EEG Feature Robustness Under Cross-Dataset Distribution Shift Using Classical Models

Version 2 (Revised): Updated abstract, clearer methodology, expanded per-subject analysis, statistical testing, and improved discussion of cross-dataset distribution shift.

This preprint investigates whether classical EEG feature representations differ in robustness under cross-dataset distribution shift. Using PhysioNet EEG Motor Imagery and BCI Competition IV Dataset 2a, it compares time-domain statistics, band power, and FFT magnitude features across 9 feature-model combinations under strict zero-shot transfer.

The main finding is that no classical feature representation achieves robust zero-shot cross-dataset transfer. In the informative BCI2a → PhysioNet direction, all feature sets showed similarly moderate degradation, and pairwise Wilcoxon tests found no statistically supported ranking between them. The results also show a strong subject-level bimodality in motor imagery performance, with a minority of strong responders and many near-chance subjects.

The study identifies a 9× amplitude difference between recording systems as a major barrier to transfer and shows that feature dimensionality imbalance does not explain the cross-dataset conclusions.

Code: https://github.com/baris-talar/eeg-feature-robustness