In six OpenNeuro ds005415 SEEG subjects, stimulus-locked features did not decode modality, format or number

We tested whether a linear decoder could read stimulus modality, stimulus format or number (1–9) from SEEG in the OpenNeuro ds005415 numbers task, using run 1 from six subjects with SEEG. We used corrected trial onsets (Cue for visual trials and the k-th Cue Start for the k-th auditory trial) and compared them with naive Cue onsets for all trials. The plan was written before neural data were read. Features were bipolar 70–150 Hz log analytic envelope and 0.5–8 Hz voltage, in 10 bins of 100 ms after onset relative to baseline. Modality and format used dual ridge decoding, stratified 5-fold cross-validation with 10 repeats and 200 label permutations; number used ridge regression and Spearman correlation within and across formats. Group tests were exact one-sided Wilcoxon or sign-flip tests over six subjects, with Holm correction. With corrected onsets, modality AUC was 0.37–0.55 and exceeded the permutation 95th percentile in 0/6 subjects (planned threshold: at least 5/6). Corrected onsets did not improve modality decoding (W = 7, p = 0.78) or median format-pair decoding (W = 8, p = 0.72). Across 10 format pairs, mean AUC was 0.46–0.54 (all Holm p ≥ 0.93). Number decoding was at or below chance within formats (mean rho −0.27 to −0.11; all Holm p = 1) and across formats (mean rho 0.00–0.14; all Holm p ≥ 0.37). The held-out evoked high-gamma control also failed in all subjects (median t −0.2 to 1.8 for visual trials and −0.2 to 1.6 for auditory Cue Start trials; Cue Start vs Cue p = 0.28). Naive onsets gave modality AUC 0.62–0.68 in subjects 73 and 78, but auditory epochs began about 1.2 s later in the trial, so baseline windows differed between classes. These results show that these onsets and features did not provide a usable stimulus-locked signal; they do not show that number is absent from the brain. A mismatch between events.tsv and the EDF clock (463 ms median offset in subject 73 in a prior timing report), or nonresponsive implanted sites, are possible causes; the study did not determine which. Only run 1, one feature set and one linear decoder were tested, and a small effect is not excluded. Features were z-scored over all subject trials without labels to permit a precomputed kernel, deviating from the plan; in subject 73, per-fold z-scoring gave the same AUC (0.394 vs 0.393).

SEEG · stimulus-locked decoding · neural decoding · intracranial EEG · OpenNeuro · negative result · number processing