Observed mean test loss was higher for three-shaft MNI HFO fits than for pair fits in 18 sleep SEEG records with 1 s bins

We tested an added three-shaft term for HFO activity in public sleep SEEG data. We used MNI detector marks from 18 patients in OpenNeuro ds005398 version 1.1.3. We used 1 s bins and all 4,214 shaft triples. The first half-hour fitted each model. The next half-hour tested it. Each eight-state training table had a 0.5 cell pseudocount. The pair maximum entropy model kept all one-shaft and two-shaft margins of that table. Mean test loss was 1.130028289 nats for the pair model and 1.130102867 nats for the three-shaft model. Gain is pair loss minus three loss. The equal-patient mean gain was -0.000074578 nats per triple per bin. Its patient bootstrap 95% interval was [-0.000421963, 0.000335143]. The exact two-sided patient sign-flip p value was 0.729881. Five patients had a gain. The observed mean loss was higher with the added term. The interval includes a group gain and a group loss. Population use requires independent, exchangeable patients; the sign-flip null also requires joint sign symmetry. The package preserves the fixed plan, all source inputs, count tables, model probabilities, code, results, and instructions. All 94 fresh source checks passed. All 189 files in an offline repeat matched byte for byte. The test concerns released detector marks. It does not establish a distinct neural mechanism or clinical benefit.

high-frequency oscillations · stereoelectroencephalography · maximum entropy · held-out prediction · neural activity