WaveSleepNet: An interpretable deep convolutional neural network for the continuous classification of mouse sleep and...

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computational Biology

Background:

• Manual sleep scoring is a time-consuming bottleneck in sleep research.
• Deep learning offers potential solutions for automating sleep scoring.
• Accurate sleep scoring is crucial for understanding sleep disorders and brain function.

Purpose of the Study:

• To introduce and evaluate WaveSleepNet (WSN), a deep convolutional neural network for automated sleep scoring in mice.
• To compare WSN performance against manual scoring and conventional machine learning classifiers.
• To assess WSN's robustness in the presence of sleep fragmentation and on independent datasets.

Main Methods:

• WaveSleepNet (WSN) utilizes wavelet transformed images of mouse electroencephalogram (EEG) and electromyography (EMG) signals.
• WSN employs a deep convolutional neural network (CNN) architecture for sleep stage classification.
• Performance was evaluated using epoch-by-epoch accuracy, F1 scores, and comparison with random forest classifiers.

Main Results:

• WSN achieved a mean accuracy of 0.86 and F1 score of 0.82 compared to human experts.
• On an independent dataset, WSN reached an accuracy of 0.91.
• WSN demonstrated higher accuracy than random forest classifiers and learned visually relevant spectral features.

Conclusions:

• WaveSleepNet (WSN) effectively automates sleep scoring in mice, learning features consistent with manual criteria.
• WSN shows potential as a valuable tool to supplement or potentially replace manual sleep scoring.
• Further research may explore WSN's application in various sleep research contexts and conditions.