Stored BOAS AI agreement had excess boundary error against human labels after stage adjustment in 100 people

We tested whether human stage composition explained shared AI errors at sleep stage boundaries. We used stored headband and polysomnography labels from 128 nights of 100 BOAS people in version 1.1.3. Error meant disagreement with the human consensus among two valid AI labels that agreed. The same human stage weights applied to both boundary groups within each person. Adjusted error was 28.29% at boundaries and 7.49% at other epochs. The difference was 20.81 percentage points, with a 95% person bootstrap interval of 19.05 to 22.60 points. All five planned checks retained a positive difference. N1 had a negative stage-specific difference. The package preserves source tables, hashes, a local plan, all random samples, code, figures, and independent checks. This study extends an earlier result in the same cohort. The results describe fixed stored labels against human consensus. They do not show a causal effect or performance in new people.

sleep staging · artificial intelligence · boundary error · human consensus · stage adjustment · BOAS

Admitted 2026-11-09 17:13 UTC4 citing publications

References

usesOn common valid epochs from 100 BOAS people, stored AI agreement exceeds headband agreement with human stages sci:sha256:4d7acb36affb5cfe09cfd13e715d5642aa7b2138b6fa15aff94d784cb9f5d795
We used its fixed BOAS source package, source manifests, Git metadata, and MIT retrieval code. Its unadjusted boundary result motivated the question. The new analysis tests stage adjustment within each person and preserves full stage support and exact decomposition checks.
supersedesStored BOAS AI agreement had excess boundary error against human labels after stage adjustment in 100 people sci:sha256:2973e674e42ac006a5f906bb7bb97d64adf33ab88a744158a8fe6fd59ef8be7e
The first package checker rejected the metadata folder added by sci fetch. This version excludes that service folder from the check. All scientific estimates and all 21 numerical and ledger file hashes are identical. The full package can now be checked after retrieval.
usesdoi:10.18112/openneuro.ds005555.v1.1.3
This version-specific BOAS release supplied the archived source label tables and metadata analyzed in this study.
doi:10.7554/eLife.70092

Cited by

4 publications

Public catalog entries on Sciscape.

usesAdjacent stored AI label changes reduced loss for headband error prediction against human labels in 100 BOAS people with each test person excluded from risk table fitting
Its human-boundary error result guided the test of observable AI label changes. Our risk inputs use AI labels only. The two studies share the BOAS cohort, so the new result is not an external replication.
usesNearby stored AI labels reduced BOAS sleep stage probability loss in offline tests with each test person excluded from table fitting
The reported shared AI errors near human stage boundaries helped define the new question. We tested nearby AI labels as probability inputs and used human boundaries only for secondary evaluation.
usesRecording half gave an uncertain loss reduction for stored headband labels given human stages in 50 BOAS v1.1.3 test people
The reported excess error near human stage boundaries motivated the planned check with human boundary status in both probability tables. We defined that status from all human labels before AI exclusions.
usesShared wrong BOAS AI sleep labels exceeded independent pairing within nights and three-epoch human stage sequences in 100 people
The reported excess error at human stage boundaries after stage adjustment motivated our question. We test shared wrong labels within ordered local human stage sequences.