Across 10 ds004789 FR1 SEEG patients, cross-patient recall decoding was weak and not significant within serial position

We compared a calibration-free, cross-patient decoder (XP) with a patient-specific decoder (WP) for predicting later free recall from word-encoding SEEG. The analysis used archived per-word log band power from 10 OpenNeuro ds004789 FR1 patients: seven patients with Morlet features and three with Welch features, harmonized to five bands and z-scored across words within each session. In session 0, XP had a median AUC of 0.547 (9/10 patients above 0.5; one-sided Wilcoxon p=0.0029). Its median AUC within serial position was 0.536 (6/10; p=0.097), so the cross-patient signal was not significant when serial position was held constant. A fixed high-gamma-minus-theta score had median AUC 0.571 (9/10; p=0.0029). WP had median AUC 0.667 overall (9/10; p=0.0020) and 0.619 within serial position (9/10; p=0.014). WP exceeded XP by a median 0.072, but this difference was not significant (two-sided p=0.084). With one labeled list (12 words), WP matched or exceeded XP on later words at the median; no learning-curve comparison was significant. In the three-patient session-1 secondary test, XP AUCs were 0.565, 0.536, and 0.673. These results suggest a small cross-patient signal, with primacy explaining part of the unadjusted association. The analysis is limited by 10 patients, one session per patient in the primary test, different spectral estimators, lack of anatomical matching, low power for the WP–XP comparison, and shrinking test sets at larger training-list counts. The analyzed OpenNeuro release is not established by the archived provenance; source URLs, ETags, and feature-file hashes are retained in the package.

SEEG · free recall · memory decoding · cross-patient decoding · patient-specific calibration · serial position · AUC · memory BCI