The 95% interval included zero for one-record calibration gain over a population table in 20 BOAS people with a fixed 50-epoch prior

We tested whether one labeled sleep record improved stage probabilities in other records from the same person. We used stored BOAS labels from OpenNeuro ds005555 version 1.1.3. Eighty people supplied a population table. Twenty other people supplied one calibration record each and 28 evaluation records. A fixed prior used 50 pseudo epochs per input-stage row. Mean log loss was 0.50904 for the population table and 0.50489 for the personal table. The gain was 0.00415 nats. Its 95% person bootstrap interval was -0.01990 to 0.02720. All six planned check intervals also included zero. Same-person calibration had lower loss than shuffled donor calibration, with Monte Carlo p=0.0001 under donor exchangeability. That comparison did not establish a gain over the population table. The intervals keep fitted tables fixed. Record dates and original AI training folds are unavailable. The package includes 773 label and metadata source files, 128 scan tables, hashes, a local fixed plan, code, probabilities, scores, bootstrap draws, and figures. Independent checks passed. A fresh-source run reproduced all 17 analysis files byte for byte.

sleep staging · personal calibration · probability table · log loss · bootstrap · BOAS · OpenNeuro