total_perspective_vortex

Project Description

The goal of this project is to create a ML BCI classifier able to distinguish between motor/imagery tasks as well as resting state. For the scope of this project I chose to focus on 3 classification tasks:

• Movement of left fist vs right fist
• Movement of feet vs fists
• Any movement vs resting state

The educational focus of this project is to implement a dimensionality reduction algorithm, Common Spatial Patterns (CSP) in my case. I also implemented a wrapper that performs CSP on selected set of bands, and stacks learned filters. Call it Multi-Frequency Band CSP if you like.

Learned filters are then used to extract features from the data. The features are then used to train a classifier. Several classifiers were tested, Linear Discriminant Analysis (LDA) is the default choice, at least in my implementation.

Setup

Dataset used is EEG Motor Movement/Imagery Dataset from PhysioNet (https://physionet.org/content/eegmmidb/1.0.0/). You can download the dataset with the following command:

wget -r -N -c -np -P "/DIRECTORY_OF_YOUR_CHOICE/data" https://physionet.org/files/eegmmidb/1.0.0/

The Jupyter notebooks were written to run with Python 3.12.7. Requirements can be installed using pip:

pip install -r requirements.txt

Notebooks

• explore.ipynb: Loading the data for selected subject(s), plotting the data in several introspective ways.

• train.ipynb: Training the classifier for selected experiment. The classifiers are then saved to a file in /model directory.

• train_and_evaluate_multi: Training subject-specific classifiers for multiple subjects at a time. Displays average test accuracy afterwards.

• evaluate.ipynb: Evaluating the classifier performance on selected subjects (useful if you have subject-non-specific models, I don't).

• simulate_stream.ipynb: Simulating a real-time stream of data. The classifier is used to predict the class of the data in "real-time".

Results

Depending on the subject, the results vary. Subject-specific classification accuracy varies from 50% (chance) to 90% (e.g. subjects 1 and 4). The average accuracy for the subject-specific classifiers is around 70%. The non-subject specific classifiers perform worse (barely 60% accuracy), which is the threshold for the project. Every brain is different, why bother with generalized models? Not for me.