Dual-Teacher Feature Distillation: A Transfer Learning Method for Insomniac PSG Staging.

Area of Science:

• Sleep Medicine
• Artificial Intelligence
• Machine Learning

Background:

• Insomnia is a prevalent sleep disorder with significant long-term health consequences.
• Accurate sleep staging is vital for diagnosing insomnia, but automatic methods are underdeveloped.
• Current transfer learning approaches struggle due to feature distribution differences between healthy and insomniac populations.

Purpose of the Study:

• To propose a novel dual-teacher cross-domain knowledge transfer method for improved automatic sleep staging in insomnia patients.
• To address the limitations of existing transfer learning methods by effectively integrating features from diverse subject groups.
• To enhance the diagnostic accuracy of automatic sleep staging for insomnia disorder.

Main Methods:

• Developed a dual-teacher cross-domain knowledge transfer framework utilizing feature-based knowledge distillation.
• Implemented an insomnia teacher for domain-specific feature learning and a health domain teacher for domain-generic feature learning.
• Employed the Overhaul of Feature Distillation (OFD) method for constructing the health domain teacher.

Main Results:

• The proposed method achieved an average sleep staging accuracy of 80.56% on the CAP-Database.
• Experimental validation using Sleep-EDF (source) and CAP-Database (target) demonstrated superior performance compared to advanced techniques.
• The method showed promising results on a private dataset, indicating generalizability.

Conclusions:

• The dual-teacher cross-domain knowledge transfer method significantly improves automatic sleep staging for insomnia.
• Effective integration of features from healthy and insomniac subjects is crucial for robust sleep staging models.
• This approach offers a promising direction for developing more accurate and reliable tools for insomnia diagnosis.