A fixed pair update reduced cross-date HFO label loss by 3.0 percent over an additive shaft and contact order model in seven SEEG patients

We tested whether an electrode pair retains HFO label information across acquisition dates. We used 135 public SEEG sleep records from seven patients in OpenNeuro ds003498 version 1.1.1. Each target date was excluded from training. The label indicated fast ripple overlap on a ripple mark. A fixed pair update improved an additive shaft and contact order model. Mean Brier loss fell from 0.074581 to 0.072343. The absolute gain was 0.002237, or 3.00 percent of baseline loss. All seven patient gains were positive. The patient bootstrap 95 percent interval was 0.001020 to 0.003580. This interval assumes independent, exchangeable patients. The exact patient sign-flip p value was 0.015625 under independent, symmetric null effects. All twelve specified checks had a positive mean gain. Fresh retrieval of 1,131 source files reproduced all 42 main outputs exactly. The package preserves source bytes, provenance, code, predictions, figures, and repeat instructions. The result concerns released detector labels and the tested models. It does not establish a neural cause or clinical benefit.

high frequency oscillations · SEEG · cross-date prediction · Brier loss · electrode pairs