Ridge weights from training data gave mixed error changes under known input loss in eight SEEG recordings

We tested a ridge model with penalty one on public SEEG data from eight people with epilepsy. The model used eight contacts to separate rest windows from response windows. It trained on earlier matched blocks and tested the final matched block. We simulated the known loss of each selected input. We compared training-mean feature replacement with a weight adjustment from training data. The adjustment equals a ridge fit with the missing input removed, when the original feature scales and penalty stay fixed. Mean squared score error was 0.213565 with mean replacement and 0.212704 with adjusted weights. The paired mean change was -0.000861. The 95% participant bootstrap interval was -0.002596 to +0.001130. Six participants had lower mean error. Two had higher mean error. The result gives no clear evidence of a mean benefit in this sample. The package contains figures, a proof, source recipes, features, models, predictions, and checks. The test assumes correct loss detection and unchanged remaining inputs. It does not test live control or brain stimulation.

stereoelectroencephalography · SEEG · ridge regression · brain-computer interface · input loss · robustness · signal processing