Test labels in pair selection raised mean random-label SEEG accuracy in eight recordings with 32 calibration trials

We tested eight public SEEG recordings from OpenNeuro ds004473 version 1.0.2. Each decoder used eight bipolar pairs and 32 early calibration trials. One pair-selection rule used later test labels. The other used calibration labels only. Both rules fitted scales and logistic coefficients on calibration trials only. We made 500 independent within-trial label swaps per person and repeated both pipelines. Mean balanced accuracy was 49.90% for training-only selection and 56.95% for test-informed selection. The paired increase was 7.04 percentage points. Its descriptive 95% patient bootstrap interval was 5.92 to 8.33 points. An independent classifier has exactly 50% expected accuracy under these fair test swaps. For original task-state labels, the two means were 62.61% and 68.77%. A reusable test-data check detected dependence in all 32 invalid original-label pipelines. All 32 clean pipelines passed both perturbations. The package preserves input features, exact source records, all random draws, models, predictions, three figures, and repeat instructions. Independent numerical checks and fresh public raw-range checks passed. This is an offline evaluation-method result in eight clinical recordings. It does not establish live BCI performance.

SEEG · data leakage · feature selection · random-label control · brain-computer interface · balanced accuracy · model validation