Interpretable EEG Biomarkers with Bag-of-Waves: Learning Waveform Atoms Instead of Deep Models

Overview

Field: Machine Learning Authors: Athanasios Papastathopoulos-Katsaros, Steven T. Lee, Lin Yao, Ajay Thomas, Junseok Park, Matthew J. McGinley, Zhandong Liu Posted: 2026-07-24 arXiv: 2607.22508

Summary

Electroencephalography (EEG) is widely used to diagnose neurological conditions, but its analysis usually relies on either predefined spectral features or deep neural networks. Predefined features carry a strong bias, since they fix in advance what counts as informative, while deep neural networks and foundation models are hard to interpret and need large amounts of data and compute.

The authors present bag-of-waves, an interpretable framework that learns a small dictionary of recurring EEG waveform templates, called atoms, using shift-invariant k-means without labels. The continuous EEG is then turned into a sequence of atom tokens, whose counts feed a simple downstream classifier or clustering step.

The representation is extended in two ways:

• Temporal structure: adding atom-to-atom transitions (n-grams).
• Spatial structure: extending from single-channel atoms to regional and cross-channel atoms in multichannel recordings.

Evaluation

The method was tested on three complementary datasets, each probing a different aspect:

• Low-data / temporal: single-channel genotype clustering in only sixteen animals (mice).
• Spatial: resting-state dementia classification.
• High-data: the TUEV benchmark, a six-way classification of clinical EEG events, compared against strong deep-learning and foundation-model baselines.

Findings

• On all three datasets, bag-of-waves achieved performance comparable to state-of-the-art deep and foundation models.
• It runs with only a fraction of the parameters of those models and offers full interpretability.
• Because each atom corresponds to an inspectable waveform, the method explicitly recovers known clinical morphologies that neurophysiologists can directly verify.
• Its main advantage is applicability to low-data regimes where heavyweight models are not suitable.

Original Abstract (excerpt)

> Electroencephalography (EEG) is widely used to diagnose neurological conditions, but its analysis usually relies on either predefined spectral features or deep neural networks. Predefined features carry a strong bias, since they fix in advance what counts as informative, while deep neural networks and foundation models are hard to interpret and need large amounts of data and compute. We present bag-of-waves, an interpretable framework that learns a small dictionary of recurring EEG waveform templates, called atoms, using shift-invariant k-means without labels. The continuous EEG is then turned into a sequence of atom tokens, whose counts feed a simple downstream classifier or clustering step. We extend this representation in two ways: we add atom-to-atom transitions, which we call n-grams...

--- *Auto-collected on 2026-07-28*