EEG-FM-Audit Pipeline Systematically Evaluates EEG Foundation Models
A new evaluation and analysis pipeline called EEG-FM-Audit has been developed by researchers to tackle significant shortcomings in EEG foundation model (FM) research. These issues include unclear supervised baseline tuning, unverified impacts of intricate learning paradigms, and insufficient transparency in model decision-making processes. The pipeline features three main elements: a benchmarking protocol driven by ASHA for equitable supervised baseline optimization, ablation studies at the paradigm level to evaluate the effectiveness of learning paradigms, and a neurophysiological probing (NPP) framework to determine if FMs utilize valid temporal, spatial, and spectral EEG characteristics. This pipeline was tested on four advanced EEG-FMs and five datasets, although specific findings are not provided in the abstract. The research is available on arXiv with the identifier 2605.26910.
Key facts
- EEG-FM-Audit is a new evaluation pipeline for EEG foundation models.
- It addresses three limitations: opaque baseline tuning, unverified learning paradigms, and lack of decision transparency.
- The pipeline includes ASHA-driven benchmarking, paradigm-level ablation, and neurophysiological probing.
- It was applied to four state-of-the-art EEG-FMs and five datasets.
- The research is available on arXiv (2605.26910).
Entities
Institutions
- arXiv