Time 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 tested a fixed time shift for saved headband sleep stage events in 128 BOAS records from 100 people. Shifts from minus 300 to plus 300 seconds were fitted on early events and applied to later events. A match required the same ordered stage pair within 30 seconds. Each event could be used once. Mean person recall decreased from 35.59% at zero shift to 33.49% after fitting. The paired change was minus 2.11 percentage points. Its 95% person bootstrap interval was minus 3.50 to minus 0.97 points. Recall decreased for 17 people, increased for 3, and stayed equal for 80. All nine planned sensitivity changes had negative mean gains. The interval at 60 second tolerance included zero. Even selecting the best shift on the test events gave only 36.85% recall. This is an exact bound for the tested 21-shift family on fixed events. The time split tests later events in the same records. The intervals assume independent people and fixed model outputs and human labels. They omit training and scoring uncertainty. The package preserves source labels, hashes, the saved plan, code, all candidate counts, person scores, resampling draws, independent checks, and instructions. A fresh source replay gave identical bytes for every numerical output file.

sleep staging · transition timing · wearable EEG · event matching · time correction · BOAS