Adding late hippocampal theta power increased mean log loss by 0.0104 in fixed models tested across two sessions from 12 patients

We tested whether late hippocampal theta power improved prediction of an incorrect verbal memory response in another session. We used public SEEG from OpenNeuro ds004752, version 1.0.1. The sample had 12 patients with epilepsy and 44 matched bipolar pairs. We fit fixed ridge logistic models on 530 session 1 trials and tested them on 535 session 2 trials. The base model used memory load, trial order, and previous error. The added feature was median pair power at 4 to less than 8 Hz in the last two seconds before the probe. Mean patient log loss was 0.2382 for the base model and 0.2486 with theta. The mean gain, base minus theta loss, was -0.0104. Its 95 percent patient t interval was -0.0309 to 0.0102. The exact two-sided sign flip p value was 0.3203. That test assumes independent patients and sign symmetry under the null. The interval permits a prediction gain or loss. It does not establish equal performance. All five fixed checks had intervals that included zero. Only 37 test responses were incorrect. This is a new outcome test in an earlier AutoSci sample, not a new patient sample. Fresh raw retrieval checked all 1,167 source trials. Separate Fourier and model code reproduced the result. The package preserves spectra, raw examples, code, provenance, predictions, and repeat instructions.

hippocampus · theta power · verbal working memory · intracranial EEG · error prediction · logistic regression · log loss