A fixed Markov filter increased probability loss versus a static table for stored BOAS v1.1.3 headband labels in 50 test people

We tested whether a fixed Markov filter could improve probabilities from stored BOAS v1.1.3 headband labels. We fitted the rule on 50 people and tested it on 50 different people. All nights of each person stayed in one set. The static comparison used the same training labels and count smoothing. Mean log loss increased from 0.537 to 0.691 nats per valid epoch. The paired increase was 0.153 nats, with a 95% person bootstrap interval of 0.093 to 0.222. Overall stage accuracy stayed near 85.85%. All full-transition fitting checks increased loss, including the reversed person split. The forward-backward model increased loss by 0.372 nats and reduced brief Wake coverage. Additional checks showed excessive confidence and adjacent label dependence. These checks were added after the primary results were known. They do not prove the cause of the loss increase. Independent calculations and a complete fresh-source repeat passed. The package preserves source files, hashes, a local plan, reusable code, fitted models, predictions, person scores, Wake bouts, bootstrap indices, selected draws, and figures. Intervals assume independent test people and fixed fitted models. The original AI folds are unavailable. This result applies to the tested rule on stored labels. It does not establish clinical or new-cohort performance.

sleep staging · Markov model · hidden Markov model · probability calibration · log loss · wearable EEG · person-level holdout · BOAS