An 8-band high-gamma estimator beat 1–8 Hz voltage in four SEEG subjects, but the four-subject replication remains inconclusive

An independent re-analysis tested whether speech-envelope decoding results from publication d3ce6c25 changed when using an 8-band, z-scored 70–150 Hz high-gamma (HG) estimator rather than its single-band estimator. We analyzed four subjects (SD010, SD013, SD017, SD021), eight passages per subject, and common-average reference (CAR). All 12 decoders (four subjects × three feature sets) were above chance (circular-shift p ≤ 0.005). HG exceeded 1–8 Hz voltage (LF) in all four subjects (mean difference +0.130 r; exact one-sided sign-flip p = 0.0625, the minimum possible with n = 4). For the same subjects, d3ce6c25 found HG > LF in two of four. HG+LF exceeded HG in three of four subjects (mean +0.017 r, one-sided p = 0.125), a smaller gain in the same direction as d3ce6c25 (+0.044 r in seven of eight subjects). LF correlations agreed closely with d3ce6c25 (Pearson r = 0.98 across subjects); HG correlations agreed less (r = 0.79), with the largest difference for SD017 (0.254 here, 0.054 in d3ce6c25). These results suggest that the earlier conclusion that HG does not reliably beat LF can depend on the HG estimator. The comparison also changed passage count, lag set, and alpha grid, so these effects are not separated. With four subjects, no group contrast is statistically significant. The source dataset version is uncertain. Source recordings, audio, event tables, and derived feature arrays are not included because the archived report records conflicting distribution terms; their source URLs and provenance are retained.

speech envelope · SEEG · high gamma · filter bank · common-average reference · ridge regression · independent replication · OpenNeuro ds004703