A fixed half-weight contact model increased alpha prediction error across stimulation shafts in 14 SEEG patients

We tested whether a contact-specific alpha response map could improve prediction when stimulation moved to another electrode shaft. We used public SEEG data from OpenNeuro ds005169 version 1.0.0. All recordings on the target stimulation shaft were excluded from training. The test used 83 recordings from 14 patients. A fixed model added half of the learned contact deviation to a recording-shaft mean. This increased mean squared error by 8.0%, from 9.441 to 10.198 dB squared. Mean error gain was -0.757 dB squared, with a patient bootstrap 95% interval of -1.237 to -0.350. The exact two-sided P value was 0.000977 under independent patients and symmetric null gain signs. Eleven patients had more model error. Fixed alternative weights and a pre-only check also had more error. Independent calculations checked all predictions. Fresh pinned raw ranges from all 14 patients reproduced the stored spectra. This result applies to the fixed model and short intervals in these implants. It does not establish a causal network or general alpha response instability. The package preserves spectra, source recipes, code, every prediction, results, and audits. This is a secondary analysis of the same cohort used in the cited source publications.

SEEG · intracranial EEG · alpha response · prediction error · cortical stimulation · cross-shaft prediction