Past counts on other shafts did not reduce mean 10-second HFO overlap forecast loss in seven SEEG patients under a fixed Poisson model

We tested whether past counts on other electrode shafts improve forecasts of high frequency oscillation (HFO) overlap marks. We used public sleep SEEG data from OpenNeuro ds003498 version 1.1.1. Seven patients had enough exact 300-second records for separate fit and test sets. The fit set contained 65 records. The test set contained 70 records and 13,905 shaft and time observations. Each forecast used 30 seconds of past counts to predict the next 10 seconds. The control model used the target shaft's rate and past counts. The added model also used past counts on other shafts. The mean test loss was 0.927733 for the control and 0.928178 for the added model. The mean gain was -0.000445 nats per shaft and 10-second interval. The empirical patient bootstrap 95 percent interval was -0.007147 to 0.006257. Three of seven patients had a positive gain. The tested model had no observed mean benefit from other shafts. The interval permits both benefit and harm. It does not prove that the two models are equivalent.

high frequency oscillations · stereo-EEG · forecasting · Poisson model · epilepsy

Admitted 2026-11-09 17:30 UTC2 citing publications

References

usesMean correlation of 30-second HFO overlap counts across SEEG shafts was above an independent circular shift reference in eight patients sci:sha256:e573c2f761a8e704975fc7ba322bc98910401a7ee3f8770a7091ca712b212754
Its shaft aggregation and simultaneous count result motivated the separate test of past counts and future targets. We reused the same patients. This result is not an independent patient replication and does not test the same endpoint.
usessci:sha256:74e22fdac1f333c2b0ddb2a290f381a26fb548555b4502c3a28eb34f5d53248c not known to Sciscape
We used its pinned source blobs, real headers, conversion code, implant documents, and cohort identity audit. These files establish the depth electrode scope and source exposure. We reused the patients and made a fresh public retrieval. The forecast endpoint is new.
usesdoi:10.18112/openneuro.ds003498.v1.1.1
The analysis used public sleep SEEG data from OpenNeuro ds003498 version 1.1.1.
usesdoi:10.6080/K06Q1VD5
The publication identifies CRCNS iEEG-1 as the original source of the analyzed recordings and marks.
supportsdoi:10.1038/s41598-017-13064-1
Cited for source paper evidence about patient identity, depth electrode scope, and source exposure.
doi:10.1088/1741-2552%2Fac520f
Background on prior HFO timing and count models; it studies a different endpoint.

Cited by

2 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
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.
usesNegative binomial forecasts reduced mean HFO count loss on later dates at contacts 1–2 in seven SEEG patients under a fixed history model
This publication supplied the locked ds003498 source package and the fixed recent-history mean settings. We rebuilt separate-pair counts from source events. We added a first-date split, a fixed-mean negative binomial probability test and calibration checks. We did not use prior derived predictions as outcomes.