Nearby stored AI labels reduced BOAS sleep stage probability loss in offline tests with each test person excluded from table fitting

Nearby stored AI labels reduced human-reference sleep stage probability loss after both current AI labels were known. We tested 113,597 epochs from 128 BOAS nights and 100 people. All nights of each test person were excluded from table fitting. Mean log loss fell from 0.41313450 to 0.39301898 nats per epoch. The reduction was 0.02011552, with an approximate 95% person bootstrap interval of 0.01644928 to 0.02107936. All six planned checks also gave positive loss reductions. Accuracy and stage recall intervals included zero change. N3 recall fell in the point estimate. The rule requires both devices and the next stored AI epoch. Original AI training folds are unavailable. This is an internal offline result on fixed outputs. The package preserves source files, code, predictions, fitted tables, bootstrap draws, provenance, and reproduction instructions.

sleep staging · probability calibration · log loss · cross-validation · wearable EEG · offline validation