A 41-sample variance mean reduced later normal SEEG voltage log loss with five fit pulses at 86 selected sites in eight patients

A fixed mean of 41 sample variances lowered later normal voltage log loss in public SEEG from eight patients. We used OpenNeuro ds004696 version 1.0.1 and 86 inherited selected stimulation sites. Five early good pulses fitted each site. The remaining 612 pulses tested it in the same session. The 41-sample mean spans 19.53 ms. It changed only the variance. The patient-weighted loss fell from 1.1432 to 0.9582 nats. The decrease was 0.1850 nats; its 95% patient bootstrap interval was 0.1277 to 0.2455. All eight patients had a decrease. The Student t decrease was 0.02652 nats. Its separate-variance reference also scored better than the local normal forecast. The benefit was larger before stimulation. This supports an estimation effect. Other fixed widths scored better than 41 samples. Sample density scores do not test joint response forecasts. Nominal 90% local normal limits covered 85.17% of samples and 4.76% of complete responses. These limits do not support a calibrated device alarm. The study is exploratory. It does not test new sessions, intent decoding, or treatment. The package retains exact input windows, three figures, numeric results, source pins, and check records. Fresh source checks passed for all eight patients. An independent calculation and clean replay passed historically. GPL analysis and extraction scripts are omitted, so the analysis cannot be rerun from this package.

stereoelectroencephalography · variance estimation · local averaging · voltage forecast · negative log density · Student t · calibration · bootstrap · brain stimulation · response model