Complex-valued unsupervised convolutional neural networks for sleep stage classification.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Signal Processing

Background:

• Deep learning for sleep stage classification typically requires extensive labeled data.
• Labeling sleep data is subjective, time-consuming, and challenging with wearable devices.
• Unsupervised learning is crucial for overcoming data limitations in sleep analysis.

Purpose of the Study:

• To propose a novel unsupervised method for automatic sleep stage classification.
• To introduce complex-valued unsupervised convolutional neural networks (CUCNN) for sleep analysis.
• To address the need for automated feature extraction in sleep studies.

Main Methods:

• CUCNN utilizes complex-valued inputs, outputs, and weights with a greedy layer-wise training strategy.
• The method involves a phase encoder to convert real-valued inputs to complex numbers.
• Unsupervised training employs complex-valued K-means for learning convolutional filters.

Main Results:

• CUCNN achieved 87% total accuracy and a 0.8 kappa coefficient on the UCD dataset.
• CUCNN outperformed unsupervised convolutional neural networks (UCNN) by over 12% in total accuracy on UCD and MIT-BIH datasets.
• CUCNN demonstrated significantly faster convergence compared to UCNN.

Conclusions:

• The proposed CUCNN method is fully automated and performs unsupervised feature learning.
• Unsupervised training and automatic feature extraction are feasible for sleep data.
• This approach is highly valuable for developing home sleep monitoring systems.