Interpretable EEG biomarkers with bag-of-waves: Spatial and temporal waveform dictionaries for low-data regimes

Imagine you're a neurologist trying to diagnose conditions using EEG data. Traditionally, you might rely on predefined spectral features, which can be biased and limit your understanding of the data. Alternatively, deep learning models can analyze EEG signals but often require vast amounts of data and are difficult to interpret, making it hard to validate findings against clinical knowledge. This is what's called the interpretability problem in machine learning for medical applications.

To address these challenges, the authors propose a method called bag-of-waves, which learns a small set of recurring EEG waveform templates, or atoms, without needing labels. This approach transforms continuous EEG signals into a sequence of tokens based on these atoms, allowing for easier classification and clustering. By incorporating atom-to-atom transitions and expanding from single-channel to multichannel analysis, the method captures both temporal and spatial structures in the data. The results show that bag-of-waves performs competitively with state-of-the-art deep learning models while being more interpretable and requiring fewer parameters, making it particularly useful in low-data scenarios where traditional models struggle.

The method introduces a novel interpretable framework for EEG analysis.

The approach is tested across multiple datasets, showing competitive performance.