Absolute distance had no clear mean AUC gain over the label gap for absent conditions at four SEEG pairs in an offline test of eight patients

We tested two scores for responses from a stimulation condition absent from the fit and calibration data. The source was public STReEF SEEG data from eight patients. Each patient supplied four conditions. Each fold used three known templates, four bipolar recording pairs, and voltage from 20 to 100 ms after a pulse. Absolute distance gave a mean detection AUC of 0.648. The label gap gave 0.635. The mean difference was 0.013, with a descriptive 95% patient bootstrap interval from -0.089 to 0.113. This is no clear mean gain. At separate cutoffs that accepted five of six calibration pulses, distance accepted 38.8% of absent test pulses and the gap accepted 59.4%. Distance also gave fewer correct known outputs: 46.5% of offered known pulses, compared with 60.8%. Before-pulse control scores were near chance. The test has an offline scope, a reused cohort, dependent fold evaluations, and contact order confounded with time. It does not establish device performance. The package preserves source pins, features, native byte examples, models, pulse decisions, three figures, code, and reproduction instructions. Fresh source retrieval, metadata checks, independent numerical checks, and full result reproduction passed.

stereoelectroencephalography · brain-computer interface · absent-condition detection · novelty detection · area under the curve · offline evaluation