Eight selected micro-SEEG contacts reduced mean voltage error by 4% across eight current blocks in one pig at penalty 0.1, with a conditional 95% range that included zero

We tested whether calibration data can select eight micro-SEEG contacts to predict 31 fixed omitted voltage differences. The source was OpenNeuro ds004819 version 1.0.0, with one pig implant and eight spinal stimulation current blocks. Each test excluded one whole block. Selection used inner block validation only. The main ridge penalty was 0.1. There were 79 unique test pulses. The mean error ratio was 0.9399 for eight fixed contacts and 0.9019 for eight selected contacts. This was a 4.04% decrease. Six blocks improved and two became worse. The conditional 95% gain range was -3.69 to 8.64 points of target energy. It included zero and did not include calibration or animal uncertainty. The gain was concentrated in targets with large source impedance numbers. All three main mean scores were worse than zero for the mean pulse waveform. The test measured recorded voltage and did not establish a gain in neural information. The package preserves all prepared inputs, source versions and hashes, retrieval code, splits, contact choices, validation scores, maps, predictions, errors, resampling draws, and three figures. A new raw audit matched every prepared value. Independent numerical checks and a full rerun passed. This is an offline benchmark for a later read budget. It requires full calibration and tests no human control or brain writing.

micro-SEEG · contact selection · voltage prediction · ridge regression · brain-computer interfaces · pig · cross-validation · stimulation artifact