The SEEG word-recall decoder had similar low-score and high-score tail separation in 18 patients

In an offline analysis of archived per-word band power from 18 patients in the OpenNeuro ds004789 FR1 session 0 cohort, the low-score tail of a within-patient leave-one-list-out L2 logistic decoder separated forgotten from recalled words about as well as the high-score tail separated recalled from forgotten words. At q = 0.2, median tail separation was 0.222 in the low tail and 0.223 in the high tail (median low-minus-high difference 0.000, 95% CI -0.038 to 0.040; exact two-sided Wilcoxon p = 0.70). Results were similar at q = 0.1 and 0.3, for positions 4-12, and for the binormal score-SD ratio. A serial-position-only positive control showed asymmetry (0.238 versus 0.124, p = 0.0004), indicating sensitivity to an asymmetry of this size. In the lowest 20% of decoder scores, 0.90 of words were forgotten, compared with a 0.78 base rate; in the highest 20%, 0.33 were recalled, compared with a 0.22 base rate. The result is consistent with an approximately symmetric ROC for this decoder, but smaller asymmetries remain possible. The analysis was offline, used one session per patient and non-causal session-wide z-scoring, tested one decoder, and did not test stimulation effects. No version of the source dataset is established by the recorded provenance.

SEEG · free recall · word memory · ROC curve · decoder · tail asymmetry · closed-loop stimulation · OpenNeuro ds004789