Validation of Deep Learning-based Sleep State Classification.

Area of Science:

• Neuroscience
• Computational Biology

Background:

• Deep learning models achieve high accuracy in classifying mouse sleep states from electroencephalogram (EEG) and electromyogram (EMG) data.
• Classification accuracy often decreases on independent datasets due to variations in experimental and recording conditions (distributional shift).
• Mixture z-scoring is a standardization technique proposed to mitigate these variations in biosignal processing.

Purpose of the Study:

• To validate the effectiveness of mixture z-scoring combined with deep learning for sleep state classification.
• To assess the performance of the open-source software Accusleep on an independent dataset.

Main Methods:

• Utilized the Accusleep software, which integrates mixture z-scoring with a convolutional neural network (CNN) deep learning model.
• Applied the method to 12 three-hour electroencephalogram (EEG) and electromyogram (EMG) recordings from mice in a head-fixed position.
• Classified sleep states (e.g., wake, NREM, REM) using the processed signals.

Main Results:

• Achieved sleep state classification accuracy ranging from 85% to 92% on two independent recordings.
• Obtained Cohen's kappa (κ) values between 0.66 and 0.71, indicating substantial agreement.
• Demonstrated the robustness of the mixture z-scoring and deep learning approach across different conditions.

Conclusions:

• Mixture z-scoring effectively enhances the performance of deep learning models for mouse sleep state classification.
• The Accusleep software provides a validated tool for robust sleep analysis.
• This approach holds potential for widespread adoption in sleep research due to its improved accuracy and generalizability.