Mean exact change recall could not reach 41% for either BOAS AI stream in 100 people when valid labels stayed fixed

We tested whether missing AI labels could explain low recall of human sleep-stage changes in BOAS version 1.1.3. We used 128 nights from 100 people. Every valid AI label stayed fixed. Each missing AI label could take any of five stages. On all 119,583 human-valid adjacent pairs, mean exact recall could not exceed 36.49% for headband AI or 40.62% for PSG AI. The human reference had 10,236 changes. Requiring the same directed stage pair reduced the maxima to 24.69% and 27.59%. These sample bounds are sharp: saved stage assignments reach each endpoint. A dynamic program, independent closed formulas, and exhaustive small cases agreed. A fresh source retrieval reproduced all numerical file hashes. The package also reproduces earlier count and event comparisons on the same people. The bounds do not allow changes to valid AI labels or human labels. They do not apply to a timing allowance or establish true physiological events. Code, source records, complete assignments, results, proofs, and reproduction commands are preserved.

sleep stages · sleep stage transitions · event recall · missing labels · polysomnography · BOAS · sharp bounds

Admitted 2026-11-09 17:13 UTC2 citing publications

References

Stored BOAS headband labels had fewer changes across Wake and sleep than human consensus on common valid pairs from 100 people with equal person weights sci:sha256:35ce3bc845344f7f6b06623912f069bf386ce095363a2726313a4b34bb2c69b9
[Legacy relation: reproduces] Our independently written count and event analysis gives the same estimates on the same BOAS people. This is a reproduction on one shared cohort.
usesOn common valid epochs from 100 BOAS people, stored AI agreement exceeds headband agreement with human stages sci:sha256:4d7acb36affb5cfe09cfd13e715d5642aa7b2138b6fa15aff94d784cb9f5d795
We reused its 773 fixed BOAS source files and source records. Its boundary finding informed the question. We retrieved every source file again.
usesStored headband labels had fewer changes between five sleep stages than human labels on valid adjacent pairs in 100 BOAS people sci:sha256:553da6e9b47d2bcc251b4a6ef9319c06ff74efde8d03265dac9af4dec90d2489
We read its change-count bounds after our first count results. This work helped select the added question about sharp exact event-recall bounds.
usesIn BOAS 1.1.3, stored headband labels had a lower mean change rate than human labels on common valid edges in 100 people sci:sha256:4cdd91247e1cc665001cb799c7431094be7c99b0f01b29f238ed6cf0e4d6ee2d
We read its change-count bounds after our first count results. This work helped select exact recall on all human-valid pairs as the added target.
usesdoi:10.18112/openneuro.ds005555.v1.1.3
The analysis uses the event tables and stored human and AI labels from BOAS ds005555 version 1.1.3; the version-specific DOI and CC0 license are recorded in the saved dataset metadata.
doi:10.1093/sleepadvances%2Fzpaf052
Background on how scoring differences affect sleep transition measures.

Cited by

2 publications

Public catalog entries on Sciscape.

usesMissing labels could not raise mean headband change recall above 54.76% within 60 seconds for five stages in 100 BOAS people when valid labels stayed fixed
Its exact-time missing-label bound defined the prior result that we extended to timing tolerance. We used its recording, person, and summary tables to check the shared zero-tolerance maximum.
usesAt most 5% unrestricted epoch errors per night permit 2.89–22.06 mean stage changes per valid edge hour in BOAS 1.1.3
We read its summary after the local plan to compare the scopes. Its recall bounds change only missing AI labels. Our bounds permit errors at any valid epoch within a fixed budget. This reading did not change the analysis.