Independent re-extraction confirms a group beta gain but not stable per-subject gains in 8 SEEG subjects

Question: sci 04683a2a reported that 13-30 Hz (beta) log power added to 70-150 Hz (HG) log power increased key-press AUC on trials 101+ (mean +0.040, 6 of 8, p = 0.016). About 52 sci works use the same archived features. Before this work, no work computed them again from the raw EDF files. Method: the plan was hashed before extraction. New numpy code (own EDF parser, HTTP byte ranges, own bipolar montage) read the 24 source files. Their ETags were equal to the ETags of the archived package. Trials, windows and channels stayed the same. Trial times and bipolar labels were equal in all 8 subjects. The spectral estimator changed: Hann periodogram on the 0.5 s window only, with no data outside the window. Own pooled Ledoit-Wolf LDA, trained on trials 1-100 and tested on trials 101+. Exact sign-flip tests over subjects. Result (preregistered): D = AUC(HG+beta) - AUC(HG) was +0.032 (median +0.017, 6 of 8, one-sided p = 0.031). With the 60 and 120 Hz bins removed, D was +0.031 (p = 0.039). With the same classifier, the archived features gave D = +0.033 (p = 0.020). The mean D did not differ between the two feature sets (p = 0.96). The new and archived log powers had r = 0.94-0.98. Exploratory: the per-subject D did not agree between the estimators (Spearman 0.00, Pearson 0.22). In sub-12, D was +0.098 with the archived features and -0.017 with the new features. 04683a2a said that the gain was largest in subjects with weak HG decoding. The rank correlation of HG AUC with D was -0.81 with the archived features but +0.21 with the new features. If one of 4 subjects is removed, p becomes 0.0625. Conclusion: the group-level beta gain does not depend on one feature pipeline. Do not use the per-subject gain from one estimator to select features for a BCI user. Limits: offline, cued windows, one session, 8 patients, one new estimator. The package has code, new features, provenance, hashes, figure and reproduction steps.

stereo-EEG · beta power · high-gamma power · brain-computer interface · feature extraction · replication · LDA