SEEG decoding found no detectable previous-sentence carry-over at 1.03–1.09 s gaps in one listener

In one listener and one run of 10 TIMIT sentences repeated 10 times, SEEG high-gamma decoding tested whether sentence identity remained after baseline subtraction and whether the previous sentence carried into the next pre-onset window. Across 0–1.3 s after onset (10 classes; chance 0.10), accuracy was 0.48 for common average reference (CAR), 0.47 for CAR with pre-onset baseline subtraction, 0.34 for bipolar reference, and 0.26 for bipolar with baseline subtraction (permutation p=0.002 for each). Neither the current nor previous label was above the null in the pre-onset windows, and no previous-label time window passed family-wise error control. The offset-locked control detected sentence information through +0.25 to +0.50 s after offset, then chance-level accuracy from +0.40 s; the next sentence began 1.03–1.09 s after offset. Small carry-over (accuracy below about 0.2 with 100 trials) cannot be excluded. Accuracy was speaker-dependent (39/50 male-speaker trials and 8/50 female-speaker trials correct). The study is limited to one subject, one run, and nearly constant gaps. Bipolar decoding without baseline was 0.34 here versus 0.45 in earlier work, likely due to different filters.

SEEG · high-gamma activity · speech decoding · sentence identity · carry-over · baseline subtraction · common average reference · intracranial EEG