Late delay power increased mean recall test log loss from 0.604 to 0.609 in fixed models for 12 SEEG patients

We tested whether late delay beta and high gamma power improved full recall prediction on later retained trials. The source was OpenNeuro ds006136 version 1.0.1. Fixed logistic models trained on early trials in each bipolar SEEG channel. The baseline used trial order, fixation power, and encoding power changes. The full model added two late delay power changes. In 12 epilepsy patients, mean test log loss increased from 0.60384 to 0.60889. Lower loss is better. Nine patient results showed higher loss with the added features. The 95% patient bootstrap interval for the increase was 0.00105 to 0.01005. A t interval included zero. All four fixed sensitivity estimates had the same direction, but two bootstrap intervals included zero. The result applies to these models and the selected channels and retained trials. It does not test unseen patients or the source paper's burst measures. The archive contains trial powers, raw validation samples, source hashes, pinned retrieval code, models, predictions, and reproduction instructions. Independent checks passed for all 117,702 source power values and all 274 primary model fits. A complete numerical replay also passed.

working memory · recall prediction · SEEG · beta power · high gamma · logistic regression · test log loss