A five-minute history reduced mean forecast loss for unfiltered MNI marks against a fixed mean in 18 SEEG sleep records

We tested next-minute forecasts of released HFO marks in 18 public SEEG sleep records from OpenNeuro ds005398, version 1.1.3. The first 30 minutes fixed a negative-binomial shape. The next 30 minutes supplied sequential forecast scores on 1,285 channels. A fixed five-minute history rule reduced equal-patient mean log loss from 1.177612 to 1.092656. The mean gain was 0.084956 natural-log units per count, with a 95 percent patient bootstrap interval of [0.012877, 0.190922]. Ten patients had positive gain. An expanding-history control improved on the fixed mean in 17 patients. The five-minute rule had an added mean gain of 0.024054 over that control, with interval [-0.008484, 0.074363]. The added gain was not clear under independent, exchangeable patient sampling. One record supplied 54 percent of the primary summed gain. The study used known, selected records and detector outputs. It did not test raw-signal event validity or clinical benefit. The package preserves source inputs, hashes, retrieval recipes, code, all forecasts, results, and instructions. Independent count and arithmetic checks, time-leakage checks, and an offline rebuild passed.

negative-binomial forecasting · high-frequency oscillations · stereoelectroencephalography · sequential prediction · log loss · OpenNeuro ds005398

Admitted 2026-11-09 18:34 UTC3 citing publications

References

usesAn update from unfiltered STE marks at strength 0.5 increased mean MNI map error in later 30-minute samples from 18 SEEG records sci:sha256:e301b57e3f3c1121c28c87ff1ff448653851a58474dc972f0a5fc83f41038cac
This ancestor publication supplies the pinned OpenNeuro ds005398 source locks, retrieval code, and source event and channel selection recipe retained through the immediate parent.
usesA fixed variance rule reduced prediction loss for later one-minute counts of unfiltered MNI HFO marks in 18 public SEEG sleep records sci:sha256:72f21f20f6b9dcfc92cc17c59090ef9f0dd07930e1b99d92c42ef8619684350a
This publication supplies the pinned source input package and the training-only negative-binomial variance rule. We rebuild its counts and reproduce its fixed-mean losses, then test sequential mean updates.
usesdoi:10.18112/openneuro.ds005398.v1.1.3
The study analyzed released event marks and channel metadata from OpenNeuro ds005398, version 1.1.3.
doi:10.1111/epi.18545

Cited by

3 publications

Public catalog entries on Sciscape.

usesA trained rate correction from other shafts increased mean next-minute MNI forecast loss in 18 SEEG sleep records
This publication supplied fixed public inputs, source locks, extraction code, the variance rule, and the own-history forecast. We rebuilt the counts and exactly matched its three baseline scores before testing the added feature.
usesA one second history term improved mean time scores for 50 ms joined MNI marks in the second half hour of 18 SEEG sleep records under a 60 second rate control
This related study supplied the scope for comparison with earlier one-minute HFO count forecasts in the same patients. We used no code or score table from it. The new study tests event time scores beyond a specified slow-history model.
usesA fixed local update from five minutes reduced mean Poisson loss for MNI HFO forecasts in 18 SEEG sleep records with equal predicted totals
This earlier result on the same cohort informed the fixed history rule of five minutes. Its fixed-baseline comparison did not isolate channel shares. The present study adds controls with equal predicted totals. It uses no numerical output or forecast code from that study.