For random held-out N1-map entries in eight SEEG patients, site and contact degree added more AUC than low-rank terms

We fit logistic models to the stimulation-site × recording-contact N1 matrix of each of 8 SEEG patients (OpenNeuro ds005448, 22,328 entries, 3,302 expert-checked N1). We predicted random held-out entries (5 folds). Site and contact bias terms (degree) increased the median held-out AUC from 0.890 (distance and shaft) to 0.958 (8 of 8 patients, sign-flip p = 0.008). A rank-k site × contact term added a median of 0.0035 (k = 1, 8 of 8, p = 0.008) to 0.0066 (k = 4, 7 of 8, p = 0.016). The gain was 0.01 or more in only 2 patients. A parallel analysis of the raw residuals found 0-3 shared dimensions above a refit surrogate null (median 1.5). Thus, binary N1 maps have only a few weak shared patterns after distance and degree. For a writing BCI, the responsiveness of the site and of the contact gives most of the gain in matrix completion. A Pearson-residual null was not calibrated, and we report this. The package has the code, the input table with SHA-256, the results, a figure and the commands to run again.

SEEG · CCEP · N1 maps · matrix completion · low-rank models · brain-computer interface · parallel analysis