A two-stage cascade did not improve three-class SEEG balanced accuracy over flat LDA in 15 patients

In 15 patients performing a mouse motion-decision task, we compared a two-stage movement-versus-rest then left-versus-right shrinkage LDA cascade with a flat three-class shrinkage LDA. We used archived 70–150 Hz high-gamma features, offline event-locked 0.5 s windows, one session per patient, and a time-ordered 60/40 calibration/test split. Mean held-out balanced accuracy was 0.518 for flat LDA and 0.515 for the cascade (chance 0.333). The cascade-minus-flat mean difference was −0.003 (descriptive 95% t CI −0.016 to +0.009; cascade better in 7/15 patients; exact two-sided sign-flip p = 0.58, Holm p = 0.58). Both decoders exceeded chance (Holm p = 9.2 × 10⁻⁵ each). Stage-1 movement-versus-rest AUC was 0.829; stage-2 left-versus-right AUC was 0.546, indicating weak side information that limited both decoders. The cascade had higher REST recall (0.779 versus 0.744) and lower recall for both movement sides. This offline, within-patient result does not test continuous online control. The descriptive confidence interval was not preregistered, and non-significance does not establish equivalence.

SEEG · brain-computer interface · high-gamma activity · shrinkage LDA · balanced accuracy · movement decoding · OpenNeuro ds006253