Under assumed epoch timing, eight added high gamma features changed mean load AUC by 0.002 in 13 final SEEG sessions

We tested whether eight high gamma power features improve a fixed SEEG readout of four versus eight letters. We used OpenNeuro ds004752 version 1.0.1. Thirteen patients supplied 56 sessions. Training used all earlier sessions. Testing used 377 trials in total from the final sessions. Trials followed a correct response and had a source Artifact flag of zero. The base model used task history and eight power features from 4 to 30 Hz. The added features used 70 to 110 Hz without 98 to 102 Hz. Mean patient AUC changed from 0.570 to 0.571. The paired change was +0.0016. Its descriptive 95% patient bootstrap interval was -0.0423 to +0.0520. The exact sign-flip p value was 0.950. Five patients gained AUC and eight lost AUC. Fixed penalty checks also gave small mean changes. This test found no clear mean gain. It does not establish equal performance or the absence of high gamma information. The analysis assumes the source epoch timing. Eleven final sessions shared a date with the prior session. Earlier work used these test sessions. This is an exploratory extension. A fresh raw audit checked 448 power values in all 56 sessions. The package contains feature inputs, raw audit ranges, source records, fitted models, predictions, figures, code, and instructions.

SEEG · working memory · high gamma · AUC · feature selection · logistic regression