Joint selection reduced contact correlation but gave no clear error gain for four SEEG contacts in ten speech recordings

We tested two contact selection rules in ten public SEEG recordings of spoken words. Each rule used four contacts and the same three features per contact. One rule used each contact alone. The other rule added the contact with the largest further ridge fit gain. Earlier trials supplied fitting and penalty selection. The remaining 194 complete trials supplied test scores. The primary mean relative error reduction was -0.11%. The descriptive 95% whole-person bootstrap interval was -2.66% to 2.07%. Five people had lower error, three had higher error, and two had equal error. Mean training contact correlation fell from 0.442 to 0.297. Separate contact scores reduced error by 12.80% against a cue-time model. Joint selection reduced it by 12.58%. These results do not establish equivalence or a neural cause. The test reused a small clinical cohort and assumed the source clock repair. The package includes all used features, model states, test predictions, trial permutations, source records, code, four figures, and replay steps. One fresh source download gave an exact feature match. Independent fits and selection checks passed. This study tests a reading method. It gives no stimulation result.

stereo-EEG · speech production · brain-computer interface · contact selection · ridge regression · speech amplitude · trial validation