Towards precision sleep medicine: Self-attention GAN as an innovative data augmentation technique for developing...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• High-quality sleep is crucial for cognitive functions, emotional regulation, and overall health.
• Accurate and rapid diagnosis of sleep disorders is essential for well-being.
• Existing automatic sleep scoring methods often lack personalization, leading to reduced accuracy due to individual differences.

Purpose of the Study:

• To develop an improved personalized automatic sleep scoring classification method.
• To address the challenge of individual variability in sleep data.
• To enhance the accuracy and efficiency of sleep quality assessment.

Main Methods:

• Utilized electroencephalography (EEG) data, converting it into spectrograms.
• Applied self-attention generative adversarial network (SAGAN) for advanced data augmentation.
• Generated subject-specific synthetic spectrograms using SAGAN.
• Trained a personalized classifier using both real and synthesized spectrograms.

Main Results:

• Achieved an average accuracy of 95.74% with a standard deviation of 3.78% for personalized sleep scoring.
• Demonstrated an 8.08% increase in average accuracy compared to a classifier trained on general data.
• Validated the effectiveness of SAGAN-generated spectrograms in improving classification performance.

Conclusions:

• The proposed personalized approach significantly enhances automatic sleep scoring accuracy.
• SAGAN-based data augmentation effectively overcomes individual differences in sleep patterns.
• This method offers substantial savings in medical resources and time for healthcare professionals and patients.