Adding time concentration did not increase mean same-shaft label AUC in fixed 80 to 150 Hz models tested on 30 SEEG patients

We tested whether time concentration adds clinical seizure onset label ordering beyond mean power. The source was OpenNeuro ds004100 version 1.1.3. Each of 30 patients supplied two 60 s SEEG clips. Power used 80 to 150 Hz, with 118 to 122 Hz omitted. Concentration was the fraction of summed one-second power in the six largest of 60 windows. Training excluded the tested patient. The fixed models used pairwise logistic loss and an L2 penalty of 0.1. Comparisons used positive and other pairs on the same named shaft. Patients had equal mean weight. Baseline mean AUC was 0.8296. Added-model mean AUC was 0.8279. The difference was -0.0017. Its 95 percent patient bootstrap interval was -0.0219 to +0.0110, with all models fitted again in each of 2,000 samples. This approximate interval assumes independent patients and representative patient sampling. The test did not detect a mean gain. It does not prove equal performance. Planned checks without shared contacts and with contact number gave similar small negative estimates. The study extends prior work in the same cohort. It does not test surgical outcomes or an external cohort. All 68 input feature files matched a fresh extraction from 72 pinned raw ranges. Independent reader and model checks passed. The package preserves data, source versions, hashes, code, scores, bootstrap identities, figures, failure logs, and reproduction commands.

stereoelectroencephalography · SEEG · epilepsy · time concentration · high-frequency power · same-shaft comparison · patient bootstrap · AUC