In one patient, epilepsy-flagged SEEG pairs tracked speech better, but flag and temporal location were confounded

We tested whether retaining stereoelectroencephalography (SEEG) contacts marked bad|epilepsy preserves speech-envelope information in the only subject in OpenNeuro ds004703 with those flags, sub-SD012. We analyzed two sessions from the same patient using 70–150 Hz bipolar amplitude. For speech-envelope tracking, we correlated z-scored neural amplitude and envelope at lags of 0–250 ms, used 500 circular shifts as a null, applied Benjamini–Hochberg false-discovery-rate correction across pairs, and compared flagged and good pairs with 20,000 label permutations. In the short ses-01 (2 passages, 70 s), flagged pairs had a mean r_max 0.021 higher than good pairs (p=0.25). In ses-02 (16 passages, about 9.6 min), the difference was +0.030 (p=0.006), and remained +0.018 after removing the left superior temporal (LST) shaft (p=0.011). The fraction with q<0.05 was 0.35 for flagged pairs and 0.19 for good pairs in ses-02. A leave-one-passage-out ridge decoder in ses-02 had mean r=0.043 with good pairs, 0.185 with flagged pairs, and 0.338 with all pairs; all-pair decoding was better on 16 of 16 passages (exact sign-test p=3e-5). Results came from one patient, two non-independent sessions, one passive-listening task, and one neural feature. Flagged contacts were mostly temporal or hippocampal, so clinical flag and location are confounded; the results do not show that epileptic tissue encodes speech better. The source dataset metadata says CC0, while its README is reported to forbid commercial machine-learning use and re-identification. Recordings, audio, event tables, and extracted feature caches are not included. The dataset’s terms and version are linked for readers; no commercial rights to that external dataset are asserted.

stereoelectroencephalography · speech envelope · high-gamma amplitude · epilepsy flags · brain-computer interface · neural decoding · OpenNeuro ds004703