Rank-1 SEEG recall decoders did not improve calibration performance over a full L2 decoder
We compared a full channel-by-frequency L2 logistic decoder with rank-1 spatial-by-spectral decoders for offline word-recall decoding from archived SEEG features in OpenNeuro ds004789 v3.1.0 (FR1, session 0). In 17 patients with at least 16 lists, models were trained on the first 1, 2, 3, or 6 lists and tested on lists 13 onward; an 18-patient leave-one-list-out analysis was also performed. The cross-patient spectral-profile model (RANK1X) did not significantly improve median test AUC over the full decoder with 1–3 calibration lists (changes +0.001, +0.005, and −0.013; Holm-adjusted p=1.0). The rank-1 model (RANK1) was never better than the full decoder; its median AUC was lower with 6 lists (change −0.022, p=0.013) and in leave-one-list-out evaluation (−0.022, p=0.018). The full model's leave-one-list-out median AUC was 0.669, matching the archived result. All models were above chance. Spectral profiles varied between patients, especially at 40–140 Hz. In these features and designs, rank-1 structure did not reduce calibration need. This offline study is limited to one dataset and session, fixed regularization, rank 1, a single initialization for the non-convex fit, late test lists, and calibration-list z-scoring based on 12 words at k=1.
SEEG · word recall · memory decoding · brain-computer interface · logistic regression · rank-1 model · calibration · AUC
References
3- usesOne-list word-recall SEEG decoding mostly reflects serial position; three or more lists improve within-position AUC sci:sha256:6666745a5d262454b840ab848c437a2ab1e6d380dd1f88f42206228636891a04
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.
- usesCross-patient SEEG recall decoding was above chance but weaker than a three-list within-patient decoder sci:sha256:d52ba56c23a69304d837e4ef5a9f3f20bcee59ebf2cd093faea38a0334a9fa4a
This work uses the archived per-word SEEG log-power features, the logistic regression, AUC and exact Wilcoxon code, the patient inclusion rule and C = 7.2e-4 from this publication. The FULL LOLO AUC (0.669) reproduces its value.
- usesdoi:10.18112/openneuro.ds004789.v3.1.0
The archived per-word features analyzed here are identified as OpenNeuro ds004789 v3.1.0, FR1 session 0.
Cited by
2 publicationsPublic catalog entries on Sciscape.
- supportsA rank-1 SEEG recall decoder projection reduced held-out AUC in 18 patients
An independent rank-1 method (SVD truncation of the full weights, and rank 2) also found no gain over the full decoder, and a significant decrease at 3 lists and with LOLO.
- supportsRank-1 channel-by-band LDA had lower AUC than full shrinkage LDA in 15 SEEG patients
As in FR1 with 1-3 calibration lists, a rank-1 spatial x spectral decoder did not improve on the full decoder at small calibration sizes in ds004752.