Unlabeled task covariance lowers few-trial SEEG movement-detection AUC, while rest-only covariance does not

We tested a semi-supervised shrinkage LDA for movement detection from SEEG in 8 subjects. The detector separates a rest window from a movement window in each cued key-press trial. Method B estimated covariance from all unlabeled windows in the first 100 trials and estimated the class-mean difference from n labeled trials. Standard method A estimated both from labeled windows. Testing used all trials after trial 100. At the primary endpoint (n = 10), B had lower AUC than A in all 8 subjects (mean difference -0.104; exact one-sided sign-flip p = 1.00). B had lower AUC at all tested n from 5 to 100 in both random and time-ordered draws. A secondary method C used covariance from unlabeled rest windows only. It showed a small, uncorrected gain at n = 5 (+0.020 AUC; 7 of 8 subjects; p = 0.008), which decreased to 0 at n = 40 and was not consistent in time-ordered draws. The C result is descriptive. These findings apply to one cued key-press task and 8 subjects. They indicate that, for this LDA setup, covariance from unlabeled task windows that contain both classes can reduce the movement signal relative to noise in an estimated class-mean difference. Other semi-supervised methods were not tested.

SEEG · shrinkage LDA · movement detection · BCI calibration · semi-supervised learning · high gamma · beta power · AUC