Background: Recent advancement in deep learning provides a pivotal opportunity to potentially supplement or supplant the limiting step of manual sleep scoring. New method: In this paper, we characterize the WaveSleepNet (WSN), a deep convolutional neural network (CNN) that uses wavelet transformed images of mouse EEG/EMG signals to autoscore sleep and wake. Results: WSN achieves an epoch by epoch mean accuracy of 0.86 and mean F1 score of 0.82 compared to manual scoring by a human expert. In mice experiencing mechanically induced sleep fragmentation, an overall epoch by epoch mean accuracy of 0.80 is achieved by WSN and classification of non-REM (NREM) sleep is not compromised, but the high level of sleep fragmentation results in WSN having greater difficulty differentiating REM from NREM sleep. We also find that WSN achieves similar levels of accuracy on an independent dataset of externally acquired EEG/EMG recordings with an overall epoch by epoch accuracy of 0.91. We also compared conventional summary sleep metrics in mice sleeping ad libitum. WSN systematically biases sleep fragmentation metrics of bout number and bout length leading to an overestimated degree of sleep fragmentation. Comparison with existing methods: In a cross-validation, WSN has a greater macro and stage-specific accuracy compared to a conventional random forest classifier. Examining the WSN, we find that it automatically learns spectral features consistent with manual scoring criteria that are used to define each class. Conclusion: These results suggest to us that WSN is capable of learning visually agreeable features and may be useful as a supplement to human manual scoring.
|Journal||Journal of Neuroscience Methods|
|State||Published - 1 Aug 2021|
- Computer vision
- Machine learning
- Sleep disruption