Shrinkage QDA had lower held-out AUC than shrinkage LDA in all 8 SEEG subjects using trials 1–100 for training and later trials for testing

We compared shrinkage QDA (one Ledoit-Wolf covariance for each class) with shrinkage LDA (pooled covariance) for key-press vs rest detection. We used the archived ds004473 features of sci 04683a2a (8 subjects, 100-112 bipolar channels, high-gamma and beta log power). The decoders were trained on trials 1-100 and tested on trials 101 to the end. The plan was hashed before the analysis. QDA gave a lower AUC than LDA in all 8 subjects (mean difference -0.074, exact two-sided sign-flip p = 0.0078). The loss was smaller for a 50/50 mix of class and pooled covariance (-0.028) and for 8 selected channels (-0.013, p = 0.063). Move and rest windows had almost the same total high-gamma variance (median ratio 0.96). Thus, at this calibration size, class-specific covariance added estimation error and no useful information. Limits: one dataset, one session, offline, cued movement, Ledoit-Wolf shrinkage only.

shrinkage QDA · shrinkage LDA · intracranial EEG · SEEG · brain-computer interface · motor decoding · held-out AUC · OpenNeuro ds004473