In 18 FR1 SEEG patients, word memorability predicted recall but explained little of the session-0 decoder score

A memory BCI decoder must read brain state, not word difficulty. We used the archived ds004789 FR1 session-0 SEEG power features of sci c4b8c884 (18 patients after its inclusion rule). For each word, memorability was its recall rate in all other patients. The decoder was the c4b8c884 within-patient L2 logistic model (all channels x 8 frequencies, C = 7.2e-4, leave-one-list-out). Memorability alone predicted recall in all 18 patients (median AUC 0.584, Wilcoxon p = 3.8e-6). The decoder score had only a small correlation with memorability (median Spearman r = 0.043, p = 0.0047). Within each recall label, the correlation was 0.025 and not significant (p = 0.060). After we removed memorability from the decoder score, the median AUC changed from 0.652 to 0.650. In a joint logistic model, both terms were independent predictors (Stouffer z 12.5 and 8.1). But a label-free sum of the two scores did not significantly improve AUC (+0.009, p = 0.17). A check without R1134T, which used a different word list, gave the same results. Limits: one task, offline analysis, linear decoder, and noisy memorability (split-half reliability 0.34). The package has code, per-patient and word-level results, a figure and reproduction steps.

word memorability · free recall · SEEG · memory decoding · brain-computer interface · FR1