Random-frame cross-validation inflated speech-envelope decoding by 0.072 in eight SEEG subjects

In a completed pre-registered test using eight subjects from one dataset, random-frame cross-validation (CV) increased the mean held-out Pearson r of a linear speech-envelope decoder from 0.233 with leave-one-passage-out CV to 0.305 (+0.072; 95% bootstrap CI 0.061 to 0.086; 8/8 subjects; exact one-sided sign-flip p = 0.0039; Holm p = 0.0078). After circularly shifting each passage envelope by half its length to destroy the stimulus-brain relation while retaining autocorrelation, random-frame CV still gave r = 0.109 (8/8; p = 0.0039; Holm p = 0.0078); passage CV gave r = 0.008. The four new subjects alone showed a mean difference of +0.074 (4/4), consistent with +0.071 in the four archived subjects; rerunning SD010 gave identical r values. Random chunks of 0.25 s reduced the excess by approximately 70%, but excess of about +0.015 to +0.02 remained for 0.25-4 s chunks. A guard gap equal to chunk length removed that residual for 1 s and 2 s chunks. All three ridge penalties yielded positive FRAME inflation in 8/8 subjects. Results are limited to one passive-listening dataset, eight subjects, 4.62 minutes per subject, one offline linear ridge decoder and one frequency band (70-150 Hz); chunk and guard estimates used one seed and include split noise. The dataset README's noncommercial and no-re-identification statements conflict with cached CC0 metadata. Source data and source-derived features are not redistributed here.

cross-validation · speech-envelope decoding · SEEG · brain-computer interface · autocorrelation · data leakage · ridge regression