Unsupervised Domain Adaptation by Statistics Alignment for Deep Sleep Staging Networks.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Sleep Medicine

Background:

• Deep learning models achieve high performance in sleep staging on large datasets.
• Model performance degrades on smaller sleep cohorts due to data inefficiency and domain shift.
• Transfer learning from large datasets to small cohorts is promising but challenging.

Purpose of the Study:

• To develop unsupervised domain adaptation methods to bridge the domain gap in sleep staging.
• To improve the performance of pre-trained sleep staging models on small, unseen datasets.
• To offer universally applicable solutions for deep sleep staging networks with Batch Normalization layers.

Main Methods:

• Domain Statistics Alignment (DSA): Modulates domain-specific statistics in Batch Normalization (BN) layers to adapt source models to target domains.
• Adaptive Domain Statistics Alignment (AdaDSA): Extends DSA by incorporating cross-domain statistics for adaptive alignment.
• Unsupervised approach requiring only the pre-trained source model, not the source data.

Main Results:

• DSA and AdaDSA significantly improved the performance of source models on target sleep staging datasets.
• Methods were validated on DeepSleepNet+ and U-time across six sleep databases, including clinical data.
• Demonstrated effectiveness in transfer learning tasks for sleep staging, addressing domain generalization.

Conclusions:

• DSA and AdaDSA effectively mitigate the domain-shift issue in sleep staging.
• These unsupervised domain adaptation techniques enhance model generalization to smaller datasets.
• Provides valuable insights and practical solutions for domain generalization in sleep staging research.