In 8 SEEG subjects, PCA-16 reduced mean AUC by 0.013, but non-inferiority to full features was not shown

Question: can a label-free linear compression of SEEG motor band power keep the decoding AUC of the full feature set? This is important for implanted BCIs that must send few numbers per frame. Data: archived, hash-checked features of sci 04683a2a (OpenNeuro ds004473, 8 subjects, cued key press; rest and move windows; log 70-150 Hz and 13-30 Hz power on 100-112 bipolar channels, 200-224 features). Method (plan hashed before analysis): train on trials 1-100 and test on trials 101+. We fit PCA to the pooled calibration windows without labels, kept k = 1-64 components, and trained shrinkage LDA on them. We compared this with the full feature set and with Gaussian random projections (20 matrices). We used exact sign-flip tests over subjects. Results: the full set gave a mean test AUC of 0.828. PCA-16 gave 0.816 (mean change -0.013, median -0.018). The primary non-inferiority test at a 0.02 margin was not significant (p = 0.22; 4 of 8 subjects above -0.02). PCA-64 gave the same AUC as the full set (+0.001). With 1-4 components, AUC decreased in 8 of 8 subjects (mean -0.09 to -0.22; p_Holm = 0.055, the minimum possible for this test). PCA was better than random projection at k = 2-32 (mean +0.08 to +0.15; 8 of 8 subjects for k = 4-32; p_Holm = 0.020). Exploratory: the first 16 PCs held 46% of the variance but 82% of the squared move-rest mean difference. Conclusion: a fitted label-free PCA can cut the decoder input to 16-64 numbers with a small mean loss, but 8 subjects do not show that the loss is below 0.02. A random projection without fitting loses much more. Limits: offline, one session, windows aligned to the known response, linear decoder only, and a balanced rest/move pool for PCA.

SEEG · brain-computer interface · principal component analysis · linear discriminant analysis · dimensionality reduction · motor decoding · AUC · random projection