Fitting on other patients chose zero added distance shrinkage for later 1–80 Hz SEEG covariance shape at a 10% ridge in 36 patients

We tested whether contact distance improved later SEEG covariance shape after 25 seconds of local calibration. We used two interictal files from each of 36 epilepsy patients in OpenNeuro ds004100 version 1.1.3. The models used a fixed 10% ridge. We tested six shrinkage weights and five Gaussian distance lengths. Each test fold fitted settings on the other 35 patients. Both model families chose zero added shrinkage in all 36 primary folds. The mean gain was zero nats per contrast. A refitted bootstrap of 20,000 patient resamples gave a 95% interval of −0.02276 to zero. At a fixed weight of 0.5, distance had lower loss than independence in 35 patients. The same fixed distance model had higher loss than the selected zero-weight model in 34 patients. Thus, the tested distance family gave no primary gain after the comparator was fitted. This result is limited to the tested grid, ridge, blocks, and reused cohort. It does not show that distance has no signal information. All 72 inputs were rebuilt from pinned public raw bytes and matched the archived arrays exactly. The package preserves the data, raw example, source metadata, code, fitting decisions, uncertainty draws, results, and reproduction instructions.

stereo-electroencephalography · covariance prediction · spatial prior · distance kernel · shrinkage · epilepsy