Sign-constrained LDA reduced held-out movement-versus-rest AUC across calibration sizes in 15 SEEG patients

We tested a pre-registered hypothesis: an LDA decoder with weights forced to the motor prior (high-gamma 70-150 Hz weight >= 0, beta 13-30 Hz weight <= 0) gives higher held-out AUC than shrinkage LDA when calibration has few trials. Data: archived band-power features from 15 SEEG patients (OpenNeuro ds004473 key press vs rest, n = 8; ds006253 mouse movement vs rest, n = 7). Calibration used 5, 10, 20, 40 trials (20 random subsets) or all train trials; the test block was later in time. Result: the hypothesis was rejected. At N = 10, mean AUC was 0.617 for the sign-constrained LDA and 0.699 for LDA (difference -0.082, 95% CI -0.135 to -0.045, 0 of 15 patients better, exact sign-flip Holm p = 0.0003). The decrease was present at all N (-0.072 to -0.086) and in both datasets. Zeroing wrong-sign LDA weights was worse (-0.089 to -0.122). A fixed +HG/-beta template had AUC 0.443 (below chance). Exploratory analysis: many SEEG channels have the opposite sign (HG decrease in 4 of 7 ds006253 patients outnumbered HG increase; beta increase outnumbered beta decrease in 5 of 8 ds004473 patients). With signs taken from the calibration data, the constrained LDA was equal to LDA within 0.001 AUC. Thus the loss comes from the physiological prior, not from the constraint. Limits: no anatomical channel selection, linear band-power decoders only, binary move vs rest, offline. The package has the plan with its sha256, code, input features with hashes, per-patient results, and a figure.

stereoelectroencephalography · brain-computer interface · linear discriminant analysis · high-gamma power · beta power · movement decoding · calibration · area under the curve