Geometry alone selects SEEG stimulation pairs and targets better than chance, but a measured response map needs 2-3 times fewer pairs (ds005448, 8 patients)

We used the expert-checked N1 labels and the contact coordinates of 8 SEEG patients in OpenNeuro ds005448 v1.0.0 (24-48 stimulated pairs and 67-142 contacts for each patient). We did not use raw signals. All pairs together evoked an N1 at a median of 55% of the contacts (range 46-78%). To reach 80% of this ceiling, a greedy selection on the measured response map needed a median of 3 pairs. A distance model that we trained on the other 7 patients needed 7 pairs, spatially spread pairs needed 9.5 pairs and random pairs needed 10.5 pairs. At 5 pairs, geometry gave more coverage than random selection in 7 of 8 patients (exact Wilcoxon p=0.016). The measured map gave more coverage than geometry in 8 of 8 patients (p=0.0078). For one target contact, the pair with the highest geometry prediction evoked an N1 in 71% of 454 targets. One random pair evoked an N1 in 19%. The top 3 pairs evoked an N1 in 94%, and 3 random pairs in 45% (8 of 8 patients, p=0.0078). This ranking is almost the same as 'stimulate the nearest pair'. When we removed all cells at less than 15 mm, the targeting effect stayed (54% vs 13%), but the geometry advantage for coverage at small K became weak (K=3: 4 of 8 patients). Spatial spread was not better than random selection. Limits: 8 patients, one clinical protocol, N1 is not a behavioral effect, and the measured-map value is an in-sample upper bound.

SEEG · electrical stimulation · N1 response · brain-computer interface · stimulation targeting · coverage · electrode geometry