Pupil Dynamics-derived Sleep Stage Classification of a Head-fixed Mouse Using a Recurrent Neural Network.

Area of Science:

• Neuroscience
• Computational Biology
• Sleep Science

Background:

• Traditional sleep state classification relies on electroencephalography (EEG) and electromyography (EMG) with manual expert correction.
• Current methods face limitations including time consumption, interference with imaging, surgical risks, and incompatibility with certain experimental setups.

Purpose of the Study:

• To develop a non-invasive pupil dynamics-based method for classifying vigilance states (NREM, REM, WAKE) in head-fixed mice.
• To overcome the limitations of conventional EEG/EMG-based sleep scoring.

Main Methods:

• Utilized a long short-term memory (LSTM) model, a type of recurrent neural network, for multi-class sleep state labeling.
• Integrated EEG and EMG recordings with left eye pupillometry tracked by DeepLabCut, a markerless tracking toolbox.
• Employed pupil diameter, location, velocity, and eyelid opening as features for the LSTM model at a 10 Hz sampling rate.

Main Results:

• The LSTM model achieved higher classification performance (macro F1 score: 0.77, accuracy: 86%) compared to a feed-forward neural network.
• Pupil dynamics revealed potential subdivisions within established EEG/EMG-defined vigilance states.
• Demonstrated the feasibility of pupil dynamics for accurate sleep stage scoring.

Conclusions:

• Pupil dynamics offer a viable, non-invasive alternative for sleep stage scoring in head-fixed mice.
• This approach enhances efficiency and reduces risks associated with traditional methods.
• The findings suggest pupil dynamics could refine our understanding of vigilance states.