Tissue labels predict distance-adjusted N1 reach in held-out patients; background spectra add no predictive value

We studied 355 stimulation pairs in 8 SEEG patients from OpenNeuro ds005448 v1.0.0, using expert-checked N1 labels and 1–2 mA clinical single pulses. Reach was the mean observed minus expected N1 at recording contacts at least 10 mm from the stimulated pair; expected N1 was modeled from distance, shaft, recording tissue and, where it varied, current. In leave-one-patient-out Ridge models, stimulated-pair tissue labels predicted reach (median patient Spearman ρ = 0.294; positive in 7/8 patients; exact two-sided sign-flip p = 0.016). Background spectra gave median ρ = 0.327 with a local average reference (LAP; p = 0.078), 0.158 with a common average reference (CAR; p = 0.32), and 0.218 with the recorded reference (p = 0.38). Adding background spectra to tissue labels changed mean ρ by −0.040 to −0.132 across references. The partial LAP correlation was 0.135 (6/8 positive; p = 0.20). Delta power had mean within-patient ρ = −0.25, but no single feature passed Holm correction. For raw N1 rate, models were weaker (tissue median ρ = 0.206; 5/8; p = 0.17). The results suggest tissue labels are a more stable guide to reach than background spectra, which were at most a weak label-free substitute with LAP. Limits include 8 patients from one center, low power, post-pulse rather than true-rest background windows, dataset-provided tissue labels, non-independent nearby sites, binary N1 outcomes, and dependence on the expected-N1 model.

SEEG · single-pulse electrical stimulation · N1 response · stimulation reach · tissue labels · background spectra · intracranial EEG · leave-one-patient-out prediction