Calibration by person increased coverage and set size for saved sleep labels from 100 BOAS people

We tested candidate sleep stage sets with 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. All nights from each person stayed together. Forty people supplied the fit, thirty supplied calibration, and thirty supplied the primary test. Both rules used the same fitted probability table. Person calibration gave at least 90% coverage for 27 of 30 test people. Pooled epoch calibration gave 17 of 30. Mean set size increased from 1.05 to 1.59 stages. Across 1,000 allocations, mean target success was 91.61% versus 65.85%. Human N1 coverage remained 70.42% in the primary equal person average. With the primary fit fixed, exact random-assignment success for the remaining saved records was 91.17%. This result concerns saved records. A guarantee for new people requires exchangeable complete person records. We apply a standard conformal rank rule. The package preserves source files, hashes, the local plan, code, all random draws, fits, results, a proof, and instructions. Fresh retrieval matched all 773 source files. A full replay matched all 23 result files byte for byte.

sleep staging · conformal prediction · uncertainty sets · calibration · coverage · polysomnography · wearable EEG · BOAS