Set-size-6 trials did not improve a 4-vs-8 SEEG working-memory load decoder (ds004752, 15 patients, one session, offline band-power shrinkage LDA)

Question: a load-monitoring BCI is usually calibrated on the extreme classes only (set size 4 vs set size 8), and the set-size-6 trials are discarded. Do the set-size-6 trials improve the decoder when you train on all trials with a graded load target? Data: OpenNeuro ds004752 (verbal Sternberg task, SEEG), session 1, 15 patients, archived maintenance band power (3-6 s, bipolar, 5 bands). Method: two shrinkage LDA decoders with the same Ledoit-Wolf estimator. TWO was trained on set sizes 4 and 8. GRADED was trained on set sizes 4, 6 and 8 with w = S^-1 cov(x, load). Stratified 5-fold cross-validation, 3 repeats. AUC on held-out set-size-4 and set-size-8 trials. The plan was written before the analysis. Primary result (pre-registered H1, exact one-sided sign-flip test): GRADED minus TWO = -0.002 AUC (95% CI -0.025 to +0.024), 5 of 15 patients positive, p = 0.56. H1 is not supported; the predicted gain of +0.01 to +0.03 did not occur. Mean AUC was 0.708 (TWO) and 0.706 (GRADED). Secondary: at equal training size, set-size-6 trials were less useful than extreme trials (-0.028 AUC, CI -0.056 to -0.001, two-sided p = 0.072). With 25-75% of the extreme trials, adding set-size-6 trials gave approximately +0.01 AUC (8 of 15 positive, not significant). Deviation: 3 repeats instead of the planned 10, because of compute load. Limits: one session, offline, one window, one feature set, one linear estimator, 15 patients. The package contains the code, input features with checksums, per-patient results, a figure and reproduction steps.

brain-computer interface · intracranial EEG · working memory · load decoding · shrinkage LDA · band power · cross-validation · OpenNeuro