Task history had lower mean memory load prediction loss than history with fixation power subtraction in session tests from 11 epilepsy patients

We tested fixation subtraction for hippocampal memory-load prediction in OpenNeuro ds004752 version 1.0.1. The main tests used 11 epilepsy patients, 56 sessions, and 2565 trials. Models used six power bands and task history. Separate inner sessions chose each ridge penalty. Fixation subtraction increased mean patient log loss by 0.021231 nats over task history alone. The 95 percent patient interval for loss gain was [-0.030173, -0.013463]. All 11 patients had higher loss. The Holm-adjusted sign-flip p value was 0.001953. The mean gain over absolute maintenance power was -0.007384; its adjusted p value was 0.075195. Patient inference assumes independent sampling and symmetric null effects. These are retrospective session tests of scheduled load. The history comparison interval included zero in the fixed check that used only trials after correct responses. This secondary analysis shares its cohort with the cited parent work. The package preserves inputs, exact source versions, code, models, predictions, checks, and an exact offline repeat.

hippocampus · intracranial EEG · verbal working memory · memory load · session prediction · log loss · fixation normalization · ridge regression