EEG foundation models vary in how they use brain signal details
Probing Spatial and Spectral Grounding in EEG Foundation Models
Neural and Evolutionary Computing
Summary
Brain signal models called EEG foundation models are usually checked by how well they guess tasks, but this doesn't show if they use real brain signal features. The authors created tests that change the model inputs in specific ways to see if these models pay attention to where electrodes are placed or to types of brain wave shapes. By testing five popular models on two different brain signal tasks, they found that some models rely less on electrode locations or certain signal details than expected. These insights help understand what models really learn beyond just their accuracy scores.
What this means in practice
- •For biomedical engineers: Assess how EEG models process spatial and spectral brain signal features for more interpretable clinical diagnosis tools.
- •For neurotechnology developers: Design EEG-based devices with better understanding of which signal components models rely on during motor or seizure tasks.
Authors
Kleanthis Avramidis, Ivan Tashev, Dimitra Emmanouilidou
Abstract
EEG foundation models are advancing at an unprecedented pace, yet their evaluation is driven by downstream accuracy, which reveals little about whether their predictions rely on physiologically interpretable signal structure. We introduce a set of perturbation probes in two grounding axes: spatial grounding in channel identity and topographic locality, and spectral grounding in oscillatory content and waveform morphology. We apply these probes to five published EEG foundation models across motor-imagery and seizure recordings. Our experiments reveal differences in both representational sensitivity and downstream reliance. Electrode identity may be encoded without influencing downstream classification: permuting LaBraM's positional codes heavily displaced its embeddings yet reduced motor-imagery accuracy by only 1.2 points, while spectral details may be discriminable without producing specific responses to simulated clinical markers, as only REVE exceeded a spectral baseline on this task. These dissociations are hidden in common benchmarks, demonstrating how grounding probes can complement performance-based evaluation by revealing what EEG foundation models encode and what their predictions depend on.