A fixed model with study high gamma had higher mean later recall pattern loss within lists in 18 SEEG patients

We tested which word pairs patients recalled in the same six-item list. The source was public SEEG from OpenNeuro ds005059 release 1.0.6. The analysis reused verified features for 18 patients. Early lists supplied fitting data. Later lists supplied test data. An exact conditional model held the observed correct count fixed and removed a common list score. With fixed L2 penalty 0.1, mean later loss was 2.533308 for study position and 3.533076 with study high gamma. Loss was higher in 16 patients. The mean increase was 0.999768 natural-log unit per informative list. Its 95 percent patient bootstrap interval was 0.554090 to 1.510808. The exact sign-flip p was 0.000106812, under independent patients and symmetric null gains. Stronger regularization reduced the loss increase. Mean within-list AUC increased from 0.531297 to 0.558627. Thus the result applies to this fixed probability model and leaves open useful rank information. The test used 111 informative later lists. It does not predict the recall count or supply a real-time decoder. This is a new analysis of a reused cohort. Independent calculations checked all primary fits, labels, and conditional scores. A full repeat gave identical output hashes. Fresh pinned retrieval checked all 2334 source trial ranges and reproduced every power feature exactly. The package preserves complete feature inputs, code, proofs, source pins, hashes, results, and reproduction steps.

paired associate learning · episodic memory · stereoelectroencephalography · high gamma · conditional prediction · recall patterns · within-list prediction · OpenNeuro