In two SEEG patients, 10–13 response-selected sites carried about 2 bits per pulse, more than all sites in 50/50 offline test splits

We tested whether a smaller set of stimulation sites can carry more information per single SEEG pulse than the full set, when the letter is read from responses at unstimulated recording contacts. Using archived single-trial features from two patients (39 and 43 sites; 20 and 6 recording contacts; 8–23 trials per site), we used nearest-centroid decoding and 50 stratified half-splits per patient, with selection based only on training data. The training-selected alphabet sizes had medians of 10 and 13 sites. On held-out test halves, these alphabets carried 2.00 and 1.95 Wolpaw bits per pulse (accuracy 0.82 and 0.73), compared with 0.95 and 1.22 bits using all sites (accuracy 0.35 and 0.40). The median differences were +1.01 [5%, 95%: 0.65, 1.47] and +0.68 [0.34, 1.26] bits, positive in 50/50 splits for each patient; the pre-registered test passed in both. Response-greedy alphabets generally exceeded random alphabets at equal sizes, while alphabets spread by contact coordinates were not consistently better than random. Results are limited to two patients and one session each; splits reuse trials, and consecutive stimulation blocks may confound responses through changes in brain state or recording. Recorded responses are only a proxy for what the brain receives. This offline analysis does not establish what patients perceive or what downstream neurons decode.

SEEG · brain-computer interface · single-pulse stimulation · information per pulse · nearest-centroid decoding · stimulation-site selection · intracranial electrophysiology · response decoding