Task-free SEEG connectivity did not improve held-out CCEP N1 prediction in 8 patients

Question: can task-free SEEG recordings tell which contacts respond to single-pulse stimulation of a given pair? We used the 22,328 (stimulation pair, recording contact) rows and expert-checked N1 labels of 8 SEEG patients from sci 10ad9c19 (OpenNeuro ds005448 v1.0.0). We downloaded 10 s of signal before stimulation for each patient with HTTP byte-range requests (2 patients: 5 inter-pulse windows of 2 s). After common average reference, we calculated Pearson correlation of 1-30 Hz signals (FC_lf) and of the 70-150 Hz envelope (FC_hg). We fixed the methods before we read the signals. Result: alone, FC_lf predicts N1 with median AUC 0.75 (7/8 patients > 0.5) and FC_hg with median AUC 0.62 (8/8 > 0.5). Distance alone gives median AUC 0.894. FC_lf decreases with distance (median rho -0.43). At fixed distance, the FC_lf coefficient is positive in 6/8 patients (sign-flip p = 0.43; within-bin permutation p < 0.05 in 4/8). FC_hg is positive in 4/8 patients (p = 0.48). For held-out patients, adding FC to log distance changes AUC by a median of +0.0004 (FC_lf) and -0.0003 (FC_hg); all changes are less than 0.006. Adding both FC metrics to distance + same shaft + DTI SC gives a median change of -0.0004. Thus, for this data length and these 2 metrics, a task-free 'read' map does not add to distance for the prediction of a stimulation 'write' map. Limits: short recordings, CAR reference only, 1 center, 8 patients, patient state not known. The package has code (numpy), byte ranges and SHA-256 of all raw windows, FC matrices, per-patient tables and a figure.

SEEG · functional connectivity · CCEP · N1 response · single-pulse electrical stimulation · epilepsy · leave-one-patient-out validation