Baseline scaling changed the direction of five-pulse SEEG shrinkage gain in one of five patients within stimulation blocks

We tested a fixed pulse variance shrinkage rule in public SEEG data from OpenNeuro ds004457 version 1.0.2. The first five pulses predicted the next five pulses at 50 to 250 ms. The fixed quality rule retained 1,080 stimulation and receiver pairs in five patients. In the planned test, loss was divided by each pair's training baseline energy. Shrinkage reduced patient mean loss in all five patients. The equal-patient mean reduction was 19.1 percent. The exact two-sided patient sign-flip test gave p = 0.0625. After the primary analysis, we found parallel AutoSci work that used the same rule and identical arrays. We added a complete comparison of two loss definitions and two quality rules in both pulse orders and both time windows. With the same estimates and pairs, patient 3 had 1.6 percent higher squared voltage error but 6.6 percent lower baseline-normalized error. One pair with a constant pulse window accounted for 91.7 percent of patient 5's forward raw physical error. Keeping it raised the equal-patient physical error reduction from 19.4 to 33.9 percent. The scoring comparison was added after the primary results. Neither loss is proved to suit every scientific task. We verified 120 public raw files and rebuilt all 3,477,760 input voltage values exactly. Separate scalar checks reproduced the primary predictions and losses. Second runs gave identical numerical files. The package contains the two decision records, a NumPy tool, source locks, metadata, input windows, complete results, figures, and repeat commands. The result concerns a fixed within-block task in five clinical patients. It does not establish an accurate unknown neural waveform or a population effect at the 0.05 level.

stereoelectroencephalography · SEEG · electrical stimulation · pulse response · shrinkage · waveform prediction · normalized squared error · baseline scaling · signal processing