Heavy tailed SEEG band power did not improve robust shrinkage LDA movement decoding

A preregistered analysis tested whether robust linear discriminant analysis (LDA) improves movement decoding from real stereoelectroencephalography (SEEG) band power. We compared Ledoit-Wolf shrinkage LDA with Huber-weighted and winsorized variants on archived features from 8 participants in OpenNeuro ds004473 (cued key press vs rest) and 15 patients in ds006253 (mouse movement vs pre-stimulus rest), using time-ordered training and test splits. Within-class training features were heavy tailed: values with |z| > 4 occurred at 5–100 times the Gaussian rate, and median excess kurtosis reached 10.9 per patient in ds006253. Neither robust method improved test AUC over standard LDA. Mean paired AUC differences for Huber weighting were −0.0002 (ds004473; Holm p = 0.44) and +0.0004 (ds006253; Holm p = 1.00); for winsorizing they were −0.0007 (Holm p = 1.00) and +0.0023 (Holm p = 0.64). All four 95% confidence intervals were within ±0.006 AUC, a descriptive equivalence result within the preregistered ±0.02 threshold. The findings indicate that, in these offline archived-feature analyses, the rare extreme values did not materially change shrinkage-LDA movement decoding. Limits include one session per participant, no online test, no separation of artifacts from physiological extremes, a fixed Huber rule, and low power for small effects in the 8-participant dataset.

stereoelectroencephalography · SEEG · band power · heavy tails · linear discriminant analysis · movement decoding · robust statistics · winsorizing