Hippocampal theta patterns raised mean test loss across sessions in 13 epilepsy patients with a fixed logistic model

This study tested hippocampal theta contact patterns for memory probe prediction across separate numbered sessions. It used public SEEG from OpenNeuro ds004752 version 1.0.1. A fixed logistic model used task history and mean theta change. The full model also used differences between contacts. The primary test retained 13 epilepsy patients, 56 sessions, and 2,590 trials. Adding the contact pattern raised equal-patient mean test log loss by 0.01429 nats per trial. Twelve patients had worse loss. The descriptive patient bootstrap interval for the gain was [-0.02552, -0.00022]. The exact two-sided sign-flip p value was 0.057617 under independent effects and a symmetric null. The primary test did not meet the planned 0.05 level. All eight planned secondary checks had negative mean gains. The result concerns these fixed features and models. It does not show that theta lacks a memory role. Source checks covered pre and post windows in all 62 source sessions and two complete raw recordings. A separate solver checked all 224 fits. An offline repeat matched all 17 result files. The package preserves code, source features, provenance, predictions, model fits, uncertainty results, and reproduction commands.

SEEG · hippocampal theta · working memory · logistic regression · test log loss · epilepsy · session prediction