HFO overlap labels had lower prediction loss within records than between records at equal training counts in 135 SEEG sleep records from seven patients

We tested whether released HFO overlap labels transfer between sleep records. We used 135 records of 300 seconds from seven SEEG patients in OpenNeuro ds003498 version 1.1.1. Each ripple mark had a binary label for fast ripple overlap. Both probability estimates used the same number of training marks. We computed expected Brier loss over all uniform training subsets without replacement. Mean loss was 0.083214 within records and 0.090975 between records, with equal patient weights. The difference was 0.007762. Its patient bootstrap 95% interval was 0.002305 to 0.016625 under independent and exchangeable patient effects. All seven effects and all six planned checks favored estimation within a record. This result shows limited transfer of the tested detector label fractions. It does not establish clinical accuracy or a neural cause. The package preserves source tables, hashes, code, scores, a loss proof, figures, and repeat instructions.

high frequency oscillations · intracranial EEG · epilepsy · Brier loss · within-record validation · event overlap