eeg-microstate-variational-embedding

name

eeg-microstate-variational-embedding

description

Interpretable EEG microstate discovery via variational deep embedding with systematic architecture search and multi-quadrant evaluation. Uses deep variational methods for data-driven microstate identification instead of traditional k-means clustering on GFP peaks. Provides principled uncertainty quantification and scalable EEG analysis pipeline. Use when performing EEG microstate analysis, building interpretable EEG pipelines, or comparing microstate discovery methods. arXiv: 2605.10947 (cs.LG, q-bio.NC). Faremi, Visentin, Longo.

EEG Microstate Discovery via Variational Deep Embedding

Variational deep embedding replaces traditional k-means microstate clustering with data-driven, uncertainty-aware latent space learning for interpretable EEG analysis.

Metadata

• Source: arXiv:2605.10947
• Authors: Saheed Faremi, Andrea Visentin, Luca Longo
• Published: 2026-05-12
• Subjects: cs.LG, q-bio.NC

Core Problem

Traditional EEG microstate analysis relies on:

1. GFP (Global Field Power) peak extraction — loses temporal information
2. K-means clustering — no uncertainty quantification, sensitive to initialization
3. Manual selection of microstate number — subjective and arbitrary

Key Innovation

Variational Deep Embedding for Microstates:

• Deep variational autoencoder learns latent representation of EEG segments
• Microstates emerge as clusters in the learned latent space
• Systematic architecture search identifies optimal model configuration
• Multi-quadrant evaluation validates across interpretability, stability, accuracy, and scalability

Advantages Over Traditional Methods

• Continuous temporal modeling (not just GFP peaks)
• Principled uncertainty quantification via variational posterior
• End-to-end differentiable pipeline
• Automatic microstate discovery without arbitrary k selection

Technical Framework

Pipeline

1. Preprocessing: Standard EEG preprocessing pipeline
2. Segment encoding: Variational encoder maps EEG segments to latent space
3. Clustering: Microstates identified in latent representation
4. Evaluation: Multi-quadrant assessment (interpretability, stability, accuracy, scalability)

Architecture Search

• Systematic exploration of encoder/decoder architectures
• Latent dimensionality optimization
• Regularization strategy comparison

Applications

• EEG biomarker discovery
• Neurological disorder characterization
• Cognitive state monitoring
• Brain-computer interface feature extraction
• Large-scale EEG analysis pipelines

Pitfalls

• Requires larger datasets than traditional k-means
• Computational cost higher than classical methods
• Interpretability of latent dimensions needs careful validation
• Architecture search can be computationally expensive

Related Skills

• interpretable-eeg-biomarkers-parkinsons
• eeg-foundation-model-adapters
• eeg-hopfield-emotion-energy
• explainable-gnn-eeg-neurological