Channel availability raised mean test log loss on later trials after 60 percent training at penalty 10 in six patients with recordings marked as SEEG

We tested whether retained channels improve object recall prediction on later trials in public OpenNeuro ds006136 version 1.0.1. Seven processed recordings from six patients supplied 719 observed trials. Models used the first 60 percent of original trial IDs for training and 292 later trials for testing. The fixed primary comparison added channel presence to trial order, with ridge penalty 10. Mean patient test log loss rose from 0.6677 to 0.6692 nats per trial. Gain was -0.0015, with 95 percent t interval [-0.0221, 0.0191] and exact two-sided sign-flip p = 0.9375. Sessions had equal weight within patients. The test assumes independent patients and symmetric null gains. The interval assumes approximately normal patient gains. The result shows no clear mean benefit for this model. It does not show equivalent models or absence of rejection bias. Fixed penalty and split checks changed the gain sign. A constant model, added after the primary results, had lower mean loss than the four primary models. The source JSON marks the recordings as SEEG; channel tables mark them as EEG. Implant type was not independently verified. Whole-trial source processing limits the result to an offline test. The study reuses a prior cohort and is not an independent replication. The package preserves code, fixed source retrieval recipes, metadata, rebuilt features, models, every prediction, and separate signal and model checks.

object recall · channel availability · intracranial EEG · later-trial prediction · logistic regression · OpenNeuro