Project Report on An Improved Feature Extraction Algorithms of EEG Signals

This report presents my M.Tech Final Year Mid-Semester Project on the implementation of an EEG-based Motor Imagery Classification System for Brain–Computer Interface (BCI) applications.

The project focuses on a complete signal processing and machine learning pipeline that includes:

* Data Source: PhysioNet EEG Motor Movement/Imagery Dataset (64 channels, 160 Hz sampling rate)
* Preprocessing: Bandpass filtering (8–30 Hz) and artifact removal using Independent Component Analysis (ICA)
* Denoising: Enhanced signal quality via Coiflet-5 Wavelet Transform, achieving approximately 18 dB SNR improvement
* Feature Extraction: Application of Common Spatial Pattern (CSP) to extract discriminative features between left and right hand imagery tasks
* Classification: Linear Discriminant Analysis (LDA) used as a supervised learning model for task classification
* Performance: Achieved 64.1% accuracy with a balanced precision, recall, and F1-score (around 0.63), validated through 5-fold cross-validation

This research demonstrates how integrating signal processing and machine learning can enhance EEG-based classification accuracy, paving the way for robust and interpretable motor imagery BCI systems.

Keywords: EEG, ICA, Wavelet Denoising, CSP, LDA, Brain–Computer Interface, Signal Processing, Machine Learning, PhysioNet

1. An Improved Feature Extraction Algorithms
of EEG Signals A Project report submitted for Mid Semester EvaluationNATIONAL INSTITUTE OF TECHNOLOGY
SILCHAR2nd YEAR
Guided By – Dr. Koushik Guha

2. Contents
Introduction to EEG Signal
Introduction to BCI
Methodology
Result and Discussion
Future Work
Conclusion
References
Objectives
Dataset Description

3. Introduction to EEG Signals
Electroencephalogram (EEG) is a technique to record electrical activity of the brain
using scalp electrodes. It reflects real-time brain functions and neural communication.
EEG signals are characterized by:
Low amplitude (10–100 μV)
Frequency range of 0.5–100 Hz
Non-stationary and time-varying nature
High temporal resolution (millisecond-level)
For this project, the PhysioNet EEG Motor Movement/Imagery Dataset
(EEGMMIDB) was used.
Introduction to Brain–Computer Interface (BCI)
A Brain–Computer Interface (BCI) connects the human brain directly to external
devices, converting brain signals into actionable commands. BCI enables users to
control devices through thought, and is especially useful for individuals with motor
disabilities.
However, EEG signals are weak, non-linear, and prone to contamination by noise
sources such as EOG and ECG. Hence, robust feature extraction is essential for
accurate classification.
Introduction

4. Objectives
The primary objectives of this project were:
To replicate an EEG signal processing pipeline using Python (MNE, PyWavelets,
sklearn).
To implement an ICA + Wavelet + CSP + LDA model for motor imagery
classification.
To evaluate classification accuracy for Left vs Right Hand mental imagery tasks.
The dataset used is PhysioNet EEG Motor Movement/Imagery (EEGMMIDB), which
contains data from 109 volunteers with 64 channels.
Tasks: Left vs Right hand imagery
Sampling rate: 160 Hz
Example file: S001R03.edf for initial testing
Final model: Trained on S001–S109
Dataset Description
Problem Statement
Conventional denoising methods struggle with low-amplitude signals, often failing to
separate useful components from noise. Since EEG signals and noise overlap in both
time and frequency domains, traditional methods complicate feature extraction and
reduce classification accuracy.

5. Gi= (Pi AND Gi*) OR Gi
Pi= (Pi AND Pi*) Ci=Gi Si=Pi XOR Ci* Methodology
The proposed methodology consists of four major steps:Purpose: Separate mixed brain signals into independent components.
Removes artifacts caused by eye and muscle movements (EOG, ECG).
1.Independent Component Analysis (ICA)

6. Si=Pi XOR Ci* Purpose: Denoise EEG signals.
Decompose the signal into frequency components using Wavelet Transform.
High-frequency coefficients represent noise, while low-frequency coefficients represent
the true signal.
Noise is suppressed using thresholding, followed by Inverse Wavelet Transform to
reconstruct the clean signal.
2.Wavelet Transform (WT)

7. Purpose: Feature extraction.
CSP maximizes the variance between two classes (Left vs Right hand imagery).
Generates discriminative spatial features.
3.Common Spatial Pattern (CSP)

8. Purpose: Classification.
LDA projects data onto a line that maximally separates two classes.
Used 5-fold cross-validation, achieving ≈64.1% classification accuracy.
4.Linear Discriminant Analysis (LDA)
LDA on Single RecordingL R

9. LDA on Entire Dataset RecordingL R

10. Results and Discussion
The combined ICA–Wavelet–CSP–LDA pipeline achieved:
Mean classification accuracy: 64.10%
Improved noise robustness compared to conventional methods
Effective preservation of motor-related EEG features
The results demonstrate that the hybrid feature extraction pipeline significantly improves
the classification of motor imagery tasks, which is crucial for effective BCI systems.
Future Work
Future extensions of this work may include:
1.Improving recognition accuracy through advanced feature engineering.
2.Tuning the classifier or using more powerful classifiers such as SVM or CNN.
3.Implementing CSP and LDA on FPGA for real-time processing and deployment.
Conclusion
The proposed methodology effectively combines ICA, Wavelet Transform, CSP, and LDA
to enhance feature extraction and classification of EEG signals. The approach is robust to
noise and preserves useful motor imagery features, achieving approximately 64% accuracy.
This lays the groundwork for real-time BCI systems and further research into deep
learning and hardware acceleration.

11. References
Xiaozhong Geng et al., “An improved feature extraction algorithms of EEG signals
based on motor imagery brain-computer interface,” Alexandria Engineering Journal,
61(6), 2022, pp. 4807–4820.
International Journal for Modern Trends in Science and Technology, Vol. 9, Issue 08,
2022.
A. Mohammad, F. Siddiqui and M. Afshar Alam, "Feature Extraction from EEG
Signals: A deep learning perspective," 11th International Conference on Cloud
Computing, Data Science & Engineering, Noida, India, 2021.
Z. Ling et al., “Extraction of Evoked Related Potentials by using the Combination of
Independent Component Analysis and Wavelet Analysis,” J. Biomed. Eng., 27(4),
2010.
International Conference on Computational Intelligence and Networks. IEEE, 2016.