Three pairs selected in record A had lower mean AUC than all eligible shafts for the first 20 seizure seconds in record B from 28 SEEG patients

We tested whether three recording pairs selected from one seizure record could improve a brain-state reader in a separate record. We used 28 SEEG patient record pairs from OpenNeuro ds004100 version 1.1.3. The test used the first 20 seizure seconds and 50 pre-onset seconds. Mean AUC was 0.686 for the selected three-pair reader and 0.753 for a reader that used all eligible shafts outside the clinical onset shafts. The mean paired change was -0.068, with a 95% patient bootstrap interval of -0.112 to -0.022. The selected reader had lower AUC in 20 patients, higher AUC in seven, and the same AUC in one. Its advantage over random three-pair sets was uncertain. The reverse-record interval included zero. All 56 source records passed new bounded raw-signal checks. The package contains powers, source identities and byte-range hashes, code, saved models, window results, three figures, and commands to repeat the work. These reused retrospective event records do not establish daily alert rates, seizure prediction, or stimulation effects.

stereoelectroencephalography · seizure detection · brain-computer interface · signal processing · AUC · calibration transfer · intracranial EEG