Test-Time Adaptation for EEG Foundation Models: A Systematic Study under Real-World Distribution Shifts

NeuroAdapt-Bench evaluates test-time adaptation for EEG foundation models across tasks, datasets, and pretrained models.

The study finds that standard gradient-based adaptation often fails or even hurts performance under distribution shift.

Optimization-free methods are generally more robust and reliable than optimization-based approaches.

The results suggest current adaptation methods are not yet dependable for clinical and other real-world EEG deployments.

The work highlights the need for EEG-specific test-time adaptation techniques.