Self-training did not improve SEEG word-recall decoding with 1–5 labeled lists in 18 patients

We asked whether self-training with unlabeled words improves within-patient SEEG word-recall decoding when only 1, 2, 3, or 5 labeled lists are available. We analyzed archived band-power features from 18 patients performing delayed free recall (OpenNeuro ds004789, FR1, session 0). The primary comparison at 3 labeled lists found median test AUC 0.617 for self-training and 0.631 for supervised logistic regression; the median per-patient difference was −0.007, self-training was higher in 8/18 patients, and the exact two-sided Wilcoxon p-value was 0.265. Differences were also not significant after Holm correction at 1, 2, or 5 lists. A disjoint unlabeled pool and unlabeled PCA did not improve performance. Final-round pseudo-label accuracy was 0.57–0.59 (chance 0.5). These results suggest that confidence-based self-training did not help this weak decoder in the tested setting. The study used one dataset, task, session, linear band-power decoders, and one self-training parameter setting; all AUCs were offline.

semi-supervised learning · self-training · SEEG · word recall · memory BCI · AUC · pseudo-labels