Application of Machine Learning to Sleep Stage Classification.

Area of Science:

• Neuroscience
• Computational Biology
• Sleep Medicine

Background:

• Manual classification of polysomnography for sleep studies is labor-intensive, requires expert training, and suffers from inter-scorer variability.
• Existing automated methods often rely on multiple electroencephalogram (EEG) channels, posing challenges for small animal research.
• Accurate sleep stage classification is crucial for understanding sleep loss and its contribution to psychopathology.

Purpose of the Study:

• To develop an automated, open-access classifier for predicting rodent vigilance states using a single cortical EEG and EMG signal.
• To overcome the limitations of manual scoring and multi-channel EEG requirements in sleep studies.
• To evaluate the performance of various machine learning algorithms for this classification task.

Main Methods:

• Utilized approximately 427 hours of expert-labeled EEG, EMG, and activity data from rodents.
• Evaluated machine learning techniques including Decision Trees, Random Forests, Naive Bayes, Logistic Regression, and Artificial Neural Networks.
• Classified 10-second epochs into three vigilance states: paradoxical sleep, slow-wave sleep, and wake.

Main Results:

• Machine learning classifiers achieved accuracies ranging from 74% to 96%.
• Random Forest and Artificial Neural Network (ANN) models demonstrated high performance, with accuracies of 95.78% and 93.31%, respectively.
• The study successfully validated the potential of single-channel EEG and EMG for automated sleep stage classification.

Conclusions:

• Automated vigilance state classification using machine learning on single EEG and EMG signals is feasible and accurate.
• This approach offers a more efficient, reliable, and less invasive alternative to manual scoring and multi-channel systems.
• The developed classifier holds promise for advancing research into sleep disorders and associated psychopathology.