Random-frame cross-validation inflated decoder correlation by 0.071 in four SEEG listeners; a 1–2 s guard removed the excess

This study tested cross-validation (CV) leakage in an offline continuous speech-envelope decoder using stereo-EEG (SEEG) high-gamma features from four of eight planned listeners in OpenNeuro ds004703. A backward ridge decoder used 40 high-gamma PCA components and lags from 0 to 250 ms. At the primary ridge setting, random-frame CV gave mean held-out r = 0.332, compared with 0.262 for leave-one-passage-out CV (mean increase 0.071; all four subjects; exact one-sided sign-flip p = 0.0625, the minimum possible at n = 4). After circularly shifting each passage envelope to destroy its stimulus-brain relation while retaining autocorrelation, random-frame CV still gave mean r = 0.109, compared with 0.015 for passage CV. Random contiguous chunks of 0.25–8 s reduced the excess to about 0.02, while 1 s chunks with a 1 s guard gap and 2 s chunks with a 2 s gap reduced it to +0.001 and approximately 0. The result was similar across relative ridge settings of 0.1, 1, and 10. Because only four subjects were completed, no exact sign-flip test could reach p < 0.05. The experiment used one dataset, one linear decoder, and offline data; PCA over all passages was a shared unsupervised leak, and chunk and guard schemes used one seed. The findings support leave-one-segment-out evaluation or chunks with a guard gap of at least 1 s for continuous BCI decoders, within these limits. The original dataset README states restrictions on commercial machine-learning use and re-identification; dataset inputs are not included.

cross-validation · data leakage · stereo-EEG · speech envelope · brain-computer interface · autocorrelation · ridge regression · guard gap