In one SEEG listener, high-gamma best identified sentences with common-average reference, while low-frequency voltage remained informative with bipolar reference

In one subject listening passively to 10 sentences repeated 10 times, we compared high-gamma amplitude (HG, 70–150 Hz), low-frequency voltage (LF, 1–8 Hz), and alpha-beta amplitude (AB, 8–30 Hz) for single-trial sentence identification from stereo-electroencephalography (SEEG). With common-average reference (CAR), 10-class shrinkage LDA accuracy at 1.5 s was 0.49 for HG, 0.38 for LF, 0.15 for AB, and 0.51 for HG+LF (chance 0.10). The available CAR label-permutation null placed HG, LF, and HG+LF above its 95th percentile (0.15–0.16; p=0.01); AB was not significant (p=0.08). With bipolar reference, HG accuracy fell to 0.26, while LF reached 0.36 at 1.5 s and 0.40 at 1.0 s. At 0.5 s, LF exceeded HG with both references (CAR 0.28 vs 0.23; bipolar 0.32 vs 0.17). For CAR, 15/106 HG channels and 11/106 LF channels passed the reported false-discovery-rate threshold for sentence-specific reliability; no bipolar channel passed. HG and LF reliabilities were weakly correlated across channels (r=0.30–0.31). These results suggest LF may be useful when bipolar montages or short windows are required, but are a pilot from one subject and run. Bipolar permutation tests and the paired HG+LF comparison did not finish; there was no acoustic model, and the apparent bipolar LF result may include residual volume conduction. Sentence identity may therefore reflect duration and onset structure as well as phonetic content.

stereo-electroencephalography · speech perception · sentence identification · high-gamma · low-frequency voltage · alpha-beta · common-average reference · bipolar reference · intracranial recording · neural decoding