Ridge tuning by inner squared error gave 48.2% sound class accuracy on held-out SEEG excerpts from ten people

We tested training-only ridge penalty selection for a binary SEEG sound class model using 3,516 balanced events from ten people with epilepsy. Whole speech excerpts were held out in nested folds. With all candidate pairs, inner squared error selection reduced mean signed-label error from 2.1113 to 1.0011, while mean balanced accuracy changed from 49.37% to 48.18%. The paired person difference was -1.18 percentage points (95% bootstrap interval -2.28 to -0.20; exact sign-flip p=0.0508). An exploratory follow-up selected penalties by inner accuracy and gave 48.16% test accuracy. The finite penalty grid often reached its largest value. Lower score error did not give a mean class accuracy gain in this tested cohort. The SCI contribution retains the methods, code, result tables, model and prediction results, and figures; it omits source recordings, event tables, and source-derived feature caches because the source dataset records conflicting terms. Direct replay therefore requires external inputs and resolution of their terms. This study does not establish online brain readout.

stereoelectroencephalography · ridge regression · nested validation · speech · brain-computer interface · balanced accuracy · squared error