Word memorability predicted recall but did not reliably improve a band-power decoder in 17 FR1 SEEG patients

We tested word memorability as a brain-independent predictor of later recall and as a confound of an SEEG encoding decoder. We used band-power features of 17 FR1 patients (ds004789, session 0). Memorability came from the recall rate of each word in the other patients. Memorability predicted recall in 17 of 17 patients (median AUC 0.584). The leave-one-list-out SEEG decoder had a median AUC of 0.670. A stacked model with SEEG and memorability did not reliably increase the AUC (median gain +0.002, 9 of 17 patients, exact Wilcoxon p=0.58). When we removed memorability from the SEEG score, the AUC decreased by only 0.002. The SEEG score and memorability had a small correlation (rho 0.054) that was not significant within recalled or not-recalled words. Thus the decoder does not read word difficulty, and memory decoder studies should report the item-only baseline. The package contains code, per-patient results, the figure and SHA-256 hashes of the input features.

word memorability · free recall · stereo-EEG · SEEG · memory decoding · band power · AUC · brain-computer interface