Low-artifact scalp EEG decodes verbal working-memory load above chance, while SEEG scores higher

In 15 patients performing a verbal working-memory task, scalp EEG decoded set size above chance after removing frontal channels (mean AUC 0.624, Holm p = 0.038), temporal and ear channels (0.619, Holm p = 0.037), or using only central, parietal, and occipital channels at 4–30 Hz (0.599, Holm p = 0.037). The core 4–13 Hz result was not significant (AUC 0.567, p = 0.155). Core 4–70 Hz decoding was 0.066 AUC below the full montage (p = 0.022). On the same trials, bipolar SEEG exceeded core scalp 4–70 Hz by 0.100 AUC (13/15 patients, p = 0.030); its advantage over core scalp 4–30 Hz was 0.094 (p = 0.056). These results show that some scalp load information remains away from the eyes and temporalis muscle, while the earlier “no SEEG advantage” result depends on artifact-prone scalp channels. The analysis cannot distinguish brain activity from artifact in frontal and temporal channels. Features were average-referenced before channel removal, no EOG/EMG or ICA was used, channel removal also reduces feature count in five patients, and the analysis used one session with within-session cross-validation.

scalp EEG · SEEG · working memory · brain-computer interface · artifact ablation · AUC