Five time templates reduced SEEG site log loss with simulated grid errors up to 10 ms in five patients

We tested stimulation site reading with five possible pulse times in public SEEG data. The test used five patients and 183 sites from OpenNeuro ds004457 version 1.0.2. Artificial time errors moved the analysis window on a 5 ms bin grid. The five errors ranged from -10 to 10 ms with equal weight. Three pulses per site fitted templates and covariance. A fourth pulse fitted each reader's probability scale. Four later pulses per site tested the readers. For 732 physical test pulses, mean site log loss fell from 3.987 to 3.408 bits with likelihood averaging. Accuracy rose from 24.2% to 32.0%. Loss fell in all five patients. The exact two-sided patient test gave p = 0.0625. The maximum-score reader had 3.382 bits of loss and 32.5% accuracy. Averaging had no clear advantage over the maximum. The benefit persisted with a later window start. Pre-pulse results left possible background information. Fit and test times interleaved across sites. The evidence is limited to these offline sessions and artificial grid errors. Fresh raw extraction matched all source-bin arrays and event tables exactly. The package preserves data, code, models, probabilities, figures, source hashes and replay instructions.

stereoelectroencephalography · SEEG · brain-computer interface · pulse timing · site classification · likelihood averaging · time templates