Clinical SEEG band power decoded speech versus music blocks above a block permutation null in one of three patients

We tested whether band power from clinical stereoelectroencephalography (SEEG) could distinguish speech from music blocks during a naturalistic film. We analyzed one 6.5-minute run from each of three patients in OpenNeuro ds003688: sub-01, sub-09, and sub-21, with 13 alternating 30-second blocks and 62–98 adjacent-contact bipolar pairs per patient. A ridge-regression decoder used log band power and leave-one-block-out cross-validation; significance was assessed against 100 permutations of the block labels. For 1-second windows and all five bands, window-level AUCs were 0.627, 0.562, and 0.679, respectively. Only sub-21 exceeded the block-label null (p = 0.010); sub-01 and sub-09 did not (p = 0.079 each). The null 95th percentiles were 0.602–0.632, showing that window AUCs near 0.6 can arise from slow changes across blocks. AUC increased with window length, while theta, alpha, and beta power performed almost as well as all bands; high gamma alone was weaker. No channel had a significant high-gamma contrast after false-discovery correction. Mean decoder transition latencies were 0.25–2.25 s. These results provide limited evidence for speech-versus-music state decoding in these implants. The study includes only three patients and 13 blocks per run; visual content, attention, and arousal may differ between block types, and the data do not isolate speech processing. The package preserves code, derived band-power features, results, figures, and source-file provenance.

stereoelectroencephalography · SEEG · speech decoding · brain-computer interface · band power · block permutation · naturalistic film