A SEEG model fitted on lists with no train gave higher mean recall loss after stimulation than a rate model in 13 people at C = 0.01

We tested whether a fixed SEEG recall model fitted on lists with no stimulation could predict recall after a train. We used public OpenNeuro ds005494 version 1.0.1 and the same cohort as two cited sci studies. All fitting used 285 pairs from 57 lists in 13 people. Testing used 269 unstimulated pairs after a train from 108 other lists. There were 62 remote bipolar SEEG channels. At C = 0.01, task features plus 4 to 8 Hz and 70 to 110 Hz power gave mean log loss 0.4188 nats. A training recall-rate model gave 0.4139 nats. The mean gain was -0.00486 nats. The 95 percent person bootstrap interval was -0.01108 to +0.00007 nats, and the exact sign-flip test gave p = 0.1462. The interval includes zero. The model also had no clear gain on excluded lists with no train. Thus, this study cannot identify stimulation as the cause of the target result. All target labels were absent from fitting. The package contains native voltage, source versions and hashes, code, all models and predictions, three figures, and replay checks. This is an offline within-session test of one small model. It does not establish useful online control or a stimulation effect.

stereoelectroencephalography · memory decoding · recall prediction · electrical stimulation · context transfer