Bipolar subtraction reduced mean later sound prediction correlation for depth pairs selected with a common average reference in nine Podcast recordings

This study tested whether an early screen from common average reference (CAR) data transfers to a bipolar measurement. It used nine public Podcast recordings from OpenNeuro ds005574, version 1.0.2. Both targets came from the same cleaned voltage files. The first five minutes selected 51 of 233 eligible depth pairs. Separate sound models used those five minutes and tested the last ten minutes. Mean participant prediction correlation was 0.1587 for the CAR pair target and 0.1193 for the bipolar target on the same list. The mean difference was -0.03933. Its participant bootstrap 95 percent interval was -0.06203 to -0.01711. The exact sign-flip probability was 0.01953 under independent, sign-symmetric differences. Eight of nine differences were negative. Seven bipolar results passed the specified distant-shift alignment control. A separate early bipolar screen gave no clear group gain. A source audit found a larger amplitude mismatch and extreme cleaned-voltage values for participant 8. An additional check without this participant retained a negative mean difference. This result is conditional on the pinned voltages, one story, and the CAR selection rule. It does not measure artifact content or prove a local neural generator. The package preserves compact inputs, source versions and hashes, code, results, figures, and instructions. Two complete model runs agreed.

electrocorticography · bipolar reference · common average reference · sound prediction · depth electrodes · OpenNeuro