In 8 SEEG subjects, beta improved high-gamma key-press detection with 8 channels (mean ΔAUC 0.031; Holm p = 0.016)

Question: an implant BCI records few channels. Does 13-30 Hz beta power add key-press information to 70-150 Hz high-gamma (HG) power when the decoder uses only 4, 8 or 16 channels? Data: archived features of sci 04683a2a (OpenNeuro ds004473, 8 SEEG subjects, 100-112 bipolar channels, hashes checked). Method: pre-registered plan (PLAN.md, hashed before analysis). Shrinkage LDA is trained on trials 1-100 and tested on trials 101+. Channels and features are ranked by paired |t| on the training trials only. Results: with the same 8 channels, HG+beta minus HG = +0.031 mean test AUC (7 of 8 subjects > 0, exact sign-flip Holm p = 0.016). With a budget of 8 features, the top 8 features from the pooled HG and beta set minus the top 8 HG = +0.040 (7 of 8, Holm p = 0.016). The 8-feature pooled decoder has mean AUC 0.828. This is equal to the all-channel HG+beta decoder (0.828, about 200 features). HG alone does not improve from 8 channels to all channels (0.789 vs 0.788). At K = 4 the gain is not consistent (HG+beta 5 of 8, pooled 3 of 8). The gain is largest in subjects with weak HG and near 0 when HG AUC is above 0.93. Limits: 8 subjects, one split, offline movement detection (not finger identity), clinical bipolar channels, small and heterogeneous effects. The package has code, input hashes, results JSON, tables, a figure and reproduction steps.

stereoelectroencephalography · brain-computer interface · high-gamma · beta power · key-press detection · feature selection · linear discriminant analysis