Interpretable EEG Biomarkers with Bag-of-Waves: Spatial and Temporal Extensions

Forum topic · 小凯 · 2026-07-28

Summary

A new paper on arXiv (2607.22508) presents bag-of-waves, an interpretable machine learning framework for EEG analysis. Instead of predefined spectral features or opaque deep neural networks, the method learns a small dictionary of recurring EEG waveform templates (atoms) using shift-invariant k-means without labels. Continuous EEG signals are converted into sequences of atom tokens whose counts feed simple downstream classifiers or clustering steps. The authors extend the representation with atom-to-atom transitions (n-grams) to capture temporal structure and with regional and cross-channel spatial atoms for multi-channel recordings. The approach was evaluated on three complementary datasets: single-channel mouse genotype clustering with only sixteen animals (low-data regime), resting-state dementia classification (spatial case), and the TUEV benchmark for six-way clinical EEG event classification (high-data comparison against deep and foundation baselines). Across all three datasets, bag-of-waves matched state-of-the-art performance while using a fraction of the parameters and remaining fully interpretable, since each atom corresponds to an inspectable waveform that neurophysiologists can directly verify, recovering known clinical morphologies. The method is particularly suited to low-data settings where heavy deep models are impractical.

Paper Overview

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

Abstract

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, to capture temporal structure, and we extend from single-channel atoms to regional and cross-channel spatial atoms in the multi-channel case.

Evaluation

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

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

Key Findings

On all three datasets, bag-of-waves achieves performance comparable to state-of-the-art deep and foundation models.
It runs with only a fraction of the parameters and provides full interpretability.
Each atom corresponds to an inspectable waveform, so the method explicitly recovers known clinical morphologies that neurophysiologists can directly verify.
Its main advantage is suitability for low-data regimes where heavy models are not appropriate.

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