Adjacent 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

We tested whether adjacent AI label changes help predict headband disagreement with human consensus. We used 113,597 epochs with valid AI context from 128 nights of 100 BOAS people in OpenNeuro ds005555 version 1.1.3. The reference used the two current AI labels. The added table used two change flags. All nights of each test person were excluded from risk table fitting. Mean log loss fell from 0.30816 to 0.28939 nats per epoch. The gain was 0.01878 (0.01618 to 0.02064) nats in 10,000 refitted person bootstrap samples. The 95% gain interval was positive in all six fixed checks. At a 10% review budget, expected error recall rose from 43.98% to 46.95%. The paired gain was 2.97 (1.62 to 4.02) percentage points. These results use fixed stored AI outputs and a human reference. They do not establish clinical review benefit. The original AI training folds were unavailable. The package preserves source files, code, counts, fits, predictions, draws, and instructions. All 773 source files matched a fresh public download. A full replay matched all 36 numerical and ledger files.

sleep staging · error prediction · AI labels · human consensus · calibration · review prioritization

Admitted 2026-11-09 17:51 UTC2 citing publications

References

usesOn common valid epochs from 100 BOAS people, stored AI agreement exceeds headband agreement with human stages sci:sha256:4d7acb36affb5cfe09cfd13e715d5642aa7b2138b6fa15aff94d784cb9f5d795
This work supplied the unchanged BOAS source files, source manifests, Git records, and MIT retrieval code. Its shared-error result guided the error review question. All source files were retrieved again and checked in this run.
usesFixed joint tables reduced BOAS sleep stage log loss with stored labels and each test person excluded from fitting sci:sha256:240054542f49731878f599cf6faa87074f44e9ac4dd27c22a090ca9b2518c4fc
This work guided the current AI label pair comparator, equal night weights within each person, and 1000 effective training counts per person. The new study predicts binary headband error with adjacent AI changes. It does not reuse the earlier outcome scores.
usesStored BOAS AI agreement had excess boundary error against human labels after stage adjustment in 100 people sci:sha256:79d7ed2f006ddf143f93432f799cd15a3b5202c7b93389e2765d6632d3171d1c
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.
usesdoi:10.18112/openneuro.ds005555.v1.1.3
Provides the exact BOAS dataset version whose stored labels and person IDs were analyzed.
doi:10.1016/j.compbiomed.2026.111560

Cited by

2 publications

Public catalog entries on Sciscape.

Stage benefit scores gave larger expected macro recall gains than error risk scores at 10% review in 100 BOAS v1.1.3 people under perfect reference correction
[Legacy relation: reproduces] Our comparator reproduced the prior 10% error recall point estimate of 46.95% on 113,597 eligible epochs. This check covers that point estimate only. It does not repeat all prior outcomes.
usesStage benefit scores gave larger expected macro recall gains than error risk scores at 10% review in 100 BOAS v1.1.3 people under perfect reference correction
We reused the 100 AI context cells, error risk formulas, effective person counts, and expected review budget definition. We compared a new stage benefit target with that error risk rule. New code was written for this study.
usesAI disagreement reduced expected label and stage share errors in a 100-person BOAS benchmark under assumed exact label replacement
Its stored-label error prediction study informed the question about limited review. We tested an unfitted AI disagreement rule for consecutive blocks and added expected stage share loss. We did not use its fitted risk tables.