Fixed joint tables reduced BOAS sleep stage log loss with stored labels and each test person excluded from fitting
We tested whether a stored headband sleep label added information to a stored polysomnography (PSG) label. We used OpenNeuro ds005555 version 1.1.3. The sample had 100 people, 128 nights, and 114,578 common valid epochs. Fixed probability tables excluded all nights of each test person. Each training person supplied 1000 effective epochs. Each table cell received a smoothing count of one. The joint table reduced mean human-reference log loss from 0.4891 to 0.4145 nats per epoch. The mean gain was 0.0746 nats. Its 95% refitted person bootstrap interval was 0.0618 to 0.0872. All four planned sensitivity checks retained positive gain intervals. 98 people had a positive observed gain. The accuracy gain was 0.14 percentage points. Its interval included zero. Joint-table N1 recall was 18.56%, compared with 35.71% for stored PSG labels. The N1 difference interval included zero. Lower log loss did not establish better stage accuracy. The interval assumes independent people and fixed stored outputs. Original AI training folds were unavailable. Both device labels were required. The package preserves source files, hashes, code, every prediction, refitted draws, and reproduction checks.
sleep staging · probability calibration · log loss · polysomnography · headband · human reference · person-level bootstrap · BOAS
References
2- usesOn common valid epochs from 100 BOAS people, stored AI agreement exceeds headband agreement with human stages sci:sha256:4d7acb36affb5cfe09cfd13e715d5642aa7b2138b6fa15aff94d784cb9f5d795
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.
- usesdoi:10.18112/openneuro.ds005555.v1.1.3
Used the pinned BOAS label files for this analysis; the exact version 1.1.3 dataset metadata identifies the source and its CC0 license.
Cited by
8 publicationsPublic catalog entries on Sciscape.
- usesWake or sleep at missing headband outputs allowed both signs of mean sleep duration error against valid human BOAS labels
Its summary showed a paired comparison of the stored PSG and headband labels. This informed our paired-device duration question. We did not use its fitted probability tables.
- usesFixed ridge calibration reduced mean squared error of sleep stage shares compared with raw labels on valid pairs in 100 BOAS adults
We used its unchanged BOAS source files, source hashes, fixed Git tree, version tags, and retrieval code. Its epoch table results informed the question. We computed new share fits and scores.
- usesElapsed headband label duration reduced mean sleep stage log loss under fixed tables in 50 BOAS v1.1.3 test people
We read the joint device label table report before the local plan. It informed the probability table design. The present rule uses current headband labels and their elapsed duration.
- usesDuration shuffles within nights retained 77% of forecast loss gain in 50 BOAS test people under fixed rules
Its fixed label tables supplied a precedent for equal person weight and 1000 effective training counts per person. We used new history tables and a separate 50/50 person split to test next-stage forecasts.
- 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 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.
- usesNearby stored AI labels reduced BOAS sleep stage probability loss in offline tests with each test person excluded from table fitting
This work supplied the initial BOAS source files, fixed retrieval code, source hashes, Git records, MIT license, and current-label probability table idea. We refitted the current-label comparator on the new retained subset and wrote a new nearby-label analysis.
- usesA fixed Markov filter increased probability loss versus a static table for stored BOAS v1.1.3 headband labels in 50 test people
We reused the archived BOAS v1.1.3 event files, source hash manifest, and fixed-commit retrieval code. Its static joint-table method informed the matched probability baseline. We wrote new code to test the Markov models on separate training and test people.
- usesElapsed headband label duration reduced mean sleep stage log loss under fixed tables in 50 BOAS v1.1.3 test people
We read the joint device label table report before the local plan. It informed the probability table design. The present rule uses current headband labels and their elapsed duration.