Same-list recall ranking performs above chance in an 18-patient FR1 SEEG decoder

We tested whether a within-patient word-recall decoder’s performance came from ranking words within the same list or from differences between lists. Using archived log10 Morlet power features and decoder W_chan in 18 patients from OpenNeuro ds004789 v3.1.0 (FR1, session 0, offline), same-list AUC had a median of 0.659; 17/18 patients were above 0.5 (exact Wilcoxon p = 7.6e-6; Holm-adjusted p = 2.3e-5). All-pair AUC was 0.669, reproducing the source decoder result. The paired same-list minus all-pair difference was +0.014 (11/18 positive; p = 0.15; Holm-adjusted p = 0.30). Different-list AUC was 0.668; its difference from same-list AUC had a median of -0.015 (p = 0.15). The median fraction of above-chance AUC retained within lists was 1.08. List-mean output tracked list recall fraction (median Spearman rho = 0.15; 14/18 positive; p = 0.012), while a median 14% of output variance was between lists. Centering features by list did not change same-list AUC (median difference -0.006; p = 0.15; Holm-adjusted p = 0.30). Exploratory same-list AUC for serial positions 4–12 was 0.627 (16/18 above 0.5; p = 3.8e-5). These results support word-level decoder information and provide a basis for word-by-word closed-loop memory stimulation; removing list means was not necessary in this analysis. Limits: one session, one linear decoder, offline data, 18 patients, and a list-centered variant that uses all words in the test list and is not causal. The tests can detect only moderate AUC differences (about 0.02 or more). Recall labels came from one free-recall test per list; output order was not used.

intracranial EEG · SEEG · free recall · word recall · memory decoding · area under the curve · within-list ranking · closed-loop stimulation

Admitted 2026-11-09 17:35 UTC4 citing publications

References

supportsIn 18 ds004789 FR1 SEEG patients, previous-word decoder output weakly predicted next-word recall, while averaging consecutive outputs reduced recall ranking sci:sha256:c0dbde9e1a4edbfea60de10c703c77df04d4378d11673a3b83e57b852130cf12
It found that the lag-1 decoder signal is a slow list-level state; we show that this list-level state is small and that the decoder AUC for the current word does not depend on it.
usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
We use its archived FR1 band-power features and its unchanged W_chan decoder code; our all-pair AUC (median 0.669) reproduces its result.
usesdoi:10.18112/openneuro.ds004789.v3.1.0
The archived FR1 features used by this analysis are identified as ds004789 v3.1.0; the exact release metadata records this DOI and a CC0 license.

Cited by

4 publications

Public catalog entries on Sciscape.

uses31 sci packages yield 6 distinct ds004789 FR1 baseline result sets, differing by median 0.044–0.066 AUC per patient across 17 patients
We used column A_pooled of results/per_patient.csv as the baseline AUC (family F1).
challengesWith 18 SEEG patients, most audited null decoder claims fell below 80%-power thresholds
The null claim reports an absolute effect of 0.010 AUC with N = 18. Our C x 10 noise model gives MDE80 0.0175 at this N, so the test had less than 80% power for this effect. The result does not show equal decoders.
challengesWithin-list band power did not predict later recall in 19 pyFR SEEG sessions (median AUC 0.483)
The same kind of within-list FR1 recall decoder (median AUC 0.659) did not decode recall in an independent pyFR cohort.
usesIn six pyFR SEEG patients, an FR1 word-recall decoder did not exceed within-list permutation nulls
FR1 reference values: all-channel within-patient recall AUC (median 0.669) and the same-list AUC metric, which we applied to pyFR.