Scalp theta patterns reduced mean SEEG error across two memory sessions with ridge penalty 10 and a common average scalp reference in 14 patients

We tested whether scalp theta differences among channels reduce local SEEG prediction error after a model has mean scalp theta power and task features. We used public OpenNeuro ds004752 version 1.0.1, with 14 patients, two memory sessions per patient, and 1275 retained trials. Both models used ridge penalty 10 and a common average scalp reference. We trained in one session and tested in the other, in both directions. The mean normalized error fell from 1.195638 to 1.145301. The mean patient decrease was 0.050338 training-variance units. Nine of 14 patients improved. The patient bootstrap 95% interval was 0.005872 to 0.102789. The exact two-sided sign-flip p value was 0.067017, under independent patients and a symmetric null distribution. With penalty 1, the scalp pattern raised correlation but also raised error. With the stored scalp reference, mean error gain was 0.005879. Relative depth-contact power had a smaller descriptive gain. The contact signals have hippocampus-labelled anchors and do not isolate hippocampal sources. The result is exploratory and applies to the selected sessions and models. The package preserves code, input features, fitted models, predictions, source versions and hashes, independent checks, and a complete retrieval recipe.

scalp EEG · SEEG · theta power · working memory · ridge regression · cross-session prediction · prediction error · intracranial EEG