Background subtraction gave uncertain site log loss gains with three template pulses in five SEEG sessions

We tested whether subtracting a forecast of ongoing voltage improves site probabilities from public SEEG data. The source was OpenNeuro ds004457 version 1.0.2. The test used 94 later sites and 376 pulses from five people. Each site supplied three template pulses, a separate probability calibration pulse, and four test pulses. The background model used earlier sites and voltage before each pulse. Mean test log loss fell from 2.721 to 2.704 bits per pulse. The paired gain was 0.017 bits per pulse. Its descriptive 95 percent patient interval was minus 0.003 to plus 0.048. Mean accuracy fell from 35.71 to 35.43 percent. A matched, post hoc known-site waveform check had 3.14 percent lower error. Thus, lower waveform error did not establish a clear site readout gain in this test. The reused sample and retrospective site library limit the result. The package preserves compact voltage data, full reader probabilities, source hashes, fitted models, code, three figures, and reproduction instructions. A fresh raw extraction matched the compact data. A new process reproduced all 62 main result files.

stereoelectroencephalography · SEEG · brain-computer interface · voltage correction · probability calibration · site classification