Recording half gave an uncertain loss reduction for stored headband labels given human stages in 50 BOAS v1.1.3 test people

We tested whether recording half reduced loss for prediction of a stored headband AI label when the human sleep stage was known. We used OpenNeuro BOAS ds005555 version 1.1.3, with 128 recordings from 100 people. A fixed split used 50 people to fit both tables and 50 other people to test them. All recordings from a person stayed together. On 55,350 valid test epochs, mean person loss fell from 0.460243 to 0.453564 nats per epoch. The gain was 0.006679 nats per epoch, or 1.45 percent. The planned 95 percent refit percentile interval was -0.001584 to 0.011449. It includes zero. The planned basic interval was 0.001908 to 0.014942. It excludes zero. Thus, the conclusion depends on the interval method. The result does not establish constant error rates or a reliable time gain. All eight planned checks are reported. The analysis assumes independent people and conditions on stored labels. It does not include source classifier fitting or human scoring uncertainty. The package preserves 773 pinned source files, the local plan, code, tables, person scores, bootstrap draws, and instructions. Independent calculations passed. A fresh source replay matched all 11 result files exactly. This is an offline label audit, not a clinical sleep classifier.

label shift · sleep staging · confusion matrix · log loss · OpenNeuro · BOAS