An FR1-trained SEEG word-recall decoder transfers above chance to catFR1 in 8 patients
In eight patients with session-0 SEEG data, an L2 logistic decoder trained on uncategorized free recall (FR1) ranked later-recalled words above non-recalled words in categorized free recall (catFR1): median AUC 0.585 (range 0.569–0.776; exact Wilcoxon p = 0.0039), with all eight patient AUCs above 0.5. Transfer AUC did not differ significantly from leave-one-list-out catFR1 AUC (median difference −0.001, p = 0.95); this failure to reject is not evidence of equivalence. AUC within serial positions remained above chance (median 0.578). A fixed decision threshold did not transfer reliably without session-specific, label-free recentering. Calibration using up to five catFR1 lists alone stayed near chance. The task and session changed together, so their effects cannot be separated. This was an offline analysis of eight patients, one feature set, and one untuned decoder; no stimulation or closed-loop test was performed.
SEEG · memory decoding · free recall · brain-computer interface · cross-task transfer · intracranial EEG
References
3- usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
We use the FR1 session-0 features, the feature extraction code and the fixed decoder settings (C = 7.2e-4) of this publication without change, and we extend them to catFR1.
- usesdoi:10.18112/openneuro.ds004789.v3.1.0
FR1 session-0 features and source data.
- usesdoi:10.18112/openneuro.ds004809.v2.2.0
catFR1 dataset version recorded for this analysis; the source describes reading current S3 objects and records ETags because an exact snapshot match is uncertain.
Cited by
4 publicationsPublic catalog entries on Sciscape.
- supportsSEEG band power did not decode word category above a structure-preserving null in 8 catFR1 patients
The positive control, a different within-patient leave-one-list-out pipeline, also decodes catFR1 recall above chance (median AUC 0.59, Wilcoxon p = 0.02 over the same 8 patients).
- usesSEEG band power did not decode word category above a structure-preserving null in 8 catFR1 patients
Uses the archived per-word catFR1 band-power features (features_catfr1, with checksums) and their extraction code.
- challengesThe FR1 recall decoder was above chance in 3 of 8 pyFR SEEG sessions, with no significant group effect
Recall decoding generalization: decoder is weaker in a different cohort (pyFR) than within FR1/catFR1 patients.
- usesThe encoding decoder score predicted current word recall but not next-word recall in catFR1
Source of the archived catFR1 per-word power features, recall labels and events.tsv with category labels for 8 patients.
- challengesIn six patients, free-recall decoding transferred weakly to paired-associate recall and cross-patient spectral tilt reversed
FR1-to-catFR1 transfer worked there. Here FR1-to-PAL1 transfer was weaker and not significant (p = 0.078, n = 6), and the cross-patient spectral-tilt model had the wrong sign. Thus the transfer is not general to associative cued recall.