Transferring Structured Knowledge in Unsupervised Domain Adaptation of a Sleep Staging Network

Area of Science:

• Neuroscience
• Computer Science
• Biomedical Engineering

Background:

• Deep learning (DL) models excel at automatic sleep staging for sleep quality analysis and treatment monitoring.
• Training state-of-the-art DL models requires large-scale, manually labeled sleep datasets, which are difficult and time-consuming to acquire.
• Direct transfer learning from large source datasets to smaller target datasets is hindered by data distribution discrepancies.

Purpose of the Study:

• To develop an unsupervised domain adaptation method for sleep staging networks.
• To reduce data distribution discrepancies between source and target domains.
• To create domain-invariant features for improved sleep staging performance on smaller datasets.

Main Methods:

• Proposed an unsupervised domain adaptation approach using adversarial learning.
• Introduced local discriminators (subject and stage) alongside a classical domain discriminator to preserve intrinsic data structure.
• Implemented novel optimization schemes to enhance the effectiveness of adversarial learning for this specific training scheme.

Main Results:

• The proposed domain adaptation method significantly improved sleep staging performance compared to direct transfer learning.
• The method effectively re-aligned domains into a shared space, producing domain-invariant features.
• Performance improvements were observed across various conditions, demonstrating robustness.

Conclusions:

• Unsupervised domain adaptation is a viable solution to overcome data limitations in DL-based sleep staging.
• The proposed method effectively addresses domain discrepancies without requiring labeled data in the target domain.
• This approach facilitates the development of more generalizable and accurate sleep staging models for diverse populations and datasets.