In 18 SEEG patients, an equal-weight cross-patient ensemble did not improve few-list word-recall decoding

We tested a pre-registered question: does a cross-patient prior reduce the calibration need of an SEEG memory decoder? In OpenNeuro ds004789 FR1 session 0, we evaluated 18 patients using encoding-period power features. W_k was a patient decoder trained on the first k lists; X_cross was trained on the other 17 patients without data from the test patient; H_k was the equal-weight mean of their z-scored scores. The held-out test set was lists 5 to the end for each model. H_k did not improve on W_k at k = 1, 2, or 4: median AUC differences were -0.004 (n = 16), -0.005 (n = 18), and -0.018 (n = 18), with Holm-adjusted exact Wilcoxon p values of 1.00, 1.00, and 0.74. H_k did outperform X_cross at all three k values (median differences +0.023 to +0.043; Holm p <= 0.002). Exploratory results found that W_k alone outperformed X_cross reliably only at k = 4. The ensemble was better than X_cross in more patients and helped some cases where W_k was worse, but it did not improve median AUC over W_k. Median AUCs were low (0.54–0.62). This offline analysis used one dataset, one session, 18 patients, and features and penalties from a prior internal publication. Full-session, label-free score standardization included test words; online use would require running estimates. The results do not establish stimulation or closed-loop utility.

SEEG · word recall · memory decoding · brain-computer interface · cross-patient decoding · ensemble · AUC · calibration · free recall