In 83 SEEG subjects, encoding band-power patterns drifted with study lag but did not predict consecutive recall (ds004789 FR1, session 0)

We tested two predictions of retrieved-context theory in SEEG data from 83 subjects (OpenNeuro ds004789, FR1 free recall, session 0). We used per-word log10 band power (theta, alpha, beta, low gamma, HFA; 0-1.6 s after word onset; bipolar pairs) from AutoSci 9da8ac2f. We removed the serial-position mean of each feature and calculated the Pearson r between word patterns in the same list. The plan was written before we calculated any similarity statistic. H1 (positive control): similarity at study lag 1 was higher than at lags 4-11 (mean difference 0.034 r, 95% CI 0.028 to 0.039; 79 of 83 subjects positive; Holm p = 0.0001). Similarity decreased with lag and was lower again between adjacent lists. Behaviour showed strong temporal contiguity (lag-CRP at lag 1 minus far lags = 0.16). H2 (main): for adjacent study pairs with both words recalled, pairs recalled one after the other did not have higher residual similarity than other pairs (mean -0.003 r, 95% CI -0.025 to 0.018; 34 subjects with 5 or more pairs per class; Holm p = 0.81). A post hoc weighted test with 74 subjects gave the same result (-0.003, CI -0.019 to 0.014). HFA-only, theta-only, MTL-only and Spearman versions did not show a significant effect after Holm correction. Thus, these coarse band-power patterns drift in time, but effects larger than about one lag step of drift (0.016 r) are unlikely. The result applies only to this feature set, this window and session 0. Drift can also come from slow signal changes that are not a memory context. The package contains the plan, the code, the events with SHA-256 provenance, the hashes of the input features, all results, and the steps to reproduce.

SEEG · free recall · temporal contiguity · neural drift · band power · retrieved context