EEG-Classification for different Cognitive States

This project is to use deep learning to classify EEG Signals related to subject's state.

The goal is to use various data processing techniques, feature extraction techniqus and making deep neural network architectures to perserve both spacial and time information in the classification of EEG data.

Introduction

An electroencephalogram (EEG) is a test that detects electrical activity in your brain using small, flat metal discs (electrodes) attached to your scalp. Your brain cells communicate via electrical impulses and are active all the time, even when you're asleep. This activity shows up as wavy lines on an EEG recording.

The goal is to classify these EEG signals related to their state whether the subject is stressed by Airthmectic Operation or not.Each Subject is to carry out subtraction of two numbers in their minds and their EEG siganls are recorded.

Secondary goals include providing insight into which brain regions and frequency bands associate with each of the respective classes.There are mainly five classes Delta(1-4Hz), Theta (4-8 Hz), Alpha (8-12 Hz), Beta (12-30 Hz), and Gamma (30-100 Hz) We have to compute the Power spectrcal denisty(psd) in each frequency band.

Dataset Description

The dataset Contain edf(European data format) files of 36 subjects.Each subject had gone through two different cognitive states and each subject has 2 edf files related to rest and stress state.

The dataset Has been taken from PhysioNet

The signals are of 0.5-45 Hz, with a sampling rate 500Hz.

Approaches

1. This approach is based on deep learning Architecture of EEGnet. We have used psd as feature extraction method like we have divided the EEG epoched signal into patches of frequency bands and calculating the psd over those bands. We have also tried statistical method for feature extraction but it produced not better results than psd

Result of first Approach

This is with psd dataset Testing Accuracy-0.8221963974433469

Classification Report

Class	Precision	Recall	F1-Score	Support
0	0.82	0.97	0.89	1262
1	0.83	0.42	0.56	459

Metric	micro avg	macro avg	weighted avg	samples avg
Precision	0.82	0.82	0.82	0.82
Recall	0.82	0.69	0.82	0.82
F1-Score	0.82	0.72	0.80	0.82
Support	1721	1721	1721	1721

With referenced the result obtained from epoched array

Testing Accuracy-0.7780360255665311

Classification Report

Class	Precision	Recall	F1-Score	Support
0	0.78	0.98	0.87	1262
1	0.81	0.22	0.35	459

Metric	micro avg	macro avg	weighted avg	samples avg
Precision	0.78	0.79	0.78	0.78
Recall	0.78	0.60	0.78	0.78
F1-Score	0.78	0.61	0.73	0.78
Support	1721	1721	1721	1721

You can clearly see that how using psd as feature extraction method is much better.

2. The second approach was based on The EEG signals acquired from the dataset were augmented using a variational autoencoder (VAE).The augmented EEG signals were later used for Classifier model.There are two CLassifier models, EEGnet and other one is from research paper. The data used in paper is different so we have to upate the kernel size and layers.

Overall archiecture of Proposed model

Result of Second Approach