EEG Motor Movement/Imagery Signal Classification: Benchmark

This repository contains code and resources for the classification of EEG motor movement/imagery signals using various deep learning models. The project benchmarks the effectiveness and robustness of different models, specifically for the BCI Competition IV 2A dataset.

YouTube (Report/Demo)

• EEG Motor Movement/Imagery Signal Classification
  • Project Overview
  • Dataset
  • Methods and Models
    • Pre-processing
    • Data Augmentation
    • Robustness Testing
  • Results
  • Limitations and Future Work

Project Overview

This project evaluates multiple deep learning techniques to classify motor imagery tasks using EEG signals, aiming to support advancements in Brain-Computer Interface (BCI) technologies. The primary objectives include:

1. Benchmarking deep learning models for EEG signal classification.
2. Testing model robustness with noise to simulate real-world conditions.
3. Exploring the impact of data augmentation on model generalization.

Dataset

We utilize the BCI Competition IV 2A dataset, which includes EEG data from 9 subjects performing 4 different motor imagery tasks:

• Left Hand
• Right Hand
• Both Feet
• Tongue

The data was pre-processed with band-pass filtering (0.5-100 Hz) and sampled at 250 Hz.

Methods and Models

The following deep learning models were benchmarked:

• EEGNet (v1, v4): Convolutional model optimized for EEG data.
• ShallowConvNet and DeepConvNet: CNN models with varying complexities.
• EEG Conformer: Combines convolutional and attention layers.
• ATCNet: Integrates convolutional, attention, and temporal convolutional layers.
• EEG-ITNet: Utilizes inception and temporal convolution for robust performance.

Pre-processing

• Band-pass filtering (4-38 Hz)
• Exponential Moving Standardization
• Common Spatial Pattern (CSP) for feature extraction

Data Augmentation

Frequency shifting and sign flipping transformations were applied to improve model generalization.

Robustness Testing

Models were evaluated for robustness by introducing Gaussian noise at SNR levels of 8 dB and 15 dB.

Results

• Best-performing model: ATCNet with an accuracy of 67.82% and an F1 score of 0.6751.
• Top models under noise: ShallowFBCSPNet and EEG Conformer, showing minimal performance drop at lower SNR levels.
• Data augmentation improved performance for most models, particularly ATCNet.

Limitations and Future Work

• Compute Limitations: Constraints restricted testing for certain complex models like FusionNet.
• Dataset Size: Small sample size (9 subjects) limits generalizability.
• Future Directions:
  • Extend to other EEG datasets.
  • Explore advanced embedding techniques for feature extraction.
  • Test robustness with additional noise/distortion types.