Fixed atlas weights gave higher mean normalized error from 20 to 100 ms in five offline 6 mA SEEG sessions

We tested whether source contact atlas labels help predict a response at an excluded stimulation site. We used public SEEG data from OpenNeuro ds004457 version 1.0.2. The test had 52 sites in five patient sessions and fixed ventral temporal readouts. Each transfer model used the first four pulses at each immediate flank. It used no target pulses. The test used the last four target pulses. Equal weights gave mean normalized squared error 0.5832 from 20 to 100 ms. A fixed 2:1 atlas rule gave 0.5945. The mean skill change was -1.13 percentage points. The descriptive interval was -4.58 to +2.15 points. The exact sign-flip p value was 0.625. The rule changed weights at 17 sites. Reverse weights gave little mean difference. An exact error split showed little mean direction benefit from the atlas choice. The unnormalized voltage error had the opposite point direction. This result depends on the score and the readout set. The test is offline. Of 416 donor pulse roles, 200 occurred after the target test block. The package supplies a fixed benchmark, site errors, predictions, figures, code, source hashes, and raw retrieval commands. All 110 raw source files passed hash checks. A raw rebuild reproduced every compact array and event field exactly. A clean run matched all 14 numerical result files. The study reuses prior recordings. It does not establish a wider benefit or harm.

stereoelectroencephalography · electrical stimulation · ventral temporal cortex · response prediction · atlas labels · normalized error