A fixed SEEG spectral index predicts opposite outcomes in free recall and load-3 working-memory encoding

We tested whether a fixed high-frequency-minus-low-frequency SEEG index that predicts later word recall also predicts success in load-3 object working memory. The index was defined before analysis as the mean across contacts of z-scored high-frequency power minus low-frequency power, using 80 and 140 Hz minus 3, 5, and 8 Hz in FR1 free recall and high gamma minus theta in working memory. Archived features were used for 18 FR1 patients and 12 working-memory patients. The prediction of the same sign failed: FR1 median per-patient AUC for later recall was 0.569 (14/18 above 0.5; exact Wilcoxon p = 0.000839), whereas working-memory encoding median AUC for 3-of-3 correct was 0.429 (3/12 above 0.5; p = 0.0425). The mean AUC difference was 0.127 (permutation p = 0.00012). The working-memory fixation AUC was 0.492 (p = 0.791); encoding AUC was lower than fixation in 11/12 patients (p = 0.00098). Highest-index encoding trials had a median error rate of 0.45, compared with 0.32 in the lowest-index tertile (p = 0.020). Per-band profiles were approximately mirror images across tasks. The datasets differ in patients, hospitals, electrode coverage, success labels, and feature pipelines, so task effects cannot be separated from cohort and pipeline effects. The index averaged all contacts, working-memory results include 12 patients, only load 3 was tested, and secondary tests were not corrected. The results show that a fixed sign derived from free recall can have the wrong sign for working-memory encoding; the sign and weights require task-specific calibration.

SEEG · intracranial EEG · memory · free recall · working memory · spectral power · brain-computer interface · AUC