Rank-1 channel-by-band LDA had lower AUC than full shrinkage LDA in 15 SEEG patients

In an offline analysis of archived log10 Welch band-power features from 15 patients performing a verbal Sternberg task, we compared full shrinkage LDA with a rank-1 channel-by-band decoder for classifying set size 8 versus 4. Features covered five fixed bands from 4–150 Hz during the 3–6 s maintenance window; patients had 20–42 valid trials. With 3, 5, or 8 training trials per class, rank-1 LDA had lower mean held-out AUC than full LDA by 0.029, 0.034, and 0.039, respectively; in repeated 5-fold cross-validation the difference was 0.045. Exact sign-flip p values were 0.001, 0.001, 0.060, and 0.030. A rank-1 decoder using the mean band profile from the other 14 patients was not significantly different from full LDA (mean differences −0.024 to −0.043; p ≥ 0.15), with patient effects from −0.24 to +0.14. Full LDA reached mean AUC 0.625 with 3 trials per class. The 70–150 Hz profile weight was positive in all patients, while the mean pairwise cosine between patient profiles was 0.16. These offline results cover one task, one session, 15 patients, fixed bands, and LDA; they do not establish performance for other tasks, sessions, or decoder families.

stereo-EEG · working memory · Sternberg task · rank-1 decoder · shrinkage LDA · band power · AUC · calibration trials