Univariate channel ranking did not outperform LDA weights at four SEEG channels in 15 patients

We compared three rules for ranking bipolar SEEG channels from small calibration sets: the sum of squared Welch t statistics across five bands (U), the L2 norm of shrinkage-LDA weights (W), and the L2 norm of Haufe activation patterns (P). Random channel selection (R) was the control. We used archived log10 Welch band-power features from 15 patients performing a verbal working-memory task, comparing set sizes 8 and 4. For each patient, 50 stratified half splits were run in both directions. Calibration sets contained 10–21 trials; shrinkage LDA was trained on one half using 1, 2, 4, or 8 selected channels and evaluated by AUC on the other half. The pre-specified primary contrast, U minus W at four channels, was +0.016 (95% CI −0.002 to 0.034; 9/15 patients; exact sign-flip p = 0.078), so the analysis did not show a significant difference. At four channels, U exceeded random selection by +0.054 (95% CI 0.019 to 0.088; 13/15; Holm-adjusted p = 0.006). W versus random had an adjusted p = 0.060, and P versus W had p = 0.363. Exploratory contrasts favored U over W at one and two channels, but not at eight. No selected subset reached the mean all-channel AUC of 0.689. Split-half Spearman reliability was low for all ranking rules (U 0.080, W 0.117, P 0.070). The results are limited to offline data from one session, one task contrast, 15 patients, and small calibration sets; the 50 splits per patient were not treated as independent group observations.

SEEG · intracranial EEG · channel selection · feature ranking · brain-computer interface · shrinkage LDA · working memory · AUC · calibration data