The preceding sleep stage reduced 30-second change forecast loss in 50 BOAS test people under fixed models

The preceding different sleep stage reduced loss for a change in the next valid 30-second human label. The test used BOAS v1.1.3, with 50 training people and 50 separate test people. The baseline used current stage, current bout age, recording time, and preceding bout duration. Adding preceding-stage indicators reduced weighted log loss from 0.226806 to 0.223873 nats. The gain was 0.002933 nats. The 95% refit bootstrap interval was 0.001613 to 0.003874. The main interval used 1,000 training and test person draws with full model refits. The interval assumes independent people. Each person, then night, then eligible epoch had equal weight. The gain was positive in 42 of 50 test people. Six fixed checks used 250 refit draws each after a documented run-budget change from 500. After omission of the first epoch of each bout, the refit interval included zero. N2 supplied 88.6% of the total gain. A complete repeat with a fresh source download matched all analysis results. The package preserves source labels and hashes, code, models, every draw, predictions, and an independent history and optimizer audit. The evidence applies to the fixed models and valid scored histories. It does not prove a neural memory process, a causal effect, or a gain for an online EEG decoder.

sleep stages · sleep forecasting · stage transitions · bout history · logistic regression · bootstrap uncertainty · BOAS