Four learned coefficients gave higher site accuracy than four fixed cosine coefficients for later pulses in five SEEG sessions at 20 to 100 ms

Four coefficients from a learned temporal basis gave 19.11% mean stimulation-site accuracy. Four fixed cosine coefficients gave 17.71%. The difference was positive in all five patient sessions. Its mean was 1.40 percentage points. The exact two-sided patient sign-flip probability was 0.0625. The descriptive 95% patient bootstrap interval was 0.41 to 2.96 percentage points. These data do not establish a population effect at the 0.05 test level. The test used responses from 20 to 100 ms after a known pulse: 992 later pulses for testing and 931 earlier pulses for training, from the same stimulation-site blocks in each session. Four time averages gave 12.93% mean accuracy; all 16 original feature values gave 22.18%. Four learned coefficients reduced the output count from 16 to 4, with a loss of 3.07 percentage points from the full-feature baseline. Output count is a payload proxy at equal scalar precision; bit rate, power use, computation cost, and hardware benefit were not measured. In a 30 to 110 ms sensitivity window, two of five patient differences were negative and the exact probability was 0.6875. This is a method comparison within five sessions, not a test of transfer to another session or patient.

SEEG · electrical stimulation · temporal compression · learned basis · cosine basis · site classification · brain-computer interface · signal processing