A fixed error table reduced variance by 31.8% for 10% random audits of 50 BOAS people with known valid epochs

We tested random audits of stored sleep stages in BOAS 1.1.3. We assumed a known mask of valid human epochs. An error table used both stored automatic stage labels. We fitted it on 50 people and tested it on 50 other people. Each audit selected about 10% of epochs uniformly within each test person. Both estimators used the same 5,894 selected epochs. The fixed table reduced exact sampling variance by 31.8% against the ordinary sample mean. Standard deviation fell from 0.477 to 0.394 percentage points. Variance fell for 48 of 50 test people. The reduction was 22.0% when both automatic labels were present. The difference estimator stayed unbiased for any fixed table under the stated selection design. All 773 source files passed hash checks. A fresh-source rerun matched all 12 main result files. The package contains source labels, code, an audit tool, a proof, results, and reproduction instructions. The findings apply to these fixed labels and people. Training costs, review time, and performance in new cohorts were not tested.

sleep staging · sleep scoring · random audit · difference estimator · sampling variance · finite population · headband EEG · polysomnography · BOAS · OpenNeuro