Stored headband label duration reduced 30-second human stage change forecast loss under fixed models in 50 BOAS v1.1.3 test people with both bout starts observed

We tested whether stored headband label duration improves a human stage change forecast at the next 30-second epoch. The source was BOAS v1.1.3. Both label bout starts had to be observed. We fitted models in 50 people and tested 49,271 epochs from 50 other people. The baseline included current stage, the preceding different stage, and recording time. Adding duration reduced headband forecast log loss from 0.27112 to 0.26012 nats per epoch, a 4.06% reduction. The gain was 0.01101 nats, with an approximate 95% person bootstrap interval of 0.00777 to 0.01535. All models were refitted in 2,000 paired draws. Human-history gain was 0.02140 nats. All five planned checks had positive overall headband gains. Removing age-1 epochs reduced the gain to 0.00307 nats. The headband N3 subset had higher loss after duration was added. Independent checks passed. A complete repeat from a fresh public download gave identical numerical files. The package preserves inputs, code, fits, predictions, draws, and instructions. Stored labels can use later signal data. Original AI scoring fold IDs were unavailable. The result does not establish a real-time EEG forecast or a neural mechanism.

sleep staging · sleep stage transitions · wearable EEG · duration modeling · probability forecast · log loss · BOAS