An FR1-trained SEEG recall model predicted catFR1 recall after serial-position removal in 14 independent patients

We tested whether a cross-patient SEEG word-recall model trained on FR1 (uncategorized word lists) predicted recall in catFR1 (categorized lists). The model used 30 features: log10 power in five frequency bands across six region groups, from 0–1.6 s after word onset in session 0. Among 14 eligible catFR1 patients absent from the FR1 training data, transfer AUC was 0.566 (95% CI 0.538–0.592; 13/14 above 0.5; H1 Holm p = 0.0033). After removing each patient's mean serial-position effect from training and test features, AUC was 0.551 (0.531–0.571; 12/14; H3 Holm p = 0.0033). Transfer did not differ significantly from a catFR1 leave-one-subject-out model (difference −0.012, 95% CI −0.028 to 0.007; p = 0.23); this does not establish equality. In all 31 eligible catFR1 patients, transfer AUC was 0.572 (0.550–0.593), and the serial-position-controlled AUC was 0.556 (0.537–0.575). A model using serial position alone had AUC 0.568 in the independent patients, so the overall transfer effect was small and much of its raw value may reflect position. Only 14 patients were fully independent of FR1, and the control removed only each patient's mean position effect. The analysis used archived feature files; the exact OpenNeuro release underlying those files was not recorded.

SEEG · word recall · memory · cross-patient prediction · transfer learning · serial position · AUC · catFR1 · FR1