A low frequency rank threshold gave 61.5 percent balanced accuracy in 36 SEEG patients when each test patient was excluded from fitting

We tested a fixed low frequency score in 36 public SEEG patients. Each patient supplied two 120-second records. Fitting excluded both records of each test patient. With the stored signal, a rank threshold gave 61.5 percent mean balanced accuracy. The 95 percent patient bootstrap interval was 56.9 to 65.2 percent. A raw score threshold gave 58.8 percent. The paired rank minus raw difference was 2.7 percentage points, with an interval of -0.8 to 7.7 points. This interval includes zero. The data do not establish an accuracy gain from ranks. The rank rule called about 42 percent of contacts positive. The mean positive predictive value was 13.0 percent. This secondary analysis used one surgical epilepsy cohort and short interictal records. The original signal reference was unknown. It does not establish a treatment threshold. The package contains source band powers, labels, all test predictions, code, refit bootstrap results, conditional label controls, source versions and hashes, a signal example, and reproduction instructions.

stereoelectroencephalography · seizure onset zone · low frequency power · balanced accuracy · patient-level validation · contact ranking · electroencephalography · epilepsy