Loss of the most informative shaft reduced movement decoding AUC by 0.061 in 15 ds006253 SEEG patients

In an offline replication using archived 70–150 Hz high-gamma activity (HGA) window features from 15 patients in OpenNeuro ds006253 v1.0.3, a shrinkage linear discriminant analysis decoded movement (0–0.5 s after movement onset) against pre-stimulus rest. The model used the first 60% of trials for training and the last 40% for testing. Removing, at test time, the shaft containing the training pair with the largest absolute Welch t statistic reduced mean AUC from 0.833 to 0.772 with training-mean fill (mean loss 0.061; 15/15 patients, exact one-sided sign-flip p = 3.1e-5). Conditional-mean imputation (CMI) and refitting on surviving channels both gave mean AUC 0.782; CMI minus refit was +0.0001, within the 0.01 non-inferiority margin (15/15; Holm p = 9.2e-5). Refit recovered 0.0097 of the 0.061 loss, less than half in 14/15 patients (Holm p = 2.4e-4). The key-shaft loss exceeded same-count random-channel loss by 0.055 (13/15; Holm p = 2.4e-4). Mean loss across all shafts was 0.006, and the key shaft was the worst shaft in 10/15 patients. The findings replicate the reported ds004473 result in a second dataset, although movement coding in ds006253 is more distributed. This is an offline test with simulated test-time loss, one feature, one decoder, and one trial split; real shaft failure can add noise or artifacts. The included feature caches, code, results, and figure support rerunning the analysis. Source feature extraction is documented through derivative paths and hashes; the analysis does not establish byte-exact reconstruction from upstream recordings.

stereo-EEG · SEEG · high-gamma activity · brain-computer interface · electrode failure · linear discriminant analysis · conditional-mean imputation · AUC · replication