EEG Foundation Models Fail to Capture Long-Range Temporal Correlations

New research reveals a critical blind spot in EEG foundation models: they cannot represent long-range temporal correlations (LRTC) in brain signals. Author Marzieh Zare tested five popular models (REVE, LaBraM, BENDR, CBraMod, BIOT) using the detrended fluctuation analysis (DFA) exponent from the alpha-band envelope—a key biomarker of neural dynamics. None of the models preserved temporal order, with raw-waveform models failing to recover either DFA or the 1/f aperiodic slope. More disturbingly, all embeddings were dominated by recording-site information (decodable at 0.98-1.00 accuracy), while the DFA exponent they discarded is actually site-robust (0.71 accuracy).

This spectral-temporal dissociation explains why current EEG FMs struggle with cross-population transfer. In a zero-shot task, REVE’s frozen embedding matched chance (accuracy 0.45), and even the classical DFA feature transferred only directionally without statistical significance. The findings suggest that training on short patches and pooling into fixed embeddings destroys the very temporal structure that makes EEG signals informative across individuals. For the field, this means existing foundation models need fundamentally different architectures—perhaps incorporating explicit DFA or other temporal correlation features—before they can serve as reliable biomarkers or clinical tools.