eeg-microstate-variational-embedding - SKILL.md Agent Skill

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:

GFP (Global Field Power) peak extraction — loses temporal information
K-means clustering — no uncertainty quantification, sensitive to initialization
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

Preprocessing: Standard EEG preprocessing pipeline
Segment encoding: Variational encoder maps EEG segments to latent space
Clustering: Microstates identified in latent representation
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