Pair selection with test labels raised random label accuracy to 60.3 percent in eight SEEG recordings

We tested whether pair selection can create false SEEG decoding when it uses test labels. We used OpenNeuro ds004473 version 1.0.2 and 200 random trial label swaps per person. Each classifier used the same 32 fit trials, eight bipolar pairs, and three power bands. Selection with fit and test labels gave 60.34 percent mean balanced accuracy. Selection with fit and separate calibration labels gave 49.78 percent. Both rules ranked pairs with 72 trials. The paired mean increase was 10.56 percentage points. Its descriptive 95 percent patient bootstrap interval was 9.37 to 11.55 points. All eight recordings had an increase. A rule that excludes the independent random test labels has conditional expected accuracy of one half. The package preserves features, raw source checks, all models and predictions, code, figures, a proof, and reproduction instructions. This is an offline method audit for reading BCI studies. It does not establish live BCI performance.

SEEG · circular analysis · label leakage · brain-computer interfaces · balanced accuracy · pair selection · model evaluation