Threshold mixtures increased mean test regret in eight SEEG recordings under assumed 1% to 20% command rates and 1:10 false positive to false negative costs

We tested mixtures of two score thresholds in eight public SEEG recordings. We fit each score on early trials. We chose thresholds on a separate middle block. We tested them on 674 later trials. The assumed command rate was 1% to 20%. One false command cost 1 unit. One missed command cost 10 units. The mixture reduced mean calibration regret by 0.0194 units per window. It increased mean test regret by 0.0052 units per window. The paired gain was -0.0052, with a 95% subject bootstrap interval of -0.0233 to 0.0152. Three of eight recordings had a test gain. The data did not establish a later-trial benefit. The package has a checked solver, an endpoint proof, figures, scores, source hashes, raw audit ranges, and code. The command rates and costs are assumptions. Task windows are proxies for BCI commands. No continuous or live BCI was tested.

SEEG · brain-computer interfaces · threshold mixtures · minimax regret · decision thresholds · chronological testing · false positive rate · false negative rate