In 15 SEEG patients, label-free high-gamma rules did not detect mouse movement, while three labeled trials improved decoding

This pre-registered offline replication tested whether label-free high-gamma rules detect movement in a second stereoelectroencephalography (SEEG) dataset. We analyzed archived log 70–150 Hz features for 15 patients in OpenNeuro ds006253: mouse movement (0–0.5 s after movement onset) versus pre-stimulus rest (0.5 s). The primary rule, the first principal component (PC1) of unlabeled windows with its sign set so high gamma goes up, did not detect movement (median AUC 0.453, mean 0.472, range 0.224–0.852; 6/15 patients above 0.5; exact one-sided sign-flip p = 0.751); PC1 had the wrong sign in 9/15 patients. Other label-free rules also failed: channel mean median AUC 0.470, top-10 variance 0.484, causal PC1 fitted on trials 1–20 0.476, and PC1 fitted on the first 60% of trials 0.476 (all one-sided p ≥ 0.38). Shrinkage LDA with 3 labeled trials gave median AUC 0.564 (12/15 above 0.5, p = 0.0045); with 5, 10, and 20 trials, medians were 0.660, 0.723, and 0.743, respectively (20 trials: 15/15 above 0.5, p = 3.1e-5). Five-fold cross-validation gave 0.865. Causal PC1 was lower than LDA with 3 trials (median difference −0.119 AUC, two-sided p = 0.011) and 5 trials (−0.179, p = 6.1e-5); at 2 trials the difference was not significant (LDA 0.497 versus PC1 0.476, p = 0.32). The pre-registered interpretation was that the earlier ds004473 null result replicated. Exploratory results, per-patient values, methods, the pre-registration, and complete limitations are in the main text. Limits include one session per patient, offline windows aligned to known onsets, different task phases for rest and movement, only 70–150 Hz features, unknown hand used, and n = 15; the sign-flip test cannot exclude a small positive mean effect, and a 95% range was not reported.

stereoelectroencephalography · brain-computer interface · movement decoding · high gamma · zero calibration · linear discriminant analysis