Extra rest windows did not significantly improve 10-trial SEEG key-press calibration but improved 5-trial calibration in 8 subjects

Question: a motor BCI must have labeled movement trials for calibration, but rest data is easy to get. When only a small number of calibration key presses are available, does the addition of extra rest windows increase the test AUC of a key-press detector? Data: archived HG (70-150 Hz) and beta (13-30 Hz) bipolar log-power features of sci 04683a2a (OpenNeuro ds004473 v1.0.2, 8 subjects, CC0). Method: shrinkage LDA. Test set: all rest and move windows of trials 101 to end. BAL = rest and move windows of trials 1..n. REST+ = move windows of trials 1..n plus rest windows of trials 1..100. We wrote and hashed the plan before decoding. Primary endpoint: D10 = AUC(REST+) - AUC(BAL) at n = 10, exact one-sided sign-flip test. Result: the primary hypothesis is not supported. Mean D10 = +0.018 (median +0.020, range -0.043 to +0.104, 5/8 positive, p = 0.17). Secondary: D5 = +0.057 (7/8, Holm p = 0.016); D20 = +0.012 (6/8, Holm p = 0.13). With 20 random draws of n trials from 1-100 (descriptive, not corrected), mean D was +0.044, +0.028 and +0.018 for n = 5, 10, 20 (7/8 or 8/8 positive). The gain becomes smaller when n increases. REST+ stayed below FULL (100 key presses) in 8/8 subjects at each n (mean gap 0.108, 0.074, 0.041). In the two subjects with the lowest AUC, the time-ordered gain was negative at n = 10 and 20. Limits: offline, one session, 8 subjects, cued task; the extra rest windows come from later trials of the same session, not from a separate rest block. The package has code, input hashes, results, a figure and reproduction steps.

brain-computer interface · stereo-electroencephalography · decoder calibration · rest windows · shrinkage linear discriminant analysis · area under the curve