In one SEEG listener, sentence accuracy was similar with common average and bipolar references without baseline subtraction

In one listener (OpenNeuro ds004993, subject W2, one run), we tested whether bipolar re-referencing removes sentence information from SEEG high-gamma amplitude (70–150 Hz). A correlation nearest-centroid classifier distinguished 10 TIMIT sentences across 100 trials using 106 common-average-reference contacts or 98 bipolar pairs. With 50 ms bins from sentence onset to 1.3 s and leave-one-repeat-out cross-validation, accuracy was 0.46 with common average reference and 0.45 with bipolar reference (10% chance; 1000-permutation p=0.001 for each). With per-trial subtraction of the −0.5 to 0 s baseline, bipolar accuracy was 0.27 in leave-one-repeat-out validation and 0.31 in stratified five-fold validation repeated five times. Thus the apparent reference effect in a prior analysis is sensitive to baseline choice. Accuracy increased with training repeats and had not plateaued at nine repeats. Pre-onset sliding windows reached 0.17–0.19, above the 0.15 null threshold for one window, although these were maxima across 10 windows; slow state or presentation order may contribute to accuracy without baseline subtraction. The study cannot distinguish these causes. It includes one subject and run, does not test common average reference with baseline subtraction, and has no paired test of the references. Top-k channel selection was exploratory and not nested. Speaker sex was confounded with sentence identity. These results do not establish generalization beyond this recording.

SEEG · high-gamma · speech decoding · sentence identification · bipolar reference · common average reference · baseline subtraction · brain-computer interface