A fixed three-epoch filter reduced human stage-change recall in stored BOAS v1.1.3 headband labels from 100 people

We tested one fixed three-epoch categorical filter on stored BOAS v1.1.3 sleep labels. The shared valid sample had 114,578 epochs from 128 nights and 100 people. Each person had equal weight. Headband human stage-change recall fell from 33.069% to 26.220%. The change was -6.849 percentage points, with a 97.5% bootstrap interval from -7.770 to -5.973 points. Epoch agreement changed by +0.002 points. Its interval included zero. The filter corrected 501 labels and spoiled 503. It removed 640 exact human-change matches. Planned PSG, width, weighting, and missing-output checks also showed recall loss. The package preserves fixed source bytes, code, all outputs, all bootstrap draws, a proof, and reproduction instructions. Fresh source retrieval and independent checks passed. The result applies to this added offline filter and these stored outputs. Human consensus is the reference. The intervals assume independent people and omit model training uncertainty.

sleep staging · BOAS · OpenNeuro · stage-change recall · categorical filter · human consensus · headband EEG · polysomnography · bootstrap