Towards More Accurate Automatic Sleep Staging via Deep Transfer Learning.

Area of Science:

• Sleep Medicine
• Artificial Intelligence
• Machine Learning

Background:

• Automatic sleep staging is crucial but challenging for small sleep study cohorts due to data variability and inefficiency.
• Existing methods struggle to generalize when data is limited, hindering accurate sleep analysis.

Purpose of the Study:

• To present a deep transfer learning approach to enhance automatic sleep staging models for small cohorts.
• To address data-variability and data-inefficiency issues in sleep studies with limited data.

Main Methods:

• Developed a generic end-to-end deep learning framework for sequence-to-sequence sleep staging.
• Utilized two networks for transfer learning, pretraining on a large dataset (MASS) and finetuning on smaller target datasets (Sleep-EDF, Surrey-cEEGrid).
• Evaluated the approach on target domains with varying degrees of data mismatch to the source domain.

Main Results:

• Demonstrated significant performance improvements in automatic sleep staging on the target domains.
• The deep transfer learning approach effectively transferred knowledge from a large database to small cohorts.
• Achieved enhanced accuracy despite data variability and limited sample sizes.

Conclusions:

• The proposed deep transfer learning approach is effective in overcoming data-variability and data-inefficiency in automatic sleep staging.
• This method enables the development of higher-quality sleep staging models even with small datasets.
• Facilitates improved sleep analysis in research settings with limited data availability.