A fixed local update from five minutes reduced mean Poisson loss for MNI HFO forecasts in 18 SEEG sleep records with equal predicted totals

We tested whether recent channel counts improve HFO forecasts beyond a change in the total count rate. We used released detector marks from 18 public SEEG sleep records in OpenNeuro ds005398 version 1.1.3. The first 30 minutes set a fixed baseline. The next 30 minutes tested a causal update from five completed minutes. The local rule and global control predicted exactly the same total count each minute. Mean Poisson loss fell by 12.12 percent, from 1.71607 to 1.50800 nats per channel per minute. The paired gain was 0.20807. Its 95 percent patient bootstrap interval was 0.07545 to 0.35628 under independent, exchangeable patient sampling. Fourteen of 18 patients had a positive gain. A control with equal totals on each named shaft also gave a positive mean gain. The STE detector check remained uncertain. Equal predicted totals make the primary gain an exact channel-share log-score gain. This identity needs no Poisson-noise or independent-channel assumption. The study is a secondary analysis of a reused cohort. It does not establish event validity or clinical benefit. The package preserves pinned source bytes, counts, every forecast and score, inference arrays, code, figures, and run instructions. Independent checks passed. A clean offline run matched all 49 numerical files byte for byte.

high-frequency oscillations · stereoelectroencephalography · event counts · sequential forecasting · Poisson loss · channel shares · sleep · epilepsy