of

calculation would be relatively higher than other

algorithms because of simultaneously processing

forward and backward data. And similar to LSTM,

they both have complex network structures, including

BEFS 2025 - International Conference on Biomedical Engineering and Food Science

a quantity of variables. Moreover, if there are noises

or some artifacts including eye tracking and

electromyographic disruptions, the performance of

these models would degrade. Thirdly, overfitting

should not be neglected, because sometimes the

model is too complex, causing those unnormal values

also fit. To address the potential problem, researchers

could introduce some lightweight models to reduce

the complexity of calculations and enhance

preprocessing techniques to avoid interruptions of

irrelevant information, and they can also use the

technology of Dropout in the LSTM model to turn off

some nodes, reducing the overfitting problem.

2.3 Spatial Feature Extraction from

EEG Signals

EEG signals are distributed in various parts of human

brain, so the spatial distribution of those signals could

represent neural activities. In addition, spatial feature

extraction could identify coordinated activity patterns

in different brain regions, which could also extract

characteristics of those signals. This kind of

algorithm could improve the accuracy of

classification of those signals, so there is a need to

apply this type of algorithm on the domain of medical

recovering, motor imagery and emotion recognition.

2.3.1 Filter bank common spatial pattern

(FBCSP)

FBCSP is an algorithm which could analyse multiple

frequency bands, by applying Common Spatial

Pattern (CSP). CSP could improve spatial filter by

determining the ratio of maximized and minimized

standard deviation of EEG signals, which could raise

the accuracy of classifications. FBCSP would

decompose signals into multiple frequency bands,

and then it could take advantages of CSP to extract

the feature of each band. At last, the feature of highest

identity would be selected. In the domain of motor

rehabilitation, movement-related cortical potentials

usually appear at two bands, α bands (8-12 Hertz) and

β bands (13-30 Hertz), so advantages of FBCSP could

match the feature of these signals, which could help

patients train their brains when patients are imagining

exercising, typically such as raising hands.

In the research of applying 3 different algorithms

including FBCSP on 2 datasets about EEG signals in

motor imagery field (Meng et al., 2024), researchers

compared those data of accuracy and found that

FBCSP performed the best, with an accuracy at

91.57% for dataset A, and an accuracy with 83.32%

for dataset B. But due to the redundancy of results of

FBCSP, researchers applied a stepwise discriminant

analysis (SDA) on FBCSP, making the accuracy of

both datasets increase to 98.47% and 95.2%. The

result illustrated that FBCSP could perform an

outstanding accuracy than other algorithms, the

feature of motor imagery could be extracted

effectively, and if the algorithm could be combined

with SDA, the accuracy would be higher.

2.3.2 Limitations of FBCSP

FBCSP would generate a considerable number of

extracted features, among of which may include some

abundant information affecting the attributes of

classifications, such as discrete difference brought by

various subjects, irrelevant neural activities and

useless noises, so there is a need to apply some

methods to select features more effectively, and

applying SDA is an appropriate way. SDA could

remove features which could not contribute to

classifications to reduce complexity of calculations,

reduce noise to improve accuracy for different

subjects in the research.

3 SITUATIONS OF

APPLICATION OF BCI

BCI is widely used in multiple fields, including

medical care and entertainment, providing people an

easier way to observe data and ease mental stress.

In medical care domain, BCI technology could be

used in clinical situations, helping doctors know

about patients’ mental activity and behaviours,

making the diagnosis and treatment easier. Taking

neural science as an example, the technique could be

used to observe and evaluate patients’ brain activity

and behaviour recognition, providing a non-invasive

detection for EEG signals (Zhang et al., 2021). On the

other hand, BCI could be used as a rehabilitation

method, enabling patients to control their limbs and

recover their perceptual functions. Khan et, al

discussed some Motor-Imagery BCI strategies (Khan

et al., 2020), including functional electric stimulation,

robotics assisted systems and virtual reality technique

based hybrid models, which can help people suffering

from stoke or other diseases recover motor function,

enhance the effect and improve the immersive

experience. With the help of BCI, EEG signals would

be converted into mechanical command to help

people train their body.

Analysis of Different Algorithms for EEG Signal Feature Extraction in BCI

In entertainment domain, by collecting EEG

signals and analysing, a glut of BCI applications

could be created. There are some devices which could

draw people’s dream, by collecting their EEG signals

when those subjects are dreaming, and making

relevant drawings, which could not only demonstrate

the scene of dream to people, but also stimulate their

curiosity to explore more knowledge about dream.

Furthermore, BCI is also used in game developing.

Game developers could take advantages of various

algorithms to analyse features of EEG signals in

different game experiences to design a game with a

better interactivity and immersion.

4 PROBLEMS AND

EXPECTATIONS OF BCI

Although there are many fields introducing BCI to

help do research, some issues about the technique

should not be neglected. For example, because EEG

signals are data which record personal emotion and

memory, there is a need to protect personal privacy,

avoiding being attacked by hackers. Moreover,

regulations about how to use BCI should be

announced. Given the fact that BCI could control

brain activities, limiting the use of BCI should be

taken into account.

