Rule-Based Sleep Staging Achieves 60.5% Agreement with Human Scorers
A new deterministic, rule-based technique for automated sleep staging has been introduced as a clear alternative to complex deep learning systems. This method translates the scoring criteria of the American Academy of Sleep Medicine (AASM) into executable code and offers natural-language explanations at the epoch level through an explanation trace. When tested on 50 polysomnography recordings, it reached an agreement of 60.5% with a reference based on a 10-scorer majority-vote consensus (κ=0.42). In a dataset utilized during its development, the agreement was significantly higher at 77.1% (κ=0.61). This approach emphasizes interpretability and compliance with clinical scoring guidelines rather than relying on opaque performance metrics.
Key facts
- Method is deterministic and rule-based, operationalizing AASM scoring logic as executable code.
- Provides epoch-level natural-language justifications from an explanation trace.
- Evaluated on 50 polysomnography recordings with 10-scorer majority-vote consensus.
- Overall agreement with majority-vote reference: 60.5% of epochs (κ=0.42).
- On development dataset, agreement reached 77.1% (κ=0.61).
- Deep learning methods dominate current automated sleep staging research.
- Machine learning models achieve near-human level agreement but are opaque.
- Proposed method is designed to follow clinical scoring rules explicitly.
Entities
Institutions
- American Academy of Sleep Medicine