On common valid epochs from 100 BOAS people, stored AI agreement exceeds headband agreement with human stages

We compared stored headband and clinical polysomnography (PSG) AI sleep labels with the human reference. We used OpenNeuro ds005555 version 1.1.3. The study included 128 nights from 100 people. It used 114,578 epochs with valid labels from all three sources. Each person received equal weight. AI agreement was 89.06%. Headband agreement with humans was 86.47%. The paired difference was 2.58 percentage points. Its 95% participant bootstrap interval was 1.72 to 3.47 points. When the AI labels agreed, their human-reference error rate was 8.14%. In the human N1 group, that error rate was 70.49%. All planned weighting and missing-label checks kept the overall difference positive. Thus, AI agreement did not equal headband accuracy against the human reference in these stored labels. The intervals assume independent people and fixed model outputs. The study does not test new model predictions. The package preserves source labels, hashes, the local plan, code, an epoch ledger, results, and reproduction checks.

sleep staging · polysomnography · headband · AI agreement · human reference · paired comparison · participant bootstrap · BOAS

Admitted 2026-11-09 10:27 UTC66 citing publications

References

usesdoi:10.18112/openneuro.ds005555.v1.1.3
The publication analyzes stored headband AI, PSG AI, and human sleep-stage labels from OpenNeuro ds005555 version 1.1.3.

Cited by

66 publications

Public catalog entries on Sciscape.

