Full quadratic terms did not show lower error in ten SEEG listeners with eight pairs and 400 ms delay

This study tests whether products of SEEG power reduce heard-speech amplitude error. It uses public OpenNeuro ds004703 version 1.1.0. Ten people supply 95 short passages. Each reader uses eight adjacent contact pairs, 70 to 110 Hz power, and five time lags. The main test permits 400 ms of added signal delay. Pair choice, scaling, and ridge penalty use early passages only. Later passages test the readers. The linear reader has 40 columns. The full quadratic reader has 860 columns. Its mean normalized error reduction changes from 3.21% to 1.84%. The paired change is -1.37 percentage points, with a descriptive 95% person interval from -9.22 to +4.33. This does not establish equal performance. The planned squares-only reader has 80 columns and gives a secondary gain of +5.72 points, with an interval from +1.60 to +10.39. That result needs a new test sample. The package preserves the plan, compact data, source metadata, pinned retrieval recipe, code, fitted readers, predictions, four figures, and checks. A separate SVD check covers all 80 reader fits and 100 tuning grids. A full analysis replay matches 166 files and 1,624 arrays exactly. The scope is later heard-speech passages in the same clinical recordings. It does not test intended speech, a live device, or stimulation.

SEEG · speech amplitude · brain-computer interface · ridge regression · quadratic features · signal delay · exploratory study