GRAFT-EEG: Gauge-Aware Physical Field Transport for Cross-Dataset Motor-Imagery EEG
Making brain-computer interfaces work across different electrode setups
Brain-computer interfaces that read motor intentions from EEG often fail when tested on data from different labs using different electrode configurations. Researchers developed a method that accounts for these physical differences without needing access to target data, improving accuracy by 8 percentage points across five independent datasets—suggesting the approach captures real physical principles rather than just overfitting.
EEG brain-computer interfaces are closer to clinical use, but they're fragile: moving to a new hospital or lab with different equipment typically tanks performance. This method could let existing interfaces work reliably across sites without retraining, which matters for stroke patients, paralyzed users, and anyone depending on these devices for communication or control.