Get a New Perspective on EEG: Convolutional Neural Network Encoders for Parametric t-SNE

Area of Science:

• Computational neuroscience
• Machine learning for biomedical signals
• Data visualization and dimensionality reduction

Background:

• t-distributed stochastic neighbor embedding (t-SNE) is a common technique for visualizing high-dimensional data, often applied to electroencephalogram (EEG) analysis.
• Traditional EEG analysis with t-SNE requires manual feature engineering, and the optimal features for specific EEG characteristics remain unknown.
• Parametric t-SNE uses neural networks but still relies on pre-defined feature extraction.

Purpose of the Study:

• To develop a novel method using convolutional neural network (CNN) encoders inspired by t-SNE principles to eliminate the need for manual feature engineering in EEG analysis.
• To learn both high- and low-dimensional representations directly from raw EEG data.
• To evaluate the effectiveness of this new approach for visualizing and analyzing distinct EEG datasets.

Main Methods:

• Utilized the t-SNE principle to train CNN encoders capable of generating both high- and low-dimensional data representations.
• Applied the method to three distinct datasets from the Temple University EEG Corpus: wakefulness/sleep, interictal epileptiform discharges, and seizure activity.
• Evaluated performance by comparing CNN encoder representations against parametric t-SNE using short-time Fourier transform and wavelet features, with classification by support vector machines and clustering via k-means.

Main Results:

• CNN encoders successfully generated low-dimensional EEG representations that captured the inherent structure of different EEG states (wakefulness, sleep, epileptiform discharges, seizures).
• The learned representations generalized well to new, unseen EEG data.
• CNN encoders demonstrated comparable or superior performance to parametric t-SNE in separating data categories and achieved better clustering, both visually and quantitatively.

Conclusions:

• The developed CNN encoder approach effectively performs dimensionality reduction and feature learning for EEG data, removing the bottleneck of manual feature engineering.
• This method shows significant promise for creating versatile tools for exploring complex relationships within EEG signals.
• The principle offers a powerful, data-driven alternative for EEG analysis and visualization.