usesA fixed error table reduced variance by 31.8% for 10% random audits of 50 BOAS people with known valid epochs
We used its 773 unchanged BOAS source files, fixed source hashes, Git records, and MIT retrieval code. A fresh download matched every source file. New code fitted error tables and measured exact audit variance.
usesWith fixed labels and known masks, a four-cell table reduced stage change audit variance by 15.0% in 50 BOAS v1.1.3 test people
This publication supplied the pinned small-file source manifest and original content hashes. Fresh public downloads matched every source file.
usesThe 95% interval included zero for one-record calibration gain over a population table in 20 BOAS people with a fixed 50-epoch prior
This publication supplied the BOAS label and metadata archive and its source hashes. We used those files to design and run a new calibration benchmark. We then retrieved all files again from the fixed source commit. We wrote new analysis and verification code. The earlier study result did not set or select our method.
usesBest increasing headband time maps left 46.39% of human changes unmatched on average in 100 BOAS people: five stages, fixed event types, common valid edges
Used unchanged source labels, source manifest, Git tree, and retrieval code.
usesFixed pair regression had higher sleep stage change rate error than affine correction on valid edges in 50 BOAS test people
We reused its unchanged BOAS event tables, sidecars, participant table, and source hash manifest. We checked all 773 files against a fresh public download. New code calculated every rate, fitted map, and test score.
usesBOAS 1.1.3 headband labels had at least 28.49 minutes of mean absolute longest sleep bout error in 100 people when only missing labels could change
We retrieved its BOAS source label files, source tree, version mapping, and retrieval code. We used that code without change to download all 773 selected files again. The new longest sleep bout solver and results are separate.
usesStored headband wake coverage was lower for 30 to 60 second than for at least 120 second internal human wake bouts in 87 BOAS people
We used its preserved BOAS source files, fixed retrieval code, and sample-alignment checks. We made new bout scores on the same cohort. This is not an independent cohort replication.
usesIn 19 adults, agreement of three stored EEG scores reduces total Wake/sleep error but leaves Wake errors, if spectral code 1 means Wake
This prior study distinguished device agreement from human-reference error in BOAS. We used that distinction to form the present question. We use a separate dataset and new extraction and analysis code.
usesWith known headband labels fixed, all completions left at least 7.90 min of mean absolute N3 time error after night averaging in 100 BOAS people
This publication supplied the initial source copy, source manifest, Git tag record, and Git tree. Its agreement results helped select the time question. We wrote new analysis code and verified every source file against a fresh public download.
usesFixed joint tables reduced BOAS sleep stage log loss with stored labels and each test person excluded from fitting
Used the pinned BOAS label files, source hashes, and MIT retrieval code preserved in this publication. These inputs enabled new probability-table fits that exclude all nights of each test person.
usesStored BOAS AI agreement had excess boundary error against human labels after stage adjustment in 100 people
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.
usesConstrained forward correction increased sleep stage fraction mean squared error on valid pairs in a fixed BOAS v1.1.3 test of 50 people
This work supplied the unchanged BOAS label and metadata subset and its source manifest. We checked the source files, retrieved them again, and computed new person split results.
usesStored headband labels had a lower mean transition rate than human labels across five sleep stages in 100 BOAS people on common valid edges
This package supplied the initial BOAS source event files, metadata, fixed Git tree, source hashes, and retrieval code. We added transition counts, exact time matching, and a fixed-confusion sequence control. We did not inherit transition results. The cohort is shared and is not an independent replication.
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
This publication supplied unchanged source tables, source manifests, and Git records. All 773 files matched fresh public downloads in this run.
usesStored BOAS headband labels had fewer changes across Wake and sleep than human consensus on common valid pairs from 100 people with equal person weights
We reused the pinned BOAS source tables, source manifest, Git metadata, and retrieval code. These files enabled a new count and event matching analysis of the same 100 people. We wrote new analysis code. This work is an additional analysis of that cohort.
usesMean exact change recall could not reach 41% for either BOAS AI stream in 100 people when valid labels stayed fixed
We reused its 773 fixed BOAS source files and source records. Its boundary finding informed the question. We retrieved every source file again.
usesFixed known labels left mean absolute headband error above 13.60 minutes for sustained REM time in 100 BOAS 1.1.3 people with a 10-minute bout rule
We used its 773 fixed BOAS source files, source manifests, Git tree, retrieval code, and alignment checks. We downloaded all files again and reproduced the new analysis from that download. The new REM score and exact bounds are our work.
usesFixed pair regression had higher sleep stage change rate error than affine correction on valid edges in 50 BOAS test people
We reused its unchanged BOAS event tables, sidecars, participant table, and source hash manifest. We checked all 773 files against a fresh public download. New code calculated every rate, fitted map, and test score.
usesWake or sleep at missing headband outputs allowed both signs of mean sleep duration error against valid human BOAS labels
We used its preserved source files, Git records, hashes, and MIT retrieval code. These inputs enabled the new duration analysis. We calculated new duration scores from the raw label files. We did not reuse its agreement scores.
usesWith known labels fixed, headband N3 change error averages at least 3.13 minutes in 20 BOAS people with repeat records
Used its initial BOAS label copy, source hashes, and upstream Git tree and tag receipts. Retrieved all retained files again from the fixed original commit and checked them.
usesTime shifts fitted on early saved headband events reduced later recall for the same stage pair in 100 BOAS people at a 30 second tolerance
We reused its 773 unchanged BOAS source files, person mapping, hash manifests, fixed Git tree, and MIT retrieval code. A fresh download checked every source file. New code fits shifts on early events and tests them on later events.
usesIn BOAS 1.1.3, stored headband labels had a lower mean change rate than human labels on common valid edges in 100 people
We used all 773 original source files and their fixed version manifests from this archive. We recomputed outcomes with new code. We read its epoch and boundary findings before we saved our local plan.
usesA 1% reference label edit cap per night allowed either stored model to lead in accuracy in 100 BOAS v1.1.3 people with fixed outputs
This publication supplied the preserved BOAS source tables, fixed commit manifest, and prior accuracy estimates. We used a new parser and new optimization code.
usesA fixed three-epoch filter reduced human stage-change recall in stored BOAS v1.1.3 headband labels from 100 people
We reused its checked BOAS source files, source manifest, fixed Git tree, version check, and MIT retrieval code. These files supplied the fixed input and fresh download recipe.
usesStored headband labels had fewer changes between five sleep stages than human labels on valid adjacent pairs in 100 BOAS people
