Training selection of covariance between shafts lowered median loss but raised mean loss in 36 ds004100 SEEG patients at 1–80 Hz with 1% ridge

We tested whether covariance between named electrode shafts improves prediction of a separate SEEG file. The data came from 36 epilepsy patients in OpenNeuro ds004100 version 1.1.3. Each patient had two interictal files. Each training file selected a covariance weight from two separately filtered blocks. No scale was fit to test power. At 1–80 Hz with a 1 percent ridge, gain was positive in 33 of 36 patients. The median gain was 0.054549 nats per contrast dimension. Its 95 percent patient bootstrap interval was 0.040170 to 0.064261. The mean gain was −2.033802. A large failure in HUP140 made the observed mean loss worse. All patients were kept. This is a method extension on a known cohort. Contact eligibility used both files, and the test used short excerpts. The result does not establish causal neural interaction or clinical benefit. The archive includes all covariance inputs, code, source pins, a raw fixture, results, matrix proofs, and repeat instructions. Two complete analysis runs matched exactly. Four fresh public source rebuilds matched every saved covariance value.

stereoelectroencephalography · covariance prediction · cross-validation · signal processing · epilepsy · robustness