The 95 percent gain interval for 12-trial calibration of memory load decoding included zero in 13 SEEG patients under independent, exchangeable patient sampling

We tested whether subtraction of a mean from 12 early trials reduces later memory load decoding loss. Public SEEG from OpenNeuro ds004752 version 1.0.1 supplied 1,280 test trials in 60 sessions from 13 patients. Calibration used no load or response labels to choose its rows or calculate its means. Each test session was excluded from supervised training. Fixed logistic models used six hippocampal power bands and five history features, with penalty 10. The equal-patient mean gain was 0.006771 nats per trial. The 95 percent patient bootstrap interval was -0.011683 to 0.025154. This interval assumes independent, exchangeable patient sampling. The exact two-sided sign-flip p value was 0.484375 under independent symmetric patient effects. The primary test did not detect an improvement. Both neural models had higher mean loss than the training-load-frequency baseline. The result permits a mean cost or benefit and does not prove a zero effect. The package preserves data, code, source hashes, all model outputs, and reproduction steps. Fresh downloads at all three native sampling rates reproduced the checked source arrays.

stereoelectroencephalography · SEEG · memory load · working memory · calibration · session transfer · logistic regression · patient bootstrap · hippocampus · OpenNeuro