One-list word-recall SEEG decoding mostly reflects serial position; three or more lists improve within-position AUC
This study measured how patient-specific recall decoding changes with 1, 2, 3, 4, 6, 8, or 12 calibration lists in 17 eligible patients performing free recall (FR1) with stereo-EEG. Archived per-word log10 band-power features (0–1.6 s after word onset; 3–140 Hz) came from OpenNeuro ds004789 v3.1.0. Training used lists 1 through k and testing used lists 13 through the end. A patient-only L2 logistic decoder had mean later-list AUC 0.589 after one list and 0.647 after 12 lists. The one-list decoder was not above chance when AUC was evaluated within serial position (0.538, p = 0.085), indicating that most of its overall AUC arose from serial-position effects. Within-position AUC was above chance from three lists onward (0.574–0.590, p ≤ 0.002), although its gain over the training-free spectral-tilt score was not significant at any k. The tilt score stayed near 0.54 overall and was not significantly different from chance. The patient-only decoder exceeded tilt in 12–13 of 17 patients at each k; after Holm correction, the difference was significant only at 8 and 12 lists. A hybrid that added tilt as an almost unpenalized feature did not significantly improve on the patient-only decoder and had lower mean AUC with 1–6 lists. The result is limited to one cohort, task, and fixed feature set. The regularization value was not tuned, and testing on later lists includes possible session drift.
stereo-EEG · free recall · memory BCI · calibration · learning curve · AUC · spectral tilt · serial position · logistic regression
References
3- supportsIn 17 SEEG patients, calibration on the first eight lists showed no significant AUC loss versus time-blind training in one session sci:sha256:ff7fa7989e64afee10d3a18f0e91ebbc2520df87c6c5752ce35f51c5db5942f6
Our fixed first-8-list training gives mean AUC 0.631 on lists 13 to the end. This agrees with the 0.626 that this publication reported for first-8-list training.
- usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
All per-word band-power features, the inclusion rule, the W_chan penalty C = 7.2e-4 and the T_tilt definition come from this package.
- usesdoi:10.18112/openneuro.ds004789.v3.1.0
The archived word-level features used here were derived from FR1 session 0 of OpenNeuro ds004789 v3.1.0. The source feature package and EDF provenance are documented in the original files.
Cited by
7 publicationsPublic catalog entries on Sciscape.
- supportsDecoder selection explained most of the lower recall of stimulated words in RAM closed-loop memory experiments
In the deployed RAM closed-loop classifier, the stimulation rate depends strongly on serial position and the position-adjusted decoder separation is smaller and not significant. This agrees with its finding that much of a small-data recall decoder's signal comes from serial position.
- usesRank-1 SEEG recall decoders did not improve calibration performance over a full L2 decoder
This work uses the short-calibration design of this publication: train on the first k lists, z-score with calibration statistics, test on lists 13 to the end.
- usesAn adaptive threshold reduces stimulation-rate error in an offline memory-decoder simulation
We use a calibration of 8 and 12 lists, in the range of calibration lengths this work examined.
- usesFixed L2 penalty outperforms per-patient tuning for SEEG recall decoding in 18 FR1 patients
That work lists the untuned penalty C as a limitation; this work tests per-patient tuning of C on the same type of decoder.
- usesIn 18 SEEG patients, an equal-weight cross-patient ensemble did not improve few-list word-recall decoding
Extends its patient-only learning curve with a cross-patient ensemble at k = 1, 2, 4 lists.
- supportsA cross-patient prior did not improve recall decoding after 1–3 calibration lists in 18 SEEG patients
Our within-patient decoder trained on 1 to 6 lists gave median later-list AUC 0.588 to 0.633, in the 0.59 to 0.65 range it reported, with a different test split (fixed lists 7 to end).
- A fixed cross-patient recall prior did not improve patient SEEG decoding with 1–3 labeled lists
[Legacy relation: reproduces] The patient-only decoder trained on 1 list gives mean held-out AUC 0.589 in our 17 targets, the same value as that work.