Z-scored shaft scores reduced AUC, and no decoder improved on LDA, in 15 SEEG patients

In an offline analysis of one session from each of 15 patients, we compared Ledoit-Wolf shrinkage linear discriminant analysis (LDA), four other decoder families, and ways to combine per-shaft LDA scores for movement-versus-rest detection from 70–150 Hz high-gamma activity (HGA) window means. Trials were split in time order: the first 60% trained each decoder and the last 40% tested it. No alternative decoder was significantly better than all-channel LDA (mean AUC 0.833). L2 logistic regression and the pre-registered equal-weight mean of z-scored per-shaft scores performed worse; the latter reduced AUC by 0.057 (95% CI −0.084 to −0.031; 2/15 patients improved; Holm p = 0.004). In an exploratory analysis, summing raw per-shaft decision values gave a difference of −0.004 (95% CI −0.014 to +0.006), while z-scored scores weighted by training d′ gave −0.008. The result suggests that retaining or restoring score scale matters for this combination rule. Stimulus-direction decoding was near chance, and its SVM result was not significant after correction. This was an offline, single-session, single-center analysis with clinical electrode placement, one feature per channel, one time-ordered split, and fixed decoder settings; the combiner analysis was exploratory.

SEEG · high-gamma activity · movement detection · linear discriminant analysis · decoder scores · score combination · AUC · brain-computer interface