Latent Prototype-Based Clustering: A Novel Exploratory Electroencephalography Analysis Approach.

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Supervised learning methods for electroencephalography (EEG) data are hindered by increasing amounts of unlabeled or mislabeled data.
• This degradation impacts the performance of brain-computer interfaces (BCIs) and neurological disease diagnosis tools.

Purpose of the Study:

• To develop a novel unsupervised exploratory analysis for EEG data.
• To address the challenges posed by incomplete or unlabeled EEG datasets.

Main Methods:

• A Generative Adversarial Network (GAN) termed W-SLOGAN was developed, extending SLOGAN.
• Clustering is performed in a low-dimensional latent space using prototypes associated with each cluster.
• A composite similarity metric and Gaussian Mixture Model (GMM) are employed for robust clustering, handling imbalanced datasets.

Main Results:

• Experiments on public EEG and intracranial EEG (iEEG) epilepsy datasets showed clustering results comparable to supervised classification.
• The approach demonstrated effectiveness in identifying epileptic subtypes and enabling multi-labeling of EEG data.

Conclusions:

• W-SLOGAN offers a promising unsupervised solution for EEG data analysis, overcoming limitations of supervised methods.
• The findings highlight the practical utility of W-SLOGAN in clinical applications like epilepsy diagnosis and data annotation.