With fixed smoothing, stage history tables reduced next-stage log loss in 50 BOAS test people and increased loss after two-minute bout selection

We tested whether the preceding human sleep stage improves a forecast of the next different stage. We used BOAS, OpenNeuro ds005555 version 1.1.3: 128 records from 100 people. A fixed split used 50 people for fitting and 50 for testing. Both models used the current stage, time from record start, and preceding bout length. The history table also used the preceding stage. Mean log loss fell from 0.7976 to 0.7208 nats across 4,820 test origins. The gain was 0.0768 nats; the 95% refit-bootstrap percentile interval was 0.0531 to 0.0851. Loss fell in 49 of 50 test people. After selection of preceding and current bouts of at least two minutes, the gain was -0.0266 nats. That check had 554 test origins and used future labels for selection. A separate five-minute target retained a smaller log loss gain. Earlier research had shown sleep stage history effects. This work adds a public benchmark and shows a limit of the fixed model comparison. It does not establish a neural cause. The package preserves source hashes, code, forecasts, draws, all fixed checks, and a matching fresh-data rerun.

sleep stage forecasting · sleep staging · log loss · probability tables · person-level bootstrap · BOAS