Worst-mask SEEG training reduced mean worst-shaft log loss at C = 1 in eight tests with 32 fit trials

We tested four linear task-state readers in eight public SEEG recordings from OpenNeuro ds004473 v1.0.2. Each reader used 32 early fit trials and eight bipolar pairs. The primary test used 975 later trials after a condition change. A known fault removed all selected pairs on one shaft. Worst-mask training minimized the largest fit loss across intact and single-shaft cases. At fixed C=1, its mean worst-shaft log loss was 0.9312, compared with 1.0613 for ordinary training. The decrease was 0.1300 natural log units; the 95% person bootstrap interval was 0.0297 to 0.2430. Six of eight people had lower loss. The paired sign-flip reference was 0.0547. Mean fault balanced accuracy was 60.84% for worst-mask training and 61.59% for ordinary training. With a stronger penalty, C=0.1, ordinary mean worst-shaft loss was 0.7134. All tested cohort mean worst-shaft losses exceeded the 0.6931 loss of a constant 0.5 reader. The result supports a fixed-penalty loss benefit in this reused cohort. It does not establish an accuracy gain or superiority to a stronger ordinary penalty. Faults were simulated, their identities were given, and surviving inputs stayed intact. The package includes three figures, model parameters, all test scores, source hashes, raw signal audits, a reader function, and reproduction code.

stereoelectroencephalography · brain-computer interfaces · robust classification · shaft loss · logistic regression · worst-case training