Calibration for each observed sleep stage increased mean label set size in a 30-person BOAS test

We tested candidate sleep stage sets in saved BOAS labels from OpenNeuro ds005555 version 1.1.3. The source has 128 nights from 100 people. We retained 114,578 epochs with valid human and both AI labels. Forty people supplied the score fit, thirty supplied calibration, and thirty supplied the primary test. The new target was at least 90% coverage for each human stage present in a person’s retained records. The common score rule met this target for 29 of 30 test people, with a mean set size of 3.867 stages. A reference rule for total person coverage met the stage target for 1 of 30 people, with a mean size of 1.589. Stage scales fitted on separate people gave 28 of 30 successes and a mean size of 3.942. The common minus scaled size difference was -0.075 stages, with a paired 95% bootstrap interval of -0.130 to -0.020. Across 1,000 allocations of the same cohort, mean size was 3.852 for common scores and 4.118 for scaled scores. With the primary fit fixed, exact expected success under random calibration assignment was 95.00% and 91.13%. These are results for saved records. A new-person probability bound requires exchangeable complete records conditional on the fit. We apply a standard rank rule and do not claim a new general theorem. The package preserves source data, code, the local plan, all fits and results, a proof, and instructions. Separate code checked every repeated fit and all planned sensitivity tests. Fresh retrieval matched all 773 source files. A full replay matched all 30 result files byte for byte.

sleep staging · conformal prediction · calibration · coverage · uncertainty sets · BOAS · polysomnography · machine learning