In two SEEG patients, greedy selection reached 90% of all-contact single-pulse site information with 3-4 contacts

A brain-computer interface that writes to the brain through SEEG stimulation must verify which site it stimulated. This analysis measured how identification of a stimulated site from one 6 mA pulse changed with the number of readout contacts. Using archived OpenNeuro ds004457 features for two patients, it compared random contact subsets with greedy forward selection performed on training trials only. A Euclidean nearest-centroid decoder used 28 ten-millisecond bins from 15 to 295 ms after the pulse. For sub-1 (39 sites, 20 contacts), all contacts gave a mean 0.97 bits per pulse; greedy selection reached 90% of this value with 4 contacts, while random subsets needed 12. For sub-3 (43 sites, 6 contacts), all contacts gave 1.26 bits; greedy selection reached 90% with 3 contacts, while random subsets needed 5. One greedy contact gave 0.38 and 0.77 bits for sub-1 and sub-3, respectively, compared with 0.22 and 0.44 bits for a random single contact. Greedy subsets gave more bits than the random mean in every split for k=2-8 in sub-1 and k=1-4 in sub-3. The first selected contact varied between splits. These findings are limited to two patients, one run, offline analysis, sparse readout montages, few trials, and overlapping half splits. Stimulation of each site in one block may let slow drift add site information and raise accuracy; the split bands are not independent confidence intervals. The results do not show that adding contacts reduces performance.

SEEG · electrical stimulation · brain-computer interface · site identification · readout contacts · nearest centroid · Wolpaw bits · greedy forward selection · OpenNeuro ds004457