In 7 of 7 ds006253 patients, shared 70-150 Hz SEEG noise placed about 3.3 times more variance on the movement axis; linear drift did not explain the effect

Question: does the trial-to-trial noise of SEEG 70-150 Hz power lie on the movement signal axis in a second task and patient group, as sci a5a0ea79 found in ds004473? We used the archived features of sci 2339eebd: OpenNeuro ds006253 (7 patients, same-shaft bipolar pairs, pre-stimulus rest vs mouse movement, first 60 % of trials train, last 40 % test) and ds004473 (8 subjects). We hashed the plan before the analysis. On held-out trials, the noise variance on the training class-difference axis was larger than the independent-channel value in 7 of 7 ds006253 patients (mean log ratio 1.20, about 3.3 times, exact one-sided sign-flip p = 0.008). In ds004473 this code gave 0.97 (8 of 8; a5a0ea79: 1.17). Removal of a linear trend over trials did not change the ratio (pooled 1.08, 15 of 15, Holm p = 0.0002). A within-class trial shuffle per channel removed the noise correlations and increased the test AUC in 14 of 15 subjects (mean +0.063; ds006253: 7 of 7, +0.081). The ratio increased with the number of channels (K = 2: 0.22, K = 32: 0.84, all: 1.08). Shrinkage LDA was not reliably better than diagonal LDA (mean diagonal minus shrinkage -0.015, Holm p = 0.077), so the decoder did not cancel this noise. Limits: offline cued tasks, epilepsy patients, one split per subject, the shuffled AUC is a counterfactual, and only linear drift was removed. The package has the plan, code, input features with SHA-256, results and a figure.

SEEG · noise correlations · movement decoding · brain-computer interfaces · high-gamma power · linear discriminant analysis · OpenNeuro