bag-of-waves-eeg-biomarkers

name	bag-of-waves-eeg-biomarkers
description	Interpretable EEG biomarkers with bag-of-waves: Spatial and temporal waveform dictionaries for low-data regimes. Use when analyzing EEG data in low-data scenarios, needing interpretable biomarkers, or working with clinical EEG classification. Activation: bag-of-waves, EEG biomarkers, interpretable EEG, waveform dictionaries, low-data EEG
metadata	{"arxiv_id":"2607.22508","published":"2026-07-24","authors":"Athanasios Papastathopoulos-Katsaros, Steven T. Lee, Lin Yao, Ajay Thomas, Junseok Park, Matthew J. McGinley, Zhandong Liu","tags":["eeg","biomarkers","interpretable","bag-of-waves","low-data","neuroscience"]}
license	Complete terms in LICENSE.txt

Bag-of-Waves EEG Biomarkers

Overview

The bag-of-waves framework provides an interpretable approach to EEG analysis that learns a small dictionary of recurring EEG waveform templates (called "atoms") using shift-invariant k-means without labels. This method operates effectively in low-data regimes where deep neural networks and foundation models are poorly suited.

Core Methodology

1. Atom Learning

• Uses shift-invariant k-means clustering to learn waveform templates from continuous EEG data
• No labels required during the atom learning phase
• Each atom corresponds to an inspectable waveform that can be validated by neurophysiologists

2. Tokenization

• Continuous EEG is converted into a sequence of atom tokens
• Token counts feed simple downstream classifiers or clustering steps

3. Temporal Extensions

• Adds atom-to-atom transitions (n-grams) to capture temporal structure
• Enables modeling of sequential patterns in EEG data

4. Spatial Extensions

• Extends from single-channel atoms to regional and cross-channel spatial atoms
• Handles multichannel EEG data effectively

Applications

The method has been tested on three complementary datasets:

1. Single-channel mouse genotype clustering (16 animals) - low-data and temporal case
2. Resting-state dementia classification - spatial case
3. TUEV benchmark - six-way classification of clinical EEG events (high-data comparison)

Advantages

• Competitive performance: Achieves results comparable to state-of-the-art deep and foundation models
• Low parameter count: Operates with a fraction of the parameters of heavy models
• Full interpretability: Every atom corresponds to an inspectable waveform
• Clinical validation: Explicitly recovers known clinical morphologies
• Low-data capability: Works effectively where heavier models fail

Implementation Guidelines

When to Use

• Low-data EEG analysis scenarios (<100 samples)
• Clinical settings requiring interpretable results
• Resource-constrained environments
• When predefined spectral features are insufficient

Workflow Steps

1. Preprocess EEG data (filtering, artifact removal as needed)
2. Apply shift-invariant k-means to learn waveform atoms
3. Tokenize continuous EEG into atom sequences
4. Extract features (counts, n-grams, spatial patterns)
5. Apply simple classifier or clustering algorithm
6. Validate atoms with domain experts

Pitfalls to Avoid

• Overfitting atom dictionary size to small datasets
• Ignoring temporal dependencies in high-frequency applications
• Failing to validate clinical relevance of learned atoms
• Not comparing against baseline spectral features

References

• Original paper: Interpretable EEG biomarkers with bag-of-waves
• Related work: EEG foundation models, interpretable machine learning for neuroscience
• Clinical validation: Consult with neurophysiologists for atom interpretation

Activation Keywords

• bag-of-waves
• EEG biomarkers
• interpretable EEG
• waveform dictionaries
• low-data EEG
• shift-invariant k-means
• EEG atoms
• clinical EEG classification