Exact Gamma product predictions reduced later one-minute MNI HFO count loss in 18 SEEG records under a fixed fit rule

We tested exact count probabilities for released MNI HFO marks in 18 public SEEG records from OpenNeuro ds005398. The comparison used a fixed fit rule and a negative binomial approximation with the same mean and variance. Models fitted the first 30 minutes and scored 38,550 channel bins in the next 30 minutes. Equal-patient mean loss fell from 1.177612 to 1.169308 nats per channel-minute. The mean gain was 0.008303 nats. The 95% patient bootstrap interval was 0.000915 to 0.022025. This interval assumes independent, exchangeable patients for population use. Fifteen patients had positive gains. UCLA49 supplied 80.43% of the summed gain. The mean gain without that patient was 0.001721. All seven fixed secondary checks retained positive mean gains. The package preserves source inputs, exact probability code, a moment proof, every prediction, numerical checks, and repeat instructions. All 94 fresh public source checks passed. All 22,507 distinct probabilities passed a second integral check. A full offline repeat matched 54 generated files byte for byte. This is a local test in a reused cohort. It does not establish event validity or clinical benefit.

Gamma product · negative binomial · count prediction · high frequency oscillations · stereoelectroencephalography · marginal probability · held-out prediction · patient bootstrap