Human Motor Imagery based EEG Signals Classification for BCI Systems

📌 Project Overview

Author: Shashank Kashyap

Brain-Computer Interfaces (BCI) allow individuals suffering from paralysis to control external devices (like wheelchairs or robotic arms) using brain signals. This project focuses on Motor Imagery (MI)—the mental rehearsal of movement without physical output.

By analyzing Electroencephalography (EEG) signals from 10 subjects, this study develops a machine learning pipeline to classify whether a user is imagining Left Hand or Right Hand movement. The goal is to identify the most efficient algorithm to translate these thoughts into control commands.

---

🛠️ Methodology & Workflow

The study utilizes a 64-channel g.tec electrode system to capture raw brain signals, followed by a rigorous signal processing pipeline:

1. Data Acquisition: 64-channel EEG data collected from 10 healthy subjects (3200 total imagination trials).
2. Pre-processing: Noise reduction and signal normalization.
3. Feature Extraction: Calculation of 14 Frequency-Domain (FD) features, including:
   • Mean Frequency (MNF)
   • Peak Frequency (PKF)
   • Spectral Moments (SM1-SM5)
4. Classification: Comparison of three K-Nearest Neighbor (KNN) variants:
   • Cosine KNN
   • Coarse KNN
   • Cubic KNN (Best Performer)

Figure 1: Block diagram of the EEG classification workflow.

---

📊 Key Findings

1. Feature Performance

We evaluated 14 different frequency domain features. Peak Frequency (PKF) was identified as the most individual relevant feature, achieving the highest standalone accuracy compared to Mean Power or Median Frequency.

2. Algorithm Comparison

The Cubic KNN classifier outperformed Cosine and Coarse variants. It demonstrated robustness across 2-fold to 10-fold cross-validation methods.

Table 1: Performance metrics of the Cubic KNN classifier across different validation folds.

3. Classification Accuracy

The model achieved high precision in distinguishing between left and right-hand motor imagery.

• True Positive Rate (Class 1): 97.6%
• True Positive Rate (Class 2): 88.5%
• AUC (Area Under Curve): 0.981

Figure 2: Confusion Matrix showing high classification accuracy for Class 1 (Left Hand) and Class 2 (Right Hand).

4. ROC Analysis

The Receiver Operating Characteristic (ROC) curve confirms the model's stability, with an Area Under the Curve (AUC) of 0.981, indicating excellent separability between the two mental states.

Figure 3: ROC Curve demonstrating the classifier's high sensitivity and specificity.

---

💡 Conclusion

This project demonstrates that Frequency Domain features, specifically Peak Frequency, combined with a Cubic KNN classifier, provide a reliable method for interpreting Motor Imagery EEG signals.

This approach offers a computationally efficient solution for BCI applications, potentially enabling real-time control of assistive devices for paralyzed patients.

---

🔒 Full Report Access

This repository contains an overview of the analysis and key visualizations.

As this is a formal project report submitted for the award of Bachelor of Engineering, the full thesis document and code scripts are not publicly available in this repository.

Recruiters & Academic Collaborators:
If you wish to review the full project report, detailed feature extraction formulas, or discuss the methodology, please contact me directly:

📧 Email: kashyapshashank222@gmail.com
🔗 LinkedIn: https://www.linkedin.com/in/shashank-kashyap-805309238/

---

📚 References

• Dataset: Private dataset (NITTTR Chandigarh), 64-channel g.tec system.
• Techniques: FFT (Fast Fourier Transform), PSD (Power Spectral Density), KNN Classification.