SEEG band power did not predict later free recall above chance in five RepFR1 patients

We tested whether stereo-electroencephalography (SEEG) band power after word onset predicted later free recall in RepFR1, using session 0 from five patients and lists 0–9 (270 word presentations per patient). Words appeared one, two, or three times. A shrinkage linear discriminant analysis decoder used log band power in five frequency bands from 0–0.75 s after onset, with five-fold cross-validation grouped by list and serial position removed using training-fold fits. Mean held-out area under the receiver operating characteristic curve (AUC) was 0.598 for first presentations of one-time words (4/5 patients above 0.5; one-sided sign-flip p = 0.16), 0.519 for first presentations of three-time words (p = 0.34), 0.456 for the mean score across three presentations (p = 0.91), and 0.501 across all presentations. The mean of three scores did not improve on the first score for three-time words (difference −0.063 AUC, two-sided p = 0.19). Mean recall rates were 0.23, 0.28, and 0.41 for one-, two-, and three-time words. Thus, in this small sample, this band-power readout did not predict later recall above chance at the group level. Each patient contributed only 30 one-time words; the study had low power for AUC below 0.6. The analysis used one feature set and one decoder, had no electrode localization or region selection, and was offline. The package includes the archived features, code, results, source hashes, and figure. The EDF and event files were fetched from live OpenNeuro S3 URLs; the exact immutable dataset release used was not recorded.

stereo-electroencephalography · free recall · memory encoding · band power · linear discriminant analysis · repeated words · memory BCI · RepFR1