A 20 mm contact model reduced later beta correlation error in 1 s bipolar SEEG windows from 21 patients

We tested a fixed contact model for later beta correlation forecasts. We used all 35 public records from 21 patients in OpenNeuro ds006519, version 1.0.0. The main model had a 20 mm spatial scale. Two early 1 s windows fitted the weights. Two later 1 s windows supplied the test correlations at 13 to 30 Hz. Scored channel pairs had separate physical contacts. Each patient had equal weight. Mean squared error was 0.038353 for shrinkage toward zero and 0.038090 for shrinkage toward geometry. The paired reduction was 0.000263, or 0.69 percent of mean comparator error. Its descriptive 95 percent patient bootstrap interval was 0.000170 to 0.000355. Twenty patients had lower error. The gain was larger on nearby pairs. The check on different electrode shafts did not establish a gain. Permuted contact locations did not reproduce the improvement. The package adds a full matrix forecast tool and preserves signals, contact locations, predictions, code, three figures, source hashes, and reproduction instructions. All 355 source files passed fresh size and hash checks. A complete rebuild matched all 35 prepared records exactly. A repeat matched all 42 numerical files byte for byte. These exploratory results apply to offline voltage statistics in a reused epilepsy cohort. Brain-interface task accuracy and causal stimulation effects were not tested.

stereoelectroencephalography · SEEG · beta band · correlation forecasting · contact geometry · spatial covariance · shrinkage · epilepsy · brain-computer interfaces