A 10% calibrated SEEG threshold showed no significant time-order increase in false positives across eight subjects

This offline analysis tested whether a movement-detection threshold calibrated on trials 1–100 maintained its false-positive rate (FPR) and true-positive rate (TPR) on later trials in eight SEEG subjects performing a cued task (OpenNeuro ds004473 v1.0.2). A shrinkage LDA decoder used high-gamma and beta features. The fixed threshold was the 90th percentile of out-of-fold rest scores from five-fold cross-validation; a rest-only adaptive threshold used the last 30 rest scores. For a 10% FPR target, later-trial FPR ranged from 0.000 to 0.151 (mean 0.075). The primary endpoint, excess absolute FPR error in time order versus 100 shuffled-order splits, was +0.013 (6/8 subjects; exact one-sided sign-flip p = 0.113), so the primary test did not show a significant time-order effect. TPR decreased from 0.594 to 0.520 (p = 0.094). Adaptive thresholding reduced mean absolute FPR error from 0.048 to 0.033 (p = 0.055), while its mean FPR was 0.116. At a secondary 5% target, excess error and TPR decrease were significant in uncorrected tests. The analysis is limited to offline fixed windows, eight subjects, and secondary tests without multiple-comparison correction.

SEEG · brain-computer interface · movement detection · threshold calibration · false-positive rate · true-positive rate · shrinkage LDA · neural signal processing · OpenNeuro