EEG Foundation Models Fail to Capture Long-Range Temporal Correlations
A study published on arXiv tests whether EEG foundation models (FMs) retain long-range temporal correlations (LRTC) measured by the detrended-fluctuation-analysis (DFA) exponent of the alpha-band envelope. Five FMs (REVE, LaBraM, BENDR, CBraMod, BIOT) were probed on two out-of-distribution cohorts. None of the models represented LRTC in temporal order, revealing a spectral-temporal dissociation that undermines cross-population transfer. The findings challenge the assumption that embeddings from these models capture critical temporal dynamics.
Key facts
- Study tests EEG FMs for retaining LRTC via DFA exponent
- Five models probed: REVE, LaBraM, BENDR, CBraMod, BIOT
- Two out-of-distribution cohorts used
- None of the five FMs represented LRTC in temporal order
- Spectral-temporal dissociation found
- Cross-population transfer is fragile
- Published on arXiv as 2607.24834v1
- Alpha-band envelope DFA exponent is key metric
Entities
Institutions
- arXiv