Truncated SEEG data show weak speech–music decoding in one patient and inconclusive results with three to four blocks

This pilot examined whether SEEG high-gamma power decodes speech versus music blocks in a naturalistic film, and how accuracy varies with contact count. It used truncated prefixes of OpenNeuro ds003688 v1.0.7 for three patients. For sub-09, leave-one-block-out accuracy across 69 contacts was 0.635; the exact block-label permutation null had a 95th percentile of 0.619 and p = 0.056 (125 assignments). Random-subset mean accuracy rose from 0.48 with one contact to 0.65 with 64; top-k selection reached 0.67 at k = 8. Four of 69 single contacts (5.8%) exceeded the null 95th percentile, close to the 5% expected by chance. Accuracy was 0.232 for sub-42 and 0.262 for sub-21, with only four and three blocks; these below-chance results are not interpretable because each held-out block leaves little training data for its class. The evidence for speech–music decoding is weak and comes from one patient; it does not establish that any contact count is sufficient. The study also has truncated recordings, one frequency band and decoder, no anatomical selection, and uncorrected audiovisual film confounds.

SEEG · high-gamma · speech decoding · music · brain-computer interface · block permutation · contact count