In 15 SEEG subjects, 10–20 principal components preserved working-memory load decoding, but 1–2 components and channel means reduced it

This archived analysis asked whether SEEG working-memory load decoding could use channel-mean band power, a few principal components, or one selected channel instead of the full channel-by-band feature set. It used archived log10 Welch band-power features from 15 subjects performing a verbal Sternberg task, with 464 valid trials for set sizes 4 and 8. A shrinkage LDA decoder was evaluated with stratified 5-fold cross-validation repeated 20 times. The full feature set had mean AUC 0.704, with 14/15 subjects above 0.5. Channel-mean power reduced mean AUC by 0.105 (14/15 subjects lower; Holm-adjusted p = 0.012). The first and second unsupervised principal components reduced mean AUC by 0.155 and 0.122 (Holm-adjusted p = 0.002 and 0.020); 10 and 20 components had mean AUC 0.683 and 0.698, with no significant difference from the full decoder. Selecting one channel in each training fold reduced mean AUC by 0.070 (12/15 lower; Holm-adjusted p = 0.020), and channel selection varied across folds. These results indicate distributed decodable information that is not captured by global mean power or the dominant variance axes, while an unsupervised 10–20 dimensional subspace retained nearly all measured decoding performance. The evidence is limited to one task, one session, 15 clinical SEEG subjects, 20–42 trials per subject, fixed-window log band power, and shrinkage LDA. Small trial counts may favor low-dimensional decoders; this analysis did not test whether more trials increase the gap.

SEEG · working memory · brain-computer interface · dimensionality reduction · principal component analysis · shrinkage LDA · band power · AUC · verbal Sternberg task