Increasing Spatial Accuracy of EEG with Signal Decomposition and Machine Learning

• Vance Steele, Computer Engineering, Rose-Hulman Institute of Technology

• Austin Brockmeier, Electrical and Computer Engineering, University of Delaware

Electroencephalography (EEG) is a non-invasive method of recording brain activity used in clinical diagnostics and cognitive research. It offers high temporal resolution, capturing brain rhythms and responses in tens of milliseconds, but lacks the spatial accuracy of invasive methods involving nodes on or inserted into the cortex. Analyzing EEG involves multiple steps, including signal preprocessing and machine learning, to separate different signal sources and isolate those related to specific brain areas or processes.

In the preprocessing stage, high pass filtering and frequency resampling are applied to isolate  the time series from within the brain cortex. Independent Component Analysis reduces this time series into source signals corresponding to the EEG channels. The spatial arrangement of each independent component’s weights, or intensities, across the scalp can be mapped onto the scalp, showing the source signal’s brain origin. These source signals are split into smaller segments called ‘windows’ and characterized by short-time power-spectral density or waveform clustering. Experts label the short-time power-spectral density of each independent component as random noise or a neural signal, which is then used in subsequent EEG analysis involving feature extraction and brain decoding. These methods were applied to data from the Imagined Emotion Study (Onton & Makeig, 2022).

The summer research project uses this dataset and others to enhance the spatial localization of EEG. The goal is to understand how much information about the spatial origin of independent component signals can be inferred from their temporal or spectral patterns. This involves extracting features from the independent components of multiple subjects and applying machine learning techniques to classify signal origin and extract features predictive of spatial weights. Logistic regression and contrastive learning will be used to achieve this. The aim is to automate EEG analysis, with applications in cognitive function analysis, including seizure detection.