Speech-envelope decoder accuracy varies by passage in 8 SEEG subjects

This secondary analysis tested whether held-out speech-envelope decoder accuracy depends on the passage as well as the subject. We analyzed published per-passage Pearson r values from two earlier SCI analyses of OpenNeuro ds004703 for 8 SEEG subjects. For six ridge decoder feature and reference combinations evaluated on 16 passages, Fisher z(r) variance attributed to passage ranged from 9% to 24%; the passage effect was significant in permutation tests (p = 0.0001–0.0024). For the high-gamma common-average reference decoder, passage accounted for 22% of variance, the 8-subject passage profile had ICC = 0.91, and the profile from 7 subjects predicted the held-out subject (mean Spearman rho = 0.71; positive for 8/8; exact sign-flip p = 0.0078). Presentation position did not explain this effect, and passage duration was not associated with it (rho = 0.04, p = 0.88). The 1–8 Hz passage profile did not correlate with the high-gamma profile (rho = 0.07). A causal high-gamma plus low-frequency decoder evaluated on 8 common passages showed a smaller passage share (7%, permutation p = 0.0032). With 2–4 test passages, passage choice accounted for approximately 60% of the variance in a subject's accuracy estimate; the adjusted estimates were slightly optimistic because the passage profile included the same subject. The results support comparing decoders, references, subjects, or sessions on the same test passages, or removing the passage effect. This is a secondary analysis of published decoder results from 8 subjects, one dataset, one passive-listening task, and ridge decoders; raw SEEG was not re-extracted, and acoustic causes were not tested.

speech-envelope decoding · brain-computer interface · stereo-EEG · passage effects · decoder accuracy · reliability · cross-validation