Word length was weakly related to recall, but did not drive the SEEG recall decoder in 18 FR1 patients

A memory brain-computer interface (BCI) should read memory state, not stimulus properties. We tested whether word letter count was a shortcut for a within-patient SEEG recall decoder, using archived band-power features for 18 patients in OpenNeuro ds004789 v3.1.0 (session 0; 8 frequencies, 3–140 Hz, 0–1.6 s after word onset). The decoder was the unchanged c4b8c884 L2 logistic model with leave-one-list-out evaluation. Longer words were recalled less (median Spearman rho −0.063; 15 of 18 patients negative; Holm p = 0.031), with pooled recall of 0.27 for 3-letter words and 0.11 for words of 7 or more letters. Decoder score did not follow letter count (median rho −0.000; mean +0.002; 95% bootstrap CI −0.035 to +0.042; p = 0.90). Median recall AUC was 0.669 over all words and 0.666 within letter-count strata (p = 0.44; Holm p = 0.88). Letter count alone gave median recall AUC 0.540. A positive control found weak letter-count information in the same features (ridge prediction median rho 0.064; 14 of 18 positive; Holm p = 0.16). The averaged features therefore provide a limited test and cannot exclude other visual confounds. Seventeen of 18 patients saw words from the same 300-word pool, confounding word identity and length. This was an offline analysis of one session and one dataset; null results do not prove zero effect.

SEEG · free recall · memory decoding · brain-computer interface · word length · stimulus confounds · band power