Four donor components increased mean squared voltage error 6.58% versus noise scaling after five calibration pulses in eight SEEG sessions

We tested a voltage forecast for brain stimulation. We used public SEEG data from eight patients with epilepsy in OpenNeuro ds004696 version 1.0.1. For each patient, we learned four waveform components from early pulses in the other seven patients. These seven patients were the donors. Five early pulses at each readout set the forecast. The final five pulses tested it. The primary target was centered bipolar voltage from 20 to 500 ms after each pulse. The test used 86 sites, 886 readouts, and 430 different later pulses. The shared forecast increased mean squared error by 6.58% versus a full-resolution noise-scaled five-pulse mean. Error was lower in 0 of eight patients. The fixed four-component cosine control had 13.11% more error than the same comparator. The shared forecast reduced error by 5.22% versus this control. Positive percentages mean less error. The patient bootstrap range was -8.86% to -4.32%. The folds share training data, so this range is descriptive. The package contains source pins, compact pulse data, code, scores, checks, and four figures. The test concerns voltage in the same session. It does not test a percept, new-day transfer, or online control.

stereoelectroencephalography · brain stimulation · voltage forecast · shared waveform components · calibration pulses · human limbic system · brain-computer interface