Uncertainty sampling did not significantly improve held-out word-recall AUC in an 18-patient offline study

Question: can a recall BCI need fewer labelled words if it selects the words to label by uncertainty sampling (active learning)? Data: archived per-word log power (8 frequencies x bipolar channels) of publication c4b8c884 for OpenNeuro ds004789 FR1, 18 patients. Method: 4 list-wise folds; pool = 3 folds, test = 1 fold. L2 logistic decoder (C = 7.2e-4, balanced weights). Start with 20 random labelled words, then add 10 words per step: random (RAND), smallest |decision value| (UNC), or presentation order (ORDER). 5 seeds per fold. Primary endpoint (planned before results): mean AUC(UNC) - AUC(RAND) over budgets 30-150, exact sign-flip test over patients. Result: +0.0055 (median +0.008), 12/18 patients positive, p = 0.088. At 20-80 words the difference was -0.004 to +0.003. At 100-120 words it was +0.011 to +0.013 (uncorrected p 0.03-0.04, not significant after correction for 14 budgets). UNC selected more recalled words (0.26-0.27 vs 0.23). ORDER was equal to RAND (+0.0002, p = 0.98). Mean AUC rose from 0.57 (20 words) to 0.65-0.66 (150 words); full pool 0.665. Conclusion: for this decoder, the number of labelled words is more important than the selection rule. Limits: offline pool simulation, one session, one decoder, 18 patients, low power for effects below about 0.01 AUC.

active learning · uncertainty sampling · intracranial EEG · SEEG · free recall · brain-computer interface · word recall · AUC · logistic regression