Reproduce “EEG-PRISM: Physiologically-Grounded Interpretability of Predictions by EEG Foundation Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Deeksha M Shama, Punnisa Amornsirikul, Archana Venkataraman
EEG-PRISM, developed by researchers at Johns Hopkins and Boston University, provides a universal post-hoc method to interpret EEG foundation models by translating their raw time-channel input attributions into physiologically relevant spectral and spatial domains. The framework accurately recovers ground-truth frequency bands and localizes sources in simulated data, while also aligning with clinical insights for epilepsy and autism spectrum disorder on real-world datasets.
Key insights
- Attribution scores can be accurately and robustly translated from the raw time-channel input space to physiologically meaningful spectral and spatial domains using linear transformations and established backpropagation principles.
- Mapping attributions to spectral and spatial domains results in a significantly higher concentration of predictive information compared to the diffuse nature of information in the raw time-channel input space.
- EEG-PRISM enables physiologically grounded interpretations of model predictions, linking them to established neuroscientific principles regarding frequency band activities and localized brain regions, thereby bridging the gap between AI output and clinical intuition.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.