Training set augmentation reduced shrinkage LDA AUC in offline SEEG key press detection

In eight subjects from OpenNeuro ds004473, we compared Ledoit-Wolf shrinkage LDA with Gaussian feature jitter and within-class mixup for offline cued key-press versus rest detection. We trained on the first 10, 20, 40, or 100 trials and tested on trials 101 onward using bipolar log power in the 70–150 Hz and 13–30 Hz bands. The prespecified primary contrast, mixup minus plain LDA at 20 calibration trials, was −0.014 AUC (one of eight subjects improved; exact two-sided sign-flip p = 0.016). Jitter reduced mean AUC by 0.009–0.030 across calibration sizes. Mixup changed mean AUC by +0.000 at 10 trials and reduced it by 0.044 and 0.087 at 40 and 100 trials. All seven secondary contrasts had Holm-adjusted p ≥ 0.055. Exploratory results showed lower move-class shrinkage intensity after augmentation, consistent with augmented copies increasing apparent sample count without adding independent information. This offline result is limited to one session, eight subjects, one dataset, one feature set, fixed augmentation settings, and the tested procedure of augmentation followed by Ledoit-Wolf LDA; augmentation that retains shrinkage fitted on real samples and other decoder families were not tested.

shrinkage LDA · training set augmentation · mixup · Gaussian jitter · SEEG · AUC · OpenNeuro ds004473 · small sample learning