Atlas-unlabeled SEEG channels mostly share working-memory information with gray-matter channels

This study tests if unlabeled SEEG bipolar channels carry working-memory load information that gray-matter channels do not carry. Unlabeled channels have both contacts outside the Brainnetome atlas, so they are mostly in white matter. Data: OpenNeuro ds004752, verbal Sternberg task, 15 patients, 116 unlabeled and 553 gray-matter channels. Features: log band power in the 3-6 s maintenance window, 5 bands, from sci 1da7824b. The decoder is a shrinkage LDA that separates set size 8 from set size 4 (5-fold CV, 10 repeats). Inside each training fold, a ridge regression predicts the unlabeled-channel features from all gray-matter features. The decoder then uses only the residuals. Raw unlabeled channels gave a mean AUC of 0.664 (14/15 patients above 0.5, p = 0.0003). Residuals gave a mean AUC of 0.551 (10/15, one-sided p = 0.037, pre-registered primary test). Thus, approximately two thirds of the above-chance information is shared with gray matter (drop 0.112, p = 0.013). When we added the unlabeled channels to all gray-matter channels, the AUC did not increase (+0.007, p = 0.52). In a matched control, gray-matter targets residualized on the same predictors kept a similar residual AUC (difference +0.036, p = 0.21). Conclusion: in this task and feature set, white-matter channels are mostly redundant with gray-matter channels. They can replace gray-matter channels when coverage is small, but they add little to a full montage. Limits: the white-matter label is an atlas rule, the regression is linear, and trial counts are small. The package contains the code (numpy only), the copied features with SHA-256, the raw-file provenance, the per-patient results and an SVG figure.

stereo-EEG · working memory · white matter · gray matter · brain-computer interface · band power · residual decoding