Mean relative recall error increased by 0.70 percent when a fixed model added delay high gamma power correlation in six SEEG patients

We tested whether delay power timing improves recall prediction in public SEEG data. The test used six Utah patients and 1,159 matched trials from OpenNeuro ds006136 version 1.0.1. One added feature was the mean correlation of frontal and temporal high gamma log power across eight 499 ms delay segments. Each channel's mean training time profile was removed. A fixed ridge model already used regional power, power concentration, channel coverage, and trial position. With equal patient weight, mean squared error was 0.551270 for the baseline and 0.554788 with the added feature. Mean baseline minus added error was -0.003519. Its 95 percent patient bootstrap interval was -0.007375 to 0.000011, under independent patient sampling. The exact sign flip test gave p = 0.25 under a symmetric null. Three of six patients had lower error. Mean relative error increased by 0.70 percent. All six fixed secondary checks also gave higher mean error. This test gives no clear evidence of a prediction gain for the specified feature and model. It does not exclude other network measures or show that power timing has no memory information. The package has source hashes, retrieval code, checked segment powers, all model fits, predictions, and repeat instructions.

working memory · SEEG · high gamma power · power correlation · recall prediction · ridge regression