Joint SEEG pair selection reduced input correlation but did not establish lower speech error at eight pairs and 400 ms delay in ten listeners

We tested whether a new SEEG pair should be chosen by its joint training error decrease. The comparison ranked each pair alone. We used public OpenNeuro ds004703 version 1.1.0 data from ten speech listeners. Both rules used eight pairs, five past power values per pair, and the same penalty search. Early passages chose pairs and fitted the readers. Later passages tested heard speech amplitude at 400 ms delay. Mean error skill was 4.74% for pair ranking and 2.99% for joint selection. The paired change was -1.75 percentage points. The descriptive 95% person bootstrap interval was -9.23 to +4.08 points. Six people improved and four had more error. Mean training input correlation fell from 0.124 to 0.075. A post-result equal-penalty check reduced the training objective in all ten people but did not show a clear later score gain. The package includes code, source records, data, models, predictions, four figures, and a raw retrieval recipe. All 95 source SEEG ranges and 39 sound ranges rebuilt the stored arrays exactly. The cohort and some sounds were reused. The result does not establish equal performance, intended speech reading, live control, or a stimulation effect.

stereo-electroencephalography · brain-computer interfaces · speech amplitude · ridge regression · feature selection