Across-channel noise correlations limited linear SEEG decoding in 8 subjects

We tested whether across-channel trial-to-trial noise correlations limit linear decoding of rest versus key press from stereo-electroencephalography (SEEG) band power. Using archived features from 8 subjects in OpenNeuro ds004473 (snapshot 1.0.2), we trained shrinkage linear discriminant analysis on trials 1–100 and tested on trials 101 and later. Within each class and split, independent channel-wise trial shuffles preserved each channel’s means and variances while removing across-channel noise correlations. For high-gamma features (70–150 Hz), the shuffled condition increased held-out AUC in 7 of 8 subjects (mean difference +0.047; exact two-sided sign-flip p = 0.023). For high-gamma plus beta features (13–30 Hz), the increase was +0.076 (7 of 8; p = 0.016). On held-out trials, noise variance along the training class-difference axis exceeded the independent-channel value in all 8 subjects (mean log ratio 1.17, approximately 3.2-fold; p = 0.0078). Full-covariance LDA did not recover the loss for high-gamma features (diagonal minus full LDA mean AUC difference −0.007). The shuffle is a counterfactual, and some AUC increase may reflect easier covariance estimation. Noise includes drift, state change, and volume conduction, which were not separated. The study includes 8 subjects from one task; secondary tests were not corrected for multiple comparisons.

stereo-electroencephalography · noise correlations · linear decoding · brain-computer interface · high-gamma activity · shrinkage LDA