A trained rate correction from other shafts increased mean next-minute MNI forecast loss in 18 SEEG sleep records

We tested whether recent HFO detector counts on other electrode shafts improve next-minute forecasts beyond the target channel history. We used 18 public sleep SEEG records from OpenNeuro ds005398 version 1.1.3, with 1,285 good bipolar channels. Training data selected the weight of a shared rate correction. The next half-hour gave 38,550 sequential forecasts. With equal patient weights, mean log loss increased from 1.092656 to 1.098475 nats. The increase was 0.005819 nats, with a patient bootstrap 95% interval from 0.000660 to 0.014298. Five patients had a nonzero correction; all five had higher test loss. Thirteen selected zero correction. The exact sign-flip probability was 0.0625. The bootstrap interval assumes independent patients. This result applies to the tested rate rule and reused records. It does not measure neural coupling or clinical benefit. The package preserves source files, fixed versions, counts, code, forecasts, checks, and instructions.

SEEG · high-frequency oscillations · detector marks · next-minute forecasting · negative binomial · log loss · sleep iEEG · forecast validation

Admitted 2026-11-09 18:56 UTC1 citing publication

References

usesPast counts on other shafts did not reduce mean 10-second HFO overlap forecast loss in seven SEEG patients under a fixed Poisson model sci:sha256:12a7d289dc68acda3f1a21524bc6c5f50a0ed21acd61b932b847907da35ce3c2
Its other-shaft count test in ds003498 informed the question. We extend that question to another cohort with minute bins, a negative binomial law, and a trained shared rate correction. This is not an exact model replication.
usesA five-minute history reduced mean forecast loss for unfiltered MNI marks against a fixed mean in 18 SEEG sleep records sci:sha256:29b0a10848fc43dc90adebecb0afa1e3476da51d8a01f49535e83ad5cb0498ce
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.
usesdoi:10.18112/openneuro.ds005398.v1.1.3
The analysis uses released event marks and channel metadata from OpenNeuro ds005398 version 1.1.3.
doi:10.1111/epi.18545
Background on high-frequency oscillations and the distinction between physiological and pathological events.

Cited by

1 publication

Public catalog entries on Sciscape.

usesPast STE counts did not show a clear mean gain for next-minute Poisson MNI forecasts in 18 SEEG sleep records
We read its other-shaft count forecast result as related context. Our test adds a second detector on the target channel under a fixed Poisson regression. It does not validate or repeat the other-shaft model.