A negative binomial model gave higher test count scores than a Poisson model in six of eight sleep SEEG patients with equal model means

We tested count probability forecasts in 185 public sleep SEEG records from eight patients with temporal lobe epilepsy. The source marks are fast ripples that overlap ripples, labelled FRandR, in OpenNeuro ds003498 version 1.0.1. We fitted models on the first half of each record and scored five-second bins in the second half. Poisson and negative binomial forecasts had equal fitted means. A patient variance parameter used training counts only. The negative binomial model gave a higher test score in six patients and the same score in two. The equal-patient mean gain was 0.005692 natural log units per channel bin. The patient t-based 95 percent interval was [-0.001094, 0.012477]. A sign-flip value of 0.03125 requires unverified symmetric null differences. Empty bins supplied the gain; positive bins had a negative score contribution. The comparison with a gamma-Poisson model for uncertainty in the mean rate did not show a clear gain. A fresh download reproduced all source hashes. Separate code reproduced counts and scores, and a fresh process reproduced result bytes. The package preserves the fixed plan, source tables, code, results, figures, metadata conflict, and instructions. The result does not establish a neural mechanism, a new-patient benefit, or clinical outcome prediction.

stereoelectroencephalography · high frequency oscillations · count forecasting · negative binomial · Poisson · sleep · epilepsy · log score