PCA fit to scaled pairs reduced mean local high gamma reconstruction error at equal output counts in 33 SEEG patients

We tested three linear maps on public SEEG data from 33 patients. Each map used the same output count in each of 354 eight-contact blocks. The maps were fit on one 60-second interictal file prefix and tested on another. PCA fit to adjacent pairs scaled with training data reduced mean local high gamma reconstruction error from 0.403 to 0.243 versus contact energy PCA. The paired reduction was 0.160, with a descriptive 95 percent patient bootstrap interval of 0.147 to 0.171. All 33 patient means improved. Raw pair PCA had a mean error of 0.286. The scaled map increased total contact contrast energy error from 0.081 to 0.434. It also increased the fraction of pairs above a 0.10 error ratio from 62.0 to 65.0 percent. One scaled pair had an error ratio of 16.66. All 72 fresh raw-prefix covariance checks passed. The package preserves code, input covariances, fixed maps, results, figures, source versions, and reproduction instructions. This is a dependent analysis of a known clinical cohort. Task decoding, online use, and stimulation were not tested.

stereoelectroencephalography · principal component analysis · bipolar signals · high gamma · signal reconstruction · brain-computer interfaces