Word-locked SEEG features predict recall within serial position and remove negative state carry-over in 48 RAM FR1 subjects

The encoding-classifier analysis of RAM FR1 SEEG data used epochs timed from the events.tsv onset column, which was not EDF time in these data. We ran the original analysis code without changes on two feature sets for the same 48 eligible subjects: sample/fs-timed word-locked features and the original onset-timed features. Onset-timed results reproduced the archived per-subject values exactly. With word-locked features, classifier-state AUC was 0.604 versus 0.533 with onset timing (paired difference +0.071, 95% CI 0.044 to 0.096, sign-flip p<1e-4). Within-serial-position AUC minus 0.5 was +0.053 (p<1e-4; 36/48 subjects above 0.5), while the onset-timed value was -0.004 (p=0.64). The negative coefficient of the state for word n-1 in predicting recall of word n was absent with word-locked features (+0.004, p=0.35); it was -0.009 with onset timing (p=0.032; Holm p=0.065 in this subset), with paired change p=0.022. Lag-1 state persistence above a within-list shuffle null was present in both runs and did not differ by timing (+0.067 and +0.087, both p<1e-4; paired p=0.27); raw lag-1 correlations were near zero. The comparison covers one RAM FR1 dataset family, 48 subjects, five log-power bands, and bipolar pairs. The classifier used the original settings without tuning. The pre-run decision rule and full results are in the README and supporting files.

SEEG · memory · free recall · word-locked · encoding classifier · serial position · RAM FR1 · feature timing