Self-training did not improve held-out SEEG key-press decoding in an 8-subject offline test

In an offline analysis of archived high-gamma and beta features from eight SEEG subjects performing a cued key-press task, we tested whether label-free self-training improved a shrinkage-LDA detector. The pre-registered primary comparison used 10 labeled trials, unlabeled trials 11–100, and held-out trials 101 onward. Self-training changed mean test AUC by −0.009 (4/8 subjects improved; exact two-sided sign-flip p = 0.570). No secondary comparison was significant before or after Holm correction; the largest mean secondary gain was +0.052 for confidence-restricted self-training with five labeled trials (p = 0.156; Holm-adjusted p = 0.781). Mean pseudo-label accuracy was 0.71 for the primary method and 0.81 for confidence-restricted self-training. In the reference condition, fitting with true labels for trials 1–100 rather than five or ten labeled trials increased AUC by +0.164 or +0.091, respectively, in all eight subjects. This indicates that the unlabeled pool contained useful information, while the tested self-training procedure did not recover that gain. The study was offline, used one session, eight subjects, a known 50/50 task balance, one archived feature set, and shrinkage LDA; self-training across days and other adaptation methods were not tested.

self-training · SEEG · brain-computer interface · pseudo-labeling · shrinkage LDA · AUC