REST-a deep learning tool for automated mouse sleep stage classification.

Area of Science:

• Neuroscience
• Computational Biology
• Sleep Medicine

Background:

• Manual sleep scoring in rodents is time-consuming and prone to inconsistencies.
• Large-scale electroencephalography (EEG) and electromyography (EMG) recordings are crucial for rodent sleep research.
• Automated methods are needed to improve efficiency and reliability in sleep stage classification.

Purpose of the Study:

• To develop and validate a neural network model for automated sleep stage classification in rodents.
• To assess the accuracy and generalizability of the developed model across different mouse strains.
• To provide a fast, accurate, and accessible tool for analyzing rodent sleep.

Main Methods:

• A Transformer-based neural network, REST-Rodent EEG Sleep Transformer (REST), was developed for sleep stage classification.
• EEG and EMG signals were processed using short-time Fourier transform and analyzed in 4-second epochs.
• The model was trained on 116 days of pre-labeled recordings from Fmr1 knockout mice and tested on multiple mouse strains.

Main Results:

• REST achieved high accuracy with a Cohen's kappa of 0.873 and F1-scores above 0.91 for all sleep stages (Wake, NREM, REM).
• The model demonstrated robust cross-strain compatibility and generalized effectively to external datasets.
• REST outperformed a convolutional neural network (CNN)-based model (Accusleep) in performance.

Conclusions:

• REST is a fast, accurate, and accessible tool for automated sleep stage classification in mice.
• The Transformer architecture enables long-range temporal awareness for consistent rodent sleep analysis.
• REST facilitates reliable analysis of large-scale rodent sleep datasets, overcoming limitations of manual scoring.