SEEG band power did not decode word category above a structure-preserving null in 8 catFR1 patients

We tested whether clinical SEEG band power during word encoding carries a cross-list code for semantic category in catFR1 session 0 from 8 patients. Features were bipolar-channel Morlet-wavelet power at 8 frequencies (3–140 Hz), averaged from 0–1.6 s after word onset. The planned primary test used a leave-one-list-out correlation-centroid decoder, list-wise z-scoring, and 500 permutations that preserved the task's list, pair, and position structure. Accuracy was 0.340 versus a null mean of 0.349 (chance about 1/3; exact one-sided Wilcoxon p = 0.77; Fisher p = 0.44); 1/8 patients had p < 0.05. Naive word-level 5-fold cross-validation exceeded a word-shuffle null in 7/8 patients (Wilcoxon p = 0.004), but not a structure-preserving null (Wilcoxon p = 0.58; Fisher p = 0.75), indicating that the apparent naive effect was explained by task structure. No single frequency exceeded its null (Holm-corrected Fisher p = 1.0 for all 8); controlling for letter count and serial position also gave no group effect (Wilcoxon p = 0.63). Letter count alone decoded category above chance (Wilcoxon p = 0.004). As a positive control, the same pipeline decoded later recall (median AUC 0.59; 4/8 patients p < 0.05; Wilcoxon p = 0.020). The study is limited to 8 patients, session 0, time-averaged band power, and one correlation-centroid decoder; small effects cannot be excluded.

SEEG · catFR1 · semantic category · band power · word encoding · leave-one-list-out · temporal adjacency · memory decoding · cross-validation