Median log band power subtraction changed mean AUC by -0.0012 in a fixed test with five neighbors on 14 excluded SEEG patients

We tested whether subtraction of each patient's median log band power improved prediction of clinical seizure onset labels. We used public SEEG data from OpenNeuro ds004100, version 1.1.3. A fixed five-neighbor model used 20 training patients. The main test used record 02 from 14 other patients. No test patient label entered a fit. Mean patient AUC was 0.6023 with raw features and 0.6011 after median subtraction. The paired change was -0.0012. Its 95 percent conditional patient bootstrap interval was [-0.0790, +0.0656]. Six patients improved and eight had lower AUC. The other fixed recording and neighbor checks did not show a consistent gain. A contact-number score gave mean AUC 0.7604 in the same test patients. Physical depth was not verified. We provide a fixed benchmark, all predictions, saved model inputs, source recipes, code, and independent checks. Public byte ranges rebuilt all 70 saved common average reference spectra exactly at float32 precision. The cohort has been used in prior work. The interval keeps training fits fixed. The result does not establish equal methods, a biological cause, or clinical use.

stereoelectroencephalography · seizure onset zone · patient exclusion · median subtraction · band power · nearest neighbors · AUC · clinical labels