Up to 120 calibration words do not significantly improve SEEG word recall decoding over a fixed spectral tilt

We compared zero-calibration read-outs with within-subject logistic decoders using per-word, bipolar SEEG band-power features from 19 pyFR participants. The test set was the later half of each participant's lists; calibration used 1, 2, 3, 4, 6, or 8 earlier lists (15 words per list), either the first lists or 20 random draws. A fixed spectral-tilt read-out had median test AUC 0.538 (one-sided p = 0.072 vs 0.5); frozen FR1 weights had median AUC 0.529 (p = 0.062). With the first k lists, the channel-by-band decoder had median AUC 0.548–0.572 and was above chance from k = 2 (p ≤ 0.013). Its median gain over spectral tilt was +0.035 to +0.050 for k = 3–8, but no gain was significant after Holm correction (best corrected p = 0.093 at k = 6). Random calibration gave no corrected significant gain (all Holm p ≥ 0.29). A decoder using eight channel-mean bands showed median gains from −0.022 to +0.012. These results indicate weak recall decoding in this dataset: calibration may provide a small channel-specific gain, but up to 120 labelled words did not establish a significant improvement over the fixed read-out. Limitations include 16–19 participants per comparison, one later-half test split, session-wide label-free z-scoring, and untuned decoder settings. The dataset release used by the upstream feature publication is not identified exactly; per-file source URLs and hashes are preserved in the supporting files.

SEEG · word recall · memory decoding · brain-computer interface · calibration · band power · AUC · pyFR