Seminar Report on An improved feature extraction algorithms of EEG signals

2nd Sem End Semester Seminar report on An improved feature extraction algorithms of EEG signals

1. An Improved Feature Extraction Algorithm of
EEG Signals A Colloquim report submitted for End Semester EvaluationNATIONAL INSTITUTE OF TECHNOLOGY
SILCHARGuided by -Dr. Koushik Guha

2. Introduction
Electroencephalogram (EEG) is a technique for recording electrical activity of the brain
using electrodes in scalp . It has become a widely used tool in neuroengineering,
cognitive neuroscience, and brain-computer interface (BCI) systems due to its high
temporal resolution (millisecond-level response).
EEG signals generally have:
Low amplitude (10–100 µV),
Frequency range: 0.5–100 Hz,
Non-stationary behavior,
Sensitivity to external and internal noise (e.g., muscle artifacts, eye movement).
Types of EEG frequency bands:
Delta (0.5–4 Hz)
Theta (4–8 Hz)
Alpha (8–12 Hz)
Beta (12–30 Hz)
Gamma (30–100 Hz)
These bands are crucial for decoding cognitive states and motor imagery in BCI
applications.
Brain-Computer Interface (BCI) aims to connect the human brain with external
systems, such as wheelchairs or robotic arms, by converting neural signals into control
commands. The key challenge is robust feature extraction from noisy EEG data.

3. Methodology
To enhance the classification accuracy of motor imagery EEG signals, we use a three-
stage pipeline:
1.Independent Component Analysis (ICA)
2.Wavelet Transform (WT)
3.Common Spatial Pattern (CSP)
2.1 Independent Component Analysis (ICA)
ICA is a blind source separation technique that separates observed signals into
statistically independent components. It is used for artifact removal, especially
eliminating EOG and ECG noise.
Mathematical Formulation:
X(t)=A⋅S(t)⇒S(t)=W⋅X(t)
Where:
X(t): Observed EEG signal matrix
A: Mixing matrix
S(t): Source signals
W=A -Unmixing matrix
−1
ICA assumes that the sources are independent and non-Gaussian, making it well-suited
for EEG signal decomposition.

Fig: The Mathemetical model of ICA .[1]

4. 2.2 Wavelet Transform (WT)
Wavelet Transform enables time-frequency analysis of non-stationary signals. It
decomposes signals into wavelet domain , maipulating the signal and transform it back.
Steps:
1.Apply discrete wavelet transform to EEG signals.
2.Threshold wavelet coefficients to remove noise.
3.Reconstruct signals using inverse WT.
High-frequency coefficients are often associated with noise, while low-frequency
coefficients preserve the core EEG signal.
2.3 Common Spatial Pattern (CSP)
CSP is a supervised feature extraction technique widely used in motor imagery tasks.
CSP extracts meaningful features from the cleaned signal, helping distinguish between
different mental states with high accuracy.
3. Proposed Workflow
The full workflow includes the following:
Step 1: Apply ICA to remove noise and artifacts.
Step 2: Perform Wavelet Transform on each component.
Step 3: Apply thresholding and inverse WT to clean the signal.
Step 4: Apply CSP to extract meaningful features.
Step 5: Use machine learning (e.g., Bagging Tree) for classification.
Fig 2: The process of the propsed method. [1]

5. 5. Results and Discussion
5.1 ICA Effectiveness
ICA successfully separates artifact-related components from neural activity, improving
signal quality and reducing false detections.
5.2 WT Improvement
WT preserves relevant neural components while discarding noise. Wavelet thresholding
allows selective removal of high-frequency noise without affecting low-frequency
ERD/ERS patterns.
4. Dataset Description
Source: BCI Competition Dataset (109 participants)
Task: Motor imagery (left hand, right hand, feet)
Electrodes: 64-channel cap (10-20 system), using BCI2000 platform
Trials: ~1500 total recordings
Each trial involves a visual cue followed by a motor imagery task. The dataset is
balanced across classes.
Fig 3: Electrodes arrangement position.[1]

6. 5.3 CSP Extraction
The CSP-filtered signal emphasizes patterns that are distinct between motor tasks by
enhancing the differences in EEG signal pattern.
5.4 Classification Accuracy
The use of a Bagging Tree classifier provides robustness through ensemble learning. It
combines multiple decision trees trained on bootstrapped datasets.Fig 4: Bagging Tree
6.Future Scope
Deep Learning Models: Explore CNNs and LSTMs for end-to-end feature
learning.
Online BCI: Adapt the method for real-time processing and control.
Multi-class Tasks: Extend to multi-class motor imagery (e.g., left, right, feet,
tongue).

7. Conclusion
The proposed method integrates ICA, Wavelet Transform, and CSP to form an
efficient EEG signal feature extraction pipeline. This hybrid approach enhances signal
quality and improves classification accuracy, making it suitable for motor imagery-
based BCI applications. Future work can explore real-time implementation and deep
learning-based classifiers to further improve performance.
References
1.[1] X. Geng, D. Li, H. Chen, P. Yu, H. Yan, M. Yue, “An improved feature
extraction algorithms of EEG signals based on motor imagery brain-computer
interface,” Alexandria Engineering Journal, vol. 61, no. 6, pp. 4807-4820, 2022.
DOI: 10.1016/j.aej.2021.10.034.
2.[2] “Feature Extraction from EEG Signals: A Deep Learning Perspective,” Int. J.
for Modern Trends in Science and Technology, vol. 9, no. 08, pp. 45–50, 2023.
DOI: 10.46501/IJMTST0908008.
3.[3] A. Mohammad, F. Siddiqui, M. A. Alam, “Feature Extraction from EEG
Signals: A deep learning perspective,” in Proc. 2021 Int. Conf. on Cloud
Computing, Data Science & Engineering (Confluence), Noida, India, 2021, pp.
757-760. DOI: 10.1109/Confluence51648.2021.9377108.
4.[4] Z. Ling, C. Shuyue, S. Yuqiang, M. Zhe, “Extraction of Evoked Related
Potentials by using the Combination of Independent Component Analysis and
Wavelet Analysis,” J. Biomed. Eng., vol. 27, no. 4, pp. 741–745, 2010.
5.[5] “International Conference on Computational Intelligence and Networks,”
IEEE, 2016, pp. 84-89.