In SEEG load decoding, noise correlation fell with distance but distance did not predict within-patient pair gain

This offline study tested whether the distance between two stereo-electroencephalography (SEEG) bipolar channels predicts the decoding gain from adding a second channel for verbal working-memory load (set size 8 versus 4). It used archived five-band log10 Welch band-power features from 15 patients in OpenNeuro ds004752 and contact coordinates from the dataset. Features were copied unchanged from an earlier SCI contribution (4b15eb71-cde0-4d60-a836-070041a6ab44; its source is 6534004a-ec35-4e24-b982-a1156db5528a). Shrinkage LDA scores were evaluated with stratified five-fold cross-validation repeated five times. The primary analysis included 2,928 informative pairs in 14 patients; the secondary analysis included 15,965 pairs in 15 patients. Noise correlation between single-channel decoder scores decreased with distance: it was 0.38–0.40 below 10 mm and 0.08 above 40 mm, with negative within-patient correlations in 13/14 patients (exact sign-flip p=0.0002). Distance did not predict pair gain within patients (median Spearman rho −0.003; 7/14 positive; p=0.3232). Among informative pairs, pairs on different shafts gained 0.018 AUC more than pairs sharing a contact (12/13 patients; p=0.0005); this difference was absent for all pairs (p=0.7207). Noise correlation predicted gain (median rho −0.600; 14/14 negative; p=0.0001). Gains were small, and with 20–42 trials the pair decoder lost about 0.045 AUC on average relative to the better single channel. The results support measuring decoder-output noise correlation when selecting a second channel; distance alone was not a reliable within-patient proxy. Limits include offline data from one session and task, small trial counts, bipolar midpoint distances in dataset coordinates, use of the same trials for noise correlation and gain, one decoder and feature set, and confounding of pair category with distance.

stereo-electroencephalography · working memory · brain-computer interface · channel selection · noise correlation · decoding · shrinkage LDA · OpenNeuro ds004752