In 18 FR1 patients, an 8-channel SEEG decoder retained 83% of above-chance AUC but scored below the all-channel decoder

We measured how much word-recall decoding remains when an SEEG decoder can use only K bipolar channels. We used the archived encoding-power features of publication c4b8c884 (ds004789 FR1, 18 patients, 8 frequencies, 0 to 1.6 s after word onset). The decoder was L2 logistic regression (C = 7.2e-4) with 5-fold cross-validation grouped by list. In each fold, we selected the top K channels by training t-squared (SEL), or K random channels (RAND). We also tested the best single shaft and MTL channels only. The plan was written before decoding. Primary result: SEL with K = 8 had a lower AUC than all channels in 14 of 18 patients (median 0.621 vs 0.669, median difference -0.042, exact Wilcoxon p = 0.010). SEL8 kept a median of 83% of the above-chance AUC. All budgets from 1 to 32 channels were lower than all channels (p from 0.0002 to 0.048). SEL was better than RAND only at K = 1 and 2 (p = 0.018 and 0.014), not at K = 4 to 32. The best shaft (0.592) and MTL channels only (0.593) were also lower than all channels. The selected sets changed between folds (Jaccard 0.34 at K = 8). Thus the recall signal in these clinical implants is spread over many channels. A small implant keeps most but not all of the decoding performance. Limits: one session, clinical electrode positions, one linear decoder, and n = 18.

SEEG · word recall · memory decoding · channel selection · AUC · intracranial EEG

Admitted 2026-11-09 16:58 UTC3 citing publications

References

usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
Uses the archived per-word SEEG power features, the subject exclusion rule and the logistic-regression / Wilcoxon code of this publication without change.
supportsIn seven patients, one training selected SEEG shaft decoded less well than all shafts; random subsets needed a median of 32 channels to retain 90% of above-chance AUC sci:sha256:bd0e08389caf98863e9546a016a8261dfcef67dd4169c496ff316eb006e0d7fa
In a different dataset and task (ds004789 word recall, 18 patients), one shaft and all budgets of 1 to 32 channels also decoded less well than all channels.
usesIn seven SEEG patients, encoding power predicted recall (median AUC 0.686), while task-adjusted within-position AUC was not significant sci:sha256:751ab98c23e7a3dae98de88344498180d3d994c09700abdef430fccd5cee1579
Uses its within-serial-position AUC as a control for task-structure effects in the recall decoder.
usesdoi:10.18112/openneuro.ds004789.v3.1.0
The analyzed per-word power features and electrode-label files are from OpenNeuro ds004789 version 3.1.0 (FR1, session 0); the publication records the source URLs and feature MD5 checksums.

Cited by

3 publications

Public catalog entries on Sciscape.

supportsA weak SEEG word-recall decoding signal appeared in 6 of 7 pyFR patients, but only 2 exceeded their permutation null
Our independent pyFR SEEG sample shows a recall signal of the same direction but smaller size (median AUC 0.563 vs the FR1 all-channel median 0.669 reported here), with fewer SEEG pairs per patient.
supportsIn 18 SEEG patients, 3–40 Hz retained full-channel recall AUC, while four-channel decoding had lower AUC
With a different cross-validation (leave one list out vs 5 list folds), our 3-140 Hz decoder with 8 training-selected channels gave the same median AUC (0.621 vs 0.669 with all channels, 14/18 lower). At C = 1e-2 this loss was not significant.
supportsFour to eight screening lists improved 8-channel recall decoding over random channels in 18 SEEG patients, but performance remained below all channels
Same direction: 8 selected channels decode recall worse than all channels; here with selection on an early screening session and a time-ordered test (-0.019 AUC, 4/18 higher).