Adding band correlations increased mean later log loss for rest and pre-response states in eight SEEG recordings

We tested whether band correlations add useful information to SEEG band power. We used rest and pre-response windows from eight public recordings in OpenNeuro ds004473 version 1.0.2. Each reader learned from early trials and predicted later trials. The electrode pairs were fixed by an earlier power study. Penalties were chosen within the early period. Mean later log loss was 0.53489 with power alone and 0.56591 with added real band correlations between different electrode shafts. Six of eight recordings had higher loss with the added features. The paired loss reduction was -0.03102. Its descriptive 95% person bootstrap interval was [-0.06609, +0.00008]. The exact two-sided sign-flip p value was 0.140625. Thus, the tested reader did not show a mean gain. The interval includes zero change. Fixed controls tested imaginary components, correlation alone, phase shuffles, and a fixed penalty. The package includes exact voltage buffers, source range records, all predictions, two figures, and code. Independent checks and a full numerical replay passed. This exploratory result applies to these selected pairs and event windows. It does not test continuous control or brain stimulation.

stereoelectroencephalography · SEEG · band power · band correlation · cross spectrum · logistic regression · brain-computer interface