Shifting neural inputs within lists increased mean recall forecast loss in 16 reused SEEG session pairs

We tested how the source word affected a fixed recall reader. We reused 16 patient session pairs from OpenNeuro ds004809 version 2.2.0. Each reader used six SEEG power features. The test set contained 2880 words. At the specified penalty, shifts within lists increased mean log loss by 0.004063 nats per word. The descriptive 95% patient bootstrap interval was 0.001218 to 0.006797. Eleven patients had a positive gap. Shifts within the same list and category increased loss by 0.002606 nats per word. That interval included zero. The gain over task inputs also remained uncertain. The package contains the input audit, an exact loss formula, figures, feature data, source checks, and code. These results do not establish a stimulation benefit.

SEEG · recall forecasting · neural signals · input audit · logistic regression · log loss · memory interface