As for the future development of BCI, although

some branches of BCI research are in a nascent stage,

including applications on virtual reality, enhancing

users’ experience including communications with

analysing their EEG signals (Zhu et al., 2023) and

bidirectional BCI (BBCI), the technology has started

being used for some more realistic fields, such as

helping more people regain abilities. BBCI could not

only receive and collect signals from brain but also

transmit the response signals back to the body, which

could be used in motor rehabilitation and memory

enhancement. For instance, BBCI makes patients feel

their strength when they are touching or holding some

objects, aiming to make them realize the existence of

their assistant legs, enable them to memorize such

feeling and help them control their behaviours, and

additionally accelerate the process of recovering. On

the other hand, BBCI could also be used in mental

recovering fields, create positive dreams to provide

some virtual experience, helping post-traumatic stress

disorder (PTSD) patients release stress.

5 CONCLUSIONS

This paper focuses on BCI, explores advanced

algorithms of EEG signal feature extraction, and

combines the application of many researchers in

medical rehabilitation, emotion recognition and brain

control devices, including the three methods, which

are time-frequency analysis, spatial feature extraction

and deep learning. In the analysis, STFT, TQWT,

LSTM, FBCSP algorithms are summarized, and their

shortcomings and improvement measures are

critically considered, and the driving effects of these

algorithms in the field of neuroscience. In addition,

the potential applications of BBCI in motor

rehabilitation, neural regulation and cognitive

enhancement are also discussed. By analysing the

results of this paper, based on the summary and

analysis of existing algorithms, the concept of typical

advanced algorithms could be understood, and it

would be explicit to know when to use different

algorithms to improve the accuracy of EEG signals,

and make ideas about the field of BCI.

However, because there are too many kinds of

algorithms, this paper cannot list them all and make

analysis and evaluation. There are many other

outstanding algorithms to extract features in an

effective way. In conclusion, BBCI has high

application value in the fields of medical

rehabilitation, neural regulation and human-computer

interaction.

In the future, with the development of

neuroscience, artificial intelligence and material

engineering, the technology is expected to make

further breakthroughs to achieve human-computer

integration in a true sense, and improve human

perception, memory and movement ability.

REFERENCES

Göker, H. 2023. Automatic detection of Parkinson’s disease

from power spectral density of electroencephalography

(EEG) signals using deep learning model. Australasian

Physical & Engineering Sciences in Medicine, 46(3),

1163–1174.

Huo, C., Shao, G., Chen, T., Li, W., Wang, J., Xie, H.,

Wang, Y., Li, Z., Zheng, P., Li, L., & Li, L. 2024.

Effectiveness of unilateral lower-limb exoskeleton

robot on balance and gait recovery and neuroplasticity

in patients with subacute stroke: A randomized

controlled trial. Journal of Neuroengineering and

Rehabilitation, 21(1), 213–216.

Khan, M. A., Das, R., Iversen, H. K., & Puthusserypady, S.

2020. Review on motor imagery based BCI systems for

BEFS 2025 - International Conference on Biomedical Engineering and Food Science

upper limb post-stroke neurorehabilitation: From

designing to application. Computers in Biology and

Medicine, 123, 103843.

Lee, J., Leung, V., Lee, A.-H., Huang, J., Asbeck, P.,

Mercier, P. P., Shellhammer, S., Larson, L., Laiwalla,

F., & Nurmikko, A. 2021. Neural recording and

stimulation using wireless networks of microimplants.

Nature Electronics, 4(8), 604–614.

Lin, C.-L., & Chen, L.-T. 2023. Improvement of brain–

computer interface in motor imagery training through

the designing of a dynamic experiment and FBCSP.

Heliyon, 9(3), e13745–e13745.

Meng, Y., Su, Y., Li, D., Nan, J., & Xia, Y. 2024. A

Stepwise Discriminant Analysis and FBCSP Feature

Selection Strategy for EEG MI Recognition.

International Journal of Advanced Computer Science &

Applications, 15(5).

Ong, K. L., Lee, C. P., Lim, H. S., Lim, K. M., & Alqahtani,

A. 2023. Mel-MViTv2: Enhanced Speech Emotion

Recognition With Mel Spectrogram and Improved

Multiscale Vision Transformers. IEEE Access, 11,

108571–108579.

Zhang, R., Jia, J., & Zhang, R. 2022. EEG analysis of

Parkinson’s disease using time–frequency analysis and

deep learning. Biomedical Signal Processing and

Control, 78, 103883.

Zhang, X., Yao, L., Wang, X., Monaghan, J., McAlpine, D.,

& Zhang, Y. 2021. A survey on deep learning-based

non-invasive brain signals: Recent advances and new

frontiers. Journal of Neural Engineering, 18(3), 31002.

Zhu, H. Y., Hieu, N. Q., Hoang, D. T., Nguyen, D. N., &

Lin, C.-T. 2023. A Human-Centric Metaverse Enabled

by Brain-Computer Interface: A Survey.

Analysis of Different Algorithms for EEG Signal Feature Extraction in BCI