Two shared time waveforms gave lower mean 100 to 500 ms prediction loss than one in seven selected limbic SEEG patients under fixed source correction

We tested one shared time waveform against two in public limbic SEEG data. The test used 80 stimulation groups from seven selected patients in OpenNeuro ds004696, version 1.0.1. We fitted each model on the mean of five pulses. We scored it on the mean of five separate pulses at the same contacts. We used all 252 oriented splits of ten pulses. The tested window was 100 to 500 ms, after 45 Hz filtering and window mean removal. Groups had equal weight within patients. Patients had equal weight. The mean relative reduction in squared prediction loss was 11.50 percent. The conditional 95 percent patient bootstrap interval was 5.61 to 17.61 percent. Six of seven patients and 49 of 80 groups had lower loss. The exact two-sided patient sign-flip probability was 0.03125. It assumes independent patients and symmetric effects under the null. Before a pulse, two waveforms increased mean loss by 10.55 percent. Source correction used all exported pulses before the split. The result is conditional on that correction and the frozen selection. It does not identify two neural sources or an optimal rank. The contribution preserves voltage inputs, per-split scores, source provenance, and result files; GPL code is omitted from the SCI payload, which limits executable reproducibility. Nine numerical output files reproduced exactly.

stereoelectroencephalography · electrical brain stimulation · waveform prediction · singular value decomposition · limbic system · cross-validation · OpenNeuro