Forgotten words that later returned as intrusions had forgotten-like SEEG decoder scores

In an offline analysis of 18 patients in the FR1 delayed free recall task (OpenNeuro ds004789, version 3.1.0, session 0), we tested whether an encoding-period SEEG decoder trained to distinguish recalled from not-recalled words assigned higher scores to forgotten words that later returned as prior-list intrusions (PLIs). The decoder used L2 logistic regression on channel-by-band log-power features from 0–1.6 s after word onset, with leave-one-list-out scoring. For 80 forgotten-then-intruded words, the pooled within-patient AUC against never-intruded forgotten words was 0.520 for raw scores and 0.517 after serial-position control (one-sided p = 0.29; controlled null 95% range 0.438–0.563). A reference using recalled words from the same patient and serial-position bin had mean AUC 0.639 (95% range 0.588–0.689); the observed controlled AUC was below all 1000 reference draws (p = 0.001). Later intrusion did not change scores among recalled words (90 words; controlled AUC 0.491, two-sided p = 0.82). The median recalled-versus-not-recalled decoder AUC was 0.667. These results do not support recalled-like encoding scores for forgotten words that later intruded, but the small sample does not rule out a modest shift: the controlled gap-fraction 95% confidence interval was −0.09 to 0.83. The analysis was offline, included one session per patient, one linear decoder, and one feature set; word-text matching identified PLIs, and encoding and retrieval causes cannot be separated.

SEEG · memory · prior-list intrusion · free recall · encoding · brain-computer interface · logistic regression