Adding SEEG power increased mean test loss by 0.035 nats for recent stimulation prediction on matched PAL2 lists in 13 people at C = 0.1

We tested if SEEG power helps predict recent stimulation history in word pairs with no current train. We used public OpenNeuro ds005494 version 1.0.1. The test had 13 people, 62 remote bipolar channels, 103 matched list pairs, and 257 trials per class. Both lists in each pair were excluded from training when used for testing. A fixed logistic model used task position and list order. A second model also used beta power and power from 70 to 110 Hz in the first and third seconds. At C = 0.1, adding SEEG increased mean test loss by 0.035 nats. The 95 percent participant bootstrap interval for loss reduction was -0.073 to -0.001 nats. The exact sign-flip test gave p = 0.100 under its symmetry assumption. SEEG helped in 6 of 13 people. At C = 0.01, mean loss reduction was 0.003 nats and its interval included zero. These results depend on the model and sample. They do not show that stimulation history is absent from SEEG. The package includes native epochs, source versions, matching rules, models, test predictions, figures, code, and checks. It reuses one earlier sci cohort; it is not an independent replication.

stereoelectroencephalography · brain-computer interface · stimulation history · paired-associate memory · spectral power · logistic regression · held-out lists