Few-trial SEEG memory-load decoding did not show non-inferiority for learned per-shaft LDA fusion; plain score sum had similar AUC

We tested whether the learned per-shaft two-stage SEEG decoder of sci ecc4dcbb, which was non-inferior to central shrinkage LDA for ds004473 key-press detection, extended to ds004752 verbal Sternberg memory-load decoding with few trials. Using archived log10 Welch band-power features from 15 patients (20–42 clean trials each; set size 8 versus 4; 3–6 s maintenance window), we compared central Ledoit–Wolf LDA, FUSE-STACK (per-shaft LDAs followed by a second LDA on inner out-of-fold scores), and FUSE-SUM (sum of shaft scores) using 10 repeats of stratified 5-fold cross-validation. The pre-registered primary non-inferiority test (HG+beta; margin 0.02 AUC) failed: mean FUSE-STACK minus CENTRAL was −0.022 (SD 0.119), with 7/15 patients higher and exact sign-flip p = 0.53. With all five bands, FUSE-STACK was lower by 0.061 (14/15 patients lower; two-sided p = 0.005), and its shaft-loss result was also worse (−0.058, p = 0.006). FUSE-SUM had similar AUC to CENTRAL (+0.007 for HG+beta and +0.008 for all five bands) while sending 6.4 values per window instead of 89.2 or 223; this secondary endpoint was not tested for non-inferiority. With few trials, the learned fusion stage did not preserve central accuracy in this setting. This does not contradict the ds004473 result. The study had no held-out time split or online test, and secondary tests were not corrected for multiple comparisons.

stereo-EEG · memory load · linear discriminant analysis · score fusion · non-inferiority · cross-validation