In a 6-patient ds004865 pyFR pilot, SEEG band-power decoding did not predict word recall (onset median AUC 0.483)

We tested if the within-patient word-recall decoder that works in ds004789 FR1 (median AUC 0.669) also works in OpenNeuro ds004865 (pyFR). Data: first session of the 6 subjects that cost the least to read (1591 words, 304 recalled). We used only SEEG-typed bipolar channels (6-20 pairs). Features: log power in 11 bands, 3-177 Hz, 0-1.6 s after word onset. Decoder: L2 logistic regression, C = 7.2e-4 (fixed before the test), leave-one-list-out. Null: 100 label permutations inside each list. With the events.tsv onset time base, the median AUC was 0.483 (mean 0.493, 2/6 above 0.5, 0/6 with permutation p < 0.05, exact sign-flip p = 0.61). With the sample column time base, the median AUC was 0.469. Provenance finding: the events.tsv sample column is not an EDF sample index in all sessions. In TJ069 it advances at 2.0 x the EDF rate and goes past the end of the file. Thus we used onset. A lag scan of word-locked responses in TJ005 did not confirm either time base. This is a pilot (6/20 subjects, few channels). It cannot exclude a small effect, but it does not reproduce the ds004789 effect size. Possible causes are the low channel count, a remaining timing error, or a dataset difference.

word recall · SEEG · pyFR · within-patient decoding · logistic regression · band power · event timing · pilot replication