EEG
Project Info
What are the best ways to classify EEG data?
Research/Motivation
Our research aims to investigate whether motor imagery data can be effectively classified, enabling individuals who rely on wheelchairs to control them using their minds. We also seek to explore whether integrating common NLP techniques, such as attention mechanisms, into biomedical data classification can enhance the accuracy of predictive models. Additionally, we aim to determine if these technologies can become more accessible by developing pipelines that achieve high accuracy even with basic EEG devices. Our motivation stems from the fact that approximately 15 million people worldwide suffer from conditions, including locked-in syndrome, that entirely restrict their movement. Currently, thought-classification models predominantly utilize professional-grade brain-computer interfaces (BCIs), costing upwards of $25,000, rendering them unaffordable for many. Furthermore, these existing models typically achieve accuracy levels ranging from 70-80%. Our primary goal, therefore, is to create a model leveraging a more accessible 8-channel EEG setup that attains accuracy comparable to existing professional systems.
Background
Amyotrophic lateral sclerosis (ALS), cerebral palsy, brain stem stroke, multiple sclerosis, and spinal cord injuries are among the primary causes of locked-in syndrome. In motor imagery tasks, the alpha and beta brain wave frequency bands are the most influential, particularly when imagery involves left and right hand movements, with corresponding neural activity originating in the brain's C3 and C4 areas. To classify EEG data effectively, convolutional neural networks (CNN) and recurrent neural networks (RNN) have been proposed due to their ability to capture both spatial and temporal features inherent in biometric data. Additionally, attention mechanisms, commonly utilized in EEG signal classification tasks, enhance model performance by enabling the network to concentrate on the most relevant features and effectively capture temporal dependencies associated with specific signal categories.
Dataset/Processing
The dataset consists of EEG recordings from six subjects (two males and four females) aged between 18 and 22 years old. Participants were instructed to complete various prompts by watching a series of videos, including: physically opening and closing fists with audio and text prompts; imagining opening and closing fists guided by text-only prompts; imagining directions guided by audio and text prompts, audio-only prompts, and text-only prompts. Each prompt lasted for three seconds and was followed by a two-second rest period.
Data processing involved filtering EEG signals between 0.1 Hz and 30 Hz, replicating independent component analysis (ICA) preprocessing as proposed in existing literature, and applying an aggressive high-pass filter at 2 Hz. Outlier epochs were automatically rejected to ensure data quality. ICA artifact detection was performed using MNE algorithms to identify and remove muscle and eye blink artifacts. Due to the limitations of inexpensive EEG devices, which lack dedicated electrooculogram (EOG) channels, electrodes Fp1 and Fp2 were utilized as substitutes. Finally, data reconstruction was applied to restore signals for subsequent analysis.
Data Preparation
During data preparation, the EEG data was trimmed to include precisely 150 seconds from each trial. The start of each trial was identified by detecting characteristic spikes within the accelerometer channel. After identifying these starting points, the signals were segmented into individual components, each clearly labeled according to the corresponding subject's intended thoughts.
EEG Classification Pipeline
The EEG classification pipeline begins by segmenting EEG signals into overlapping windows, followed by normalization and feature extraction using a masked autoencoder. The autoencoder compresses input EEG data into latent representations through an encoder comprised of convolutional layers and subsequently reconstructs the original signals using a decoder trained with mean squared error (MSE) loss. Features learned by the encoder are then passed into a convolutional neural network (CNN) classifier, optimized using categorical cross-entropy loss. Model performance and generalizability across individuals are assessed through leave-one-subject-out (LOSO) cross-validation. This pipeline allows efficient extraction of EEG features and significantly enhances classification accuracy.
Our Team
Thank you to Cayden M, Professor Ali Etemad, Brandon C, and Ethan C for their advising throughout the project, as well as our EEG data subjects for offering their brainwaves! This project was possible through the support of QMIND.