Adding hippocampal fixation theta reduced mean log loss by 0.0014 for later numbered session errors in 12 SEEG patients with a fixed model

We tested whether hippocampal theta before letter presentation helps predict current trial errors in later numbered sessions. We used public SEEG from OpenNeuro ds004752 version 1.0.1. The fixed split gave 1,140 training trials and 1,271 test trials in 12 patients with epilepsy. Each patient had a ridge logistic model. The baseline used scheduled load, recent behavior, and a broad spectral line. The full model added a fixed 4 to 8 Hz contrast above that line. Mean test log loss decreased by 0.001377 natural log units per trial. Seven patients had a positive gain. The exact two sided patient sign flip p value was 0.09180. The 95% patient bootstrap interval was [0.000124, 0.002868]. It uses the fitted models and excludes training sample variation. The full model had higher loss than the simpler behavior model. The common average reference reduced the added gain to 0.000234, with an interval that included zero. These results do not establish a reliable error marker. This is an exploratory study of a reused cohort. New checks covered raw ranges in all 55 used sessions, one full source rebuild, 108 model fits, and an offline repeat of 121 outputs. The package includes the fixed plan, source spectra, metadata, raw examples, hashes, code, predictions, model arrays, results, figures, and repeat instructions.

stereoelectroencephalography · hippocampus · theta power · working memory · error prediction · logistic regression · session holdout · log loss