Equal-weight scalp EEG and SEEG fusion did not significantly improve load decoding in 15 patients

We compared shrinkage-LDA decoders of verbal working-memory load (set size 8 vs 4) from scalp EEG and SEEG in 15 patients from OpenNeuro ds004752 session 1. Features were band power from the 3–6 s maintenance window. Both decoders used the same trials and stratified five-fold cross-validation repeated 20 times. Equal-weight late fusion had mean AUC 0.731, compared with 0.709 for SEEG alone (mean difference +0.022; 10/15 patients higher; 95% bootstrap CI −0.013 to +0.060; exact sign-flip p = 0.29). Fusion exceeded scalp alone by 0.068 AUC (p = 0.023) and was similar to feature concatenation (difference −0.007, p = 0.52). The mean within-class correlation between out-of-fold scores was r = 0.13 (95% CI 0.05 to 0.22; p = 0.008). The results indicate low observed error correlation, but the equal-weight fusion gain depended on scalp decoder performance and was negative in four patients with scalp AUC below 0.5. Accuracy-based weights may help, but were not tested. This offline study is limited to one task, one contrast, band-power features, 15 epilepsy patients, and 20–42 trials per patient; its primary confidence interval includes gains up to +0.06 AUC.

scalp EEG · SEEG · late fusion · working memory · shrinkage LDA · AUC · brain-computer interface · within-class correlation