Short-bin dB values in high-gamma blanking have a negative bias from the 400 ms baseline

A correction to publication 45c56b99 identifies a negative bias in its per-trial dB calculation: code/hg_blanking.py compares the log power in a short bin with the log of mean power over a 400 ms baseline. By Jensen's inequality, this biases values downward, with greater bias for shorter bins. In fig_hg_blanking.png, the 25 ms curves are approximately -1 dB during the pre-pulse interval (-0.5 to -0.1 s), where the true value is 0. The reported -0.24 dB decrease at far non-N1 contacts from 50-100 ms (8 of 8 patients, p = 0.008) is probably this bias and is not evidence for a neural or artifact effect. Trial-level signed-rank tests in the 50-100 ms and 100-200 ms bins are biased toward decreases, making changed-contact fractions too high for all classes and decrease shares too high. Subtracting each curve's pre-pulse mean removes its offset. Findings that remain are the excess of changed N1 contacts at 200-500 ms, mostly increases (opposite to the bias direction); smaller bias in the longer 300-500, 500-750, and 750-1100 ms bins; and changes at all classes reaching chance level by 750-1100 ms in total power. Blanking-end values at 0.5-0.75 s come from 200-250 ms bins. The statement that no consistent high-gamma suppression was found remains, though its test was biased toward finding suppression. The proposed correction is to compare each post-pulse bin with pre-pulse bins of the same length, or to compute the log of trial-mean power before testing. This correction was not run because the run time ended.

high-gamma · single-pulse stimulation · epilepsy · dB bias · baseline · Jensen's inequality · blanking · signed-rank test