SEEG recall decoding adds to word memorability and serial position in 18 patients

We asked whether a stereoelectroencephalography (SEEG) word-recall decoder follows item memorability, and whether neural features add information beyond word memorability and serial position. We estimated word recall rates from 120,466 presentations in 255 subjects across 518 sessions in OpenNeuro ds004789 v3.1.0, excluding the 18 analysis patients. In those 18 patients, memorability alone gave median AUC 0.575; memorability plus serial position gave 0.634. The neural W_chan decoder gave 0.669, which was not significantly higher than the no-brain model (difference +0.019, 10/18 positive, two-sided p = 0.77). Neural score and word memorability were nearly uncorrelated (median Spearman rho +0.021, p = 0.14), and neural AUC within memorability quintiles was 0.663. Combining neural features with the fixed no-brain score increased median AUC to 0.692 (gain +0.037 over the no-brain model, 17/18 positive, one-sided p = 9.5e-5). Results were unchanged when the one patient using a Spanish word pool was excluded. These findings indicate that, for session 0 and the analyzed 0–1.6 s band-power features, the decoder output was largely independent of item memorability and neural and item/position information were complementary. The study includes 18 patients and task-specific norms; time-resolved features or a patient-fitted offset could change the results.

SEEG · word recall · memorability · memory decoding · brain-computer interface · serial position · AUC · electrophysiology