In one patient, 12 unstimulated SEEG contacts identified 37 stimulation sites at 42% top-1 accuracy

We tested whether responses at 12 unstimulated SEEG depth contacts could identify which of 37 cortical bipolar sites received an electrical pulse in one patient. In OpenNeuro ds003708 snapshot 1.0.4, a correlation-template classifier using 10–300 ms features reached 0.419 top-1 and 0.565 top-3 accuracy for one held-out pulse per probe, versus chance levels of 0.027 and 0.081. A pre-pulse control reached 0.059 top-1 and 0.164 top-3. The k=1 label-permutation null had mean accuracy 0.029 and maximum 0.072 across 100 permutations (p=0.0099, the smallest possible value with 100 permutations). Across k=1–5 averaged pulses, top-1 accuracy ranged from 0.416 to 0.489; as k increased, the true-site template had fewer pulses, so this comparison mixes probe noise and template noise. Of 430 wrong k=1 predictions, 111 selected a site sharing a contact with the true site. This is a single-patient, single-session, 6 mA result. Pulses at each site were consecutive, and pre-pulse accuracy was above chance, indicating some site information from slow drift. Features began 10 ms after stimulation, so volume-conducted artifact or its decay may contribute; neural and artifact components were not separated. The study does not establish performance in other patients or a pure pulse-count effect.

stereoelectroencephalography · electrical stimulation · evoked responses · site classification · brain-computer interface · template correlation · single-pulse stimulation