Joint Classification and Prediction CNN Framework for Automatic Sleep Stage Classification.

Area of Science:

• Computer Science
• Neuroscience
• Biomedical Engineering

Background:

• Accurate sleep staging is crucial for diagnosing and treating sleep disorders.
• Current automatic sleep staging methods face challenges in leveraging temporal dependencies between sleep epochs.

Purpose of the Study:

• To propose a novel joint classification-and-prediction framework for automatic sleep staging using convolutional neural networks (CNNs).
• To introduce an efficient CNN architecture to power this framework, enhancing diagnostic capabilities.

Main Methods:

• A joint classification-and-prediction framework was developed, utilizing CNNs to determine an epoch's label and predict neighboring epochs' labels simultaneously.
• Probabilistic aggregation techniques were employed to leverage multiple decisions generated by the single model.
• Experiments were conducted on the Sleep-EDF Expanded and Montreal Archive of Sleep Studies (MASS) datasets.

Main Results:

• The proposed framework achieved overall classification accuracies of 82.3% on Sleep-EDF and 83.6% on MASS.
• The method demonstrated superiority over baseline classification schemes and existing deep learning approaches.
• The framework effectively leverages dependencies among consecutive sleep epochs.

Conclusions:

• The joint classification-and-prediction framework offers a significant advancement in automatic sleep staging.
• This approach provides a computationally efficient method for achieving high performance, comparable to ensemble methods.
• The study opens new avenues for exploring diverse neural network architectures in automatic sleep staging.