Researcher(s)
- Eric Mans, Electrical Engineering, Saint Louis University
Faculty Mentor(s)
- Austin Brockmeier, Electrical & Computer Engineering, University of Delaware
Abstract
Electroencephalography (EEG) is a non-invasive way to record the electrical activity of the brain through electrodes placed on the scalp. These readings can then be analyzed for clinical purposes, neuroscientific investigations, and to enable brain-computer interfaces. Due to electrical conduction EEG time series readings are a combination of a multitude of sources at the same time. To separate out relevant sources from other sources, like artifacts not associated with brain waves, independent component analysis (ICA) is applied. Through ICA, the original EEG signals are described by a set of independent component signals (ICs). Each IC is characterized both spatially in terms of a topographic map (topo-map) of the signal across the electrodes on the scalp, and temporally by its autocorrelation function and power spectral density (PSD). In this work we seek to interpret how much of the spatial information is contained within the temporal (PSD/autocorrelation) features. To do this, we trained machine learning models to process both the spatial (topo-map) and temporal features. We first employed contrastive learning techniques to increase the similarity of vector-space embeddings of spatial and temporal features produced by artificial neural networks. We then used the embedding of the temporal features to predict the brain region of the equivalent current dipole fit as estimated from the topo-map by applying a classifier to the embedding vector. This classification divided the brain into 7 regions based on k-means of all the dipoles in the training set. After training, the classifier was correct 39% of the time, doubling the 17.8% change accuracy. This model lays the basis for more study, including seeing if there is more spatial information in other temporal features from the IC as compared to PSD and autocorrelation.