A per-shaft stacked LDA was non-inferior to pooled LDA for cued key-press detection in 8 SEEG subjects

We tested whether using electrode-shaft structure improves offline cued key-press detection from SEEG. Archived features from OpenNeuro ds004473 v1.0.2 covered 8 subjects and 70–150 Hz and 13–30 Hz log power from adjacent bipolar pairs. We trained on trials 1–100 (200 windows) and tested on trials 101 onward. The primary decoder used one Ledoit–Wolf shrinkage LDA per shaft and a second LDA trained on out-of-fold shaft scores; the comparator was one pooled shrinkage LDA. The primary hypothesis of higher AUC was not supported: mean difference was +0.007 (median +0.004), 5 of 8 subjects, exact two-sided sign-flip p = 0.2266. The stacked decoder was non-inferior within 0.02 AUC (one-sided p = 0.0039). With 70–150 Hz only, the difference was −0.008 (p = 0.5625), and non-inferiority was not shown (p = 0.1797). Equal-weight shaft voting was worse than pooled LDA in all 8 subjects (mean −0.064, p = 0.0078); the best single shaft was worse in 7 of 8 (mean −0.039, p = 0.0391). Secondary tests were not corrected. Exploratory weights showed that the largest level-two weight accounted for 17–46% of total absolute weight and the effective number of shafts was 3.8–9.2. Ledoit–Wolf shrinkage already controlled the feature count in this task; shaft structure gave no clear accuracy gain, and fusion weights needed to be learned. This was an offline analysis of one cued task and session, 8 epilepsy patients with clinical electrode placement, one split, and linear decoders. The small sample gave low power (minimum possible two-sided p = 0.0078).

stereo-EEG · SEEG · key-press detection · linear discriminant analysis · shrinkage LDA · electrode shafts · stacking · AUC · OpenNeuro ds004473