Adding a band contrast reduced mean test error beyond features outside the band in eight SEEG records

We tested whether a 70 to 110 Hz band feature adds task information beyond a spectral trend outside that band. We used public OpenNeuro ds004473 version 1.0.2. Eight people each supplied eight fixed bipolar pairs. Readers trained on early trials and tested on 674 later trials. Each trial supplied rest and a window before a response. The band contrast was band log power minus a line fitted at 35 to 55 and 130 to 170 Hz. Adding this contrast reduced equal-person mean Brier error from 0.2042 to 0.1840. Seven people improved. The mean change was -0.02022. Its descriptive 95 percent bootstrap interval was -0.03625 to -0.00638. All five fixed sensitivity checks had the same mean direction. Contrast alone had more error than band power alone. The result supports keeping band and trend features separate in this reader test. It does not identify a separate neural rhythm. Contacts and test trials came from earlier studies. This exploratory test does not measure new-day performance, live control, or stimulation. The package preserves voltage windows, source checks, all fits and probabilities, four figures, and repeat instructions.

stereo-EEG · brain-computer interfaces · spectral features · Brier score · logistic regression · high-frequency power · reproducibility