Cross-finger SEEG detection did not show 0.03-AUC non-inferiority in eight subjects using all-channel high-gamma power

We tested if a BCI movement switch that is calibrated with one finger detects movements of a different finger. Data: OpenNeuro ds004473 (8 SEEG subjects; left arrow key = index finger, right arrow key = middle finger of the same hand). We used the archived features of 07478163 (log 70-150 Hz and 13-30 Hz power, all adjacent bipolar channels, rest and move windows, 100 trials). We added finger labels from the beh.tsv files. A shrinkage LDA was trained on rest vs move windows of one finger and tested on held-out trials of the other finger. The training size was the same as for a same-finger model (33-40 trials). Primary endpoint (fixed before the run): AUC(cross) minus AUC(within), 70-150 Hz, all channels, non-inferiority margin 0.03, exact sign-flip test over subjects. Result: mean delta -0.008 (range -0.059 to +0.046). Non-inferiority was not shown (p = 0.066; 3 of 8 subjects below the margin). With 8 channels selected in training, delta was +0.001 (non-inferiority p = 0.0078). With 13-30 Hz, delta was +0.015 (p = 0.0039). Index vs middle finger decoding from 70-150 Hz move windows was weak (median AUC 0.54; 1 of 8 subjects significant). Conclusion: the movement signal was mostly the same for the 2 fingers, but the all-channel model with this small training set was too noisy to show transfer within 0.03 AUC. Limits: 2 adjacent fingers of 1 hand, offline window detection, small calibration sets. The package has the code, input features with SHA-256, S3 URLs of the label files, results tables and an SVG figure.

stereo-electroencephalography · brain-computer interface · cross-finger transfer · movement detection · shrinkage linear discriminant analysis · non-inferiority