In 18 SEEG patients, encoding-power PC1 did not predict word recall and underperformed a fixed frequency tilt

We tested whether the first principal component (PC1) of each patient's z-scored channel-by-frequency SEEG band power during word encoding could predict later free recall without labels. In 18 patients from OpenNeuro ds004789 FR1, session 0, PC1 had median AUC 0.485 (IQR 0.439–0.561); 6 of 18 patients were above 0.5 (exact one-sided Wilcoxon p = 0.63). PC1 also did not predict recall within serial position (median AUC 0.475) or when fitted on unlabeled lists 1–3 and tested on later lists (0.507; Holm p = 0.39). The fixed high-minus-low frequency tilt readout had median AUC 0.569 (14/18 above 0.5, p = 0.0004), and a labelled leave-one-list-out decoder had median AUC 0.669. PC1 explained a median 10.9% of feature variance and correlated with the global feature mean (median r = 0.83). Exploratory results found higher PC1 AUC among eight patients whose PC1 had a tilt-like spectral shape than among ten broadband patients (0.543 vs 0.456); this analysis was not pre-specified. The findings concern one offline task, one dataset, 18 patients, and PCA with one pre-specified sign rule; they do not exclude a small positive PC1 effect or establish performance in other settings.

SEEG · word recall · free recall · brain-computer interface · principal component analysis · band power · frequency tilt · label-free decoding · OpenNeuro ds004789 · FR1