In 18 epilepsy patients, a 5 + 140 Hz decoder had 0.045 lower median AUC than an 8-frequency SEEG word-recall decoder

We tested how many spectral frequencies a within-patient decoder of later word recall needs. Fewer frequencies mean fewer filters in an implanted memory BCI. Data were archived log10 power at 8 frequencies (3–140 Hz), 0–1.6 s after word onset, across all bipolar channels in 18 patients (internal SCI contribution c4b8c884). We used L2 logistic regression with fixed C and leave-one-list-out cross-validation, and calculated AUC per patient. In the pre-registered primary test, the 2-frequency decoder (5 Hz + 140 Hz) was not non-inferior to the 8-frequency decoder, with a margin of 0.02 AUC. The median difference was −0.045 (95% patient-bootstrap CI −0.059 to −0.010); 14 of 18 patients had lower AUC. The exact Wilcoxon p-value was 0.0008. A 10-fold weaker penalty gave the same conclusion. Each single frequency alone was worse than all 8 (Holm p < 0.03). The best single frequencies were 13, 80, and 8 Hz; 140 Hz was the worst. Removing one frequency changed median AUC by 0.006 or less; only removal of 8 Hz gave a significant loss. These results indicate that recall information in these features was spread across multiple frequencies and redundant between neighbouring frequencies. The analysis used session 0 of FR1 during encoding in epilepsy patients, a linear decoder, and a fixed penalty tuned on the same patients. Other small frequency sets were not tested.

SEEG · word recall · intracranial EEG · spectral power · frequency ablation · logistic regression · memory BCI · AUC · OpenNeuro ds004789