New study: EEG predictions flip 42% with different preprocessing choices
How you prep EEG data changes results more than the model itself.
A team led by Dengzhe Hou has exposed a critical flaw in EEG-based deep learning: the preprocessing pipeline. In a paper on arXiv, they show that simply changing how raw EEG signals are preprocessed (filtering, artifact removal, etc.) flips up to 42% of trial-level predictions across six datasets spanning four cognitive paradigms. This variability is not captured by standard model uncertainty methods, which assume a fixed pipeline. The researchers formalize preprocessing as a counterfactual intervention space and provide three tools to address it.
First, they use a Walsh-Hadamard decomposition to break down the 2^7 pipeline space, revealing that sensitivity is near-additive, enabling efficient step-by-step optimization. Second, they introduce Preprocessing Uncertainty (PU), a per-trial diagnostic that complements model-based confidence scores. Third, they propose Normalized Adaptive PGI (NA-PGI), a graph-structured regularizer that leverages the compositional nature of preprocessing interventions. These tools aim to make EEG decoding more reliable for brain-computer interfaces and clinical applications.
- Up to 42% of EEG trial predictions change solely due to different preprocessing choices across 6 datasets.
- Walsh-Hadamard decomposition efficiently maps how each preprocessing step contributes to prediction instability.
- New PU diagnostic and NA-PGI regularizer help quantify and reduce preprocessing-induced uncertainty.
Why It Matters
Brain-computer interface reliability hinges on preprocessing standardization, not just model architecture.