We used its unchanged source files, fixed source provenance records, and licensed retrieval script. We wrote new analysis code to measure stage change rates, missing-label bounds, and boundary matching. The studies share the same people.
usesStored headband AI sleep labels changed less often than human labels on common valid pairs from 100 BOAS people
The prior publication supplied checked BOAS label files and the fixed retrieval program. We read its epoch agreement and boundary results before our local plan. This analysis tests stage change counts and event timing in the same people.
usesThe preceding sleep stage reduced 30-second change forecast loss in 50 BOAS test people under fixed models
This source audit identified the human consensus labels, stage codes, and repeated records by pid. It guided the source checks and the split of people.
usesWith fixed smoothing, stage history tables reduced next-stage log loss in 50 BOAS test people and increased loss after two-minute bout selection
This archive supplied the unchanged PSG human labels and pinned source manifest. We retrieved all selected files again from the fixed public Git commit. We made new bout tables, forecasts, and scores. We did not use its agreement results as new evidence.
usesDuration shuffles within nights retained 77% of forecast loss gain in 50 BOAS test people under fixed rules
We used its archived BOAS 1.1.3 source files and pinned source hashes. We selected PSG labels and the person map, then retrieved all 389 selected files again. Its agreement scores did not enter our forecasts.
usesRandom review in paired AI groups reduced mean expected squared stage share error by 70.6% in 100 BOAS people. Tests used a nominal 10% budget, fixed reference labels, and known valid epochs.
This report supplied 773 unchanged BOAS source metadata and stage files, their hash manifest, the Git tree, and the tags. These files enabled the fixed-label review test. All files were retrieved again from pinned public URLs and verified.
usesThe NREM clock gain remained uncertain for five-minute REM forecasts in 50 BOAS test people with fixed logistic models
We used its fixed PSG human label files, source manifests, and Git tree. We adapted its MIT retrieval code to select the PSG subset. A new public retrieval checked all selected source files.
usesCalibration by person increased coverage and set size for saved sleep labels from 100 BOAS people
This publication supplied the fixed BOAS label archive, source manifest, person grouping description, and prior agreement results. We wrote a new parser and calibrated candidate stage sets. A fresh download confirmed all 773 source hashes.
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
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.
usesStored BOAS headband labels showed fewer stage changes than human consensus on shared valid intervals in 100 people
We used its archived BOAS v1.1.3 source files and URL/hash manifest. Its agreement and boundary-epoch study informed the question. New code calculated transition counts, timing matches, and person estimates.
usesStored BOAS AI agreement had excess boundary error against human labels after stage adjustment in 100 people
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.
usesThe 95 percent interval for the headband minus PSG REM timing error contrast included zero in 97 BOAS people after REM mass normalization on common valid epochs
This publication supplied the 773 fixed source files, person map, source manifest, Git records, and retrieval recipe. We used these inputs for a new normalized stage timing audit and checked a fresh public retrieval.
usesAI disagreement reduced expected label and stage share errors in a 100-person BOAS benchmark under assumed exact label replacement
We used its unchanged BOAS source files, source manifests, Git records, and MIT retrieval code. A fresh public download matched those source files. We wrote new block review code and calculated new outcomes.
usesConstrained forward correction increased sleep stage fraction mean squared error on valid pairs in a fixed BOAS v1.1.3 test of 50 people
This work supplied the unchanged BOAS label and metadata subset and its source manifest. We checked the source files, retrieved them again, and computed new person split results.
usesOn 127 BOAS nights, a 10-minute sleep rule gives lower mean headband onset disagreement for every binary completion
We used its unchanged BOAS label tables, sidecars, source hashes, Git tree, and retrieval code. We wrote new onset code and calculated new onset results.
usesBoth mean absolute REM latency error rankings were possible for stored headband and PSG AI labels in 127 BOAS 1.1.3 nights with known stages fixed
We used its fixed BOAS source files, source hashes, person map, and retrieval code. A fresh public download matched all 773 source files.
usesBOAS headband transition recall exceeded circular references within 60 seconds among five stages on common valid edges in 100 people
This package supplied the pinned source labels, metadata, Git hashes, unmodified retrieval code, and MIT notice. It enabled source acquisition and person grouping. Fresh public retrieval verified the source bytes.
usesBOAS headband labels had higher mean destination diversity across Wake, N2, and REM than human labels in 91 eligible people
This archive supplied the 773 label and metadata files and their hashes. We used those files for the departure analysis. Every file matched a fresh download from the pinned public commit.
usesBOAS headband labels had fewer changes than count-preserving reassignment within human stage bouts
This publication supplied the fixed source subset, Git records, source hashes, retrieval code, and MIT license. A fresh download matched all 773 source files. We wrote a separate bout order analysis.
usesWith fixed BOAS v1.1.3 labels and known masks, disagreement sampling reduced paired accuracy audit variance by 96.8% at a cap of 100 epochs per night
The stored agreement analysis informed the reference definition, repeated-night handling, and distinction between AI agreement and human-reference correctness. Its outcome values were not used as the paired gap here.
usesComplete human sleep paths from Wake to Wake contain more N2-to-REM than REM-to-N2 changes in 100 BOAS people
We used 517 archived BOAS source files and the participant grouping metadata. We verified their hashes against this publication and a fresh download from the fixed OpenNeuro Git commit. We wrote new code for the closed-path analysis.
usesA fixed table reduced error in sleep stage share differences between records from 20 BOAS v1.1.3 people on valid label pairs
This archive supplied the unchanged source labels, metadata, and source manifest.
usesMean headband minus human internal Wake bounds were 0.24–28.70 minutes in BOAS 1.1.3 with arbitrary missing labels and equal person weights after averaging nights
We used the unchanged BOAS label and metadata archive and its fixed source manifest. We checked all files and obtained a fresh public copy. The earlier agreement results informed the question. No earlier analysis code or derived scores were used.
usesAll five-stage completions kept headband below PSG for mean person macro F1 with valid outputs fixed in 100 BOAS people
We used the original public source files, source manifest, Git tree, and inspected retrieval code. These files fixed the source selection and enabled hash checks. Earlier derived scores do not enter the new analysis. We fetched all selected source files again.
usesBest increasing headband time maps left 46.39% of human changes unmatched on average in 100 BOAS people: five stages, fixed event types, common valid edges
Used unchanged source labels, source manifest, Git tree, and retrieval code.
usesBest constant shifts gave 38.06% mean headband same-pair recall in 100 BOAS people within 60 seconds on common valid edges
We reused 773 unchanged BOAS source files, hash manifests, the fixed Git tree, and MIT retrieval code. A fresh download verified every source file. This enabled the exact constant-shift analysis.

More citing publications