EEG-Based Brain Computer Interface Classification

Overview

This project implements an end-to-end Brain-Computer Interface (BCI) pipeline for EEG signal classification. The workflow combines signal preprocessing, artifact removal, spatial feature extraction, and machine learning to classify neural activity from multi-channel EEG recordings.

The objective is to predict feedback events from EEG signals while mitigating noise and physiological artifacts.

---

Pipeline

1. Data Loading

Training and testing EEG recordings are loaded from CSV files using chunked processing to efficiently handle large datasets.

2. Signal Preprocessing

Notch Filtering

• Removes power-line interference at 50 Hz.

Band-Pass Filtering

• Butterworth filter
• Frequency range: 1–40 Hz

These steps suppress noise while retaining relevant EEG frequency components.

3. Artifact Removal

Eye-movement artifacts are removed using EOG regression.

For each EEG channel:

• Linear regression is trained using EOG measurements.
• Predicted ocular artifacts are subtracted from EEG signals.
• Cleaned signals are retained for downstream analysis.

4. Dataset Balancing

Class imbalance is addressed through random resampling to ensure equal representation of feedback-event classes.

5. Feature Extraction

Common Spatial Patterns (CSP)

CSP is applied to cleaned EEG channels to learn spatial filters that maximize variance differences between classes.

Configuration:

• Channels used: First 8 EEG channels
• Components: 4 CSP features

The resulting features provide compact and discriminative representations of neural activity.

6. Classification

A Support Vector Machine (SVM) classifier is trained using:

• Linear kernel
• Balanced class weights
• Probability estimation enabled

The model is evaluated using a train-validation split and classification metrics.

7. Inference Pipeline

For unseen test recordings:

1. Apply preprocessing
2. Remove EOG artifacts
3. Extract CSP features
4. Generate prediction probabilities
5. Create submission file

Output format:

IdFeedBack,Prediction
FB00001,0.81
FB00002,0.24
...

Tech Stack

• Python
• Pandas
• NumPy
• SciPy
• MNE
• Scikit-Learn
• Joblib

Key Techniques

• EEG Signal Processing
• Brain-Computer Interfaces
• Common Spatial Patterns (CSP)
• Artifact Removal
• Support Vector Machines
• Signal Filtering
• Feature Engineering

Future Improvements

• Riemannian geometry-based classifiers
• Filter Bank CSP (FBCSP)
• Deep-learning-based EEG decoding
• Hyperparameter optimization
• Subject-independent evaluation