Automated sleep stage classification of wide-field calcium imaging data via multiplex visibility graphs and deep learning

Authors: Zhang X, Landsness EC, Chen W, Miao H, Tang M, Brier LM, Culver JC, Lee JM, Anastasio MA
University of Illinois at Urbana-Champaign, Urbana, IL - 61801, USA
Washington University School of Medicine
Washington University School of Engineering

Contact: xiaohui8@illinois.edu, landsness@wustl.edu, maa@illinois.edu

Abstract: Wide-field calcium imaging (WFCI) allows for monitoring of cortex-wide neural dynamics in mice. When applied to the study of sleep, WFCI data are manually scored into the sleep states of wake, non-REM (NREM) and REM by use of adjunct EEG and EMG recordings. However, this process is time consuming and often suffers from low inter- and intra-rater reliability and invasiveness. Therefore, an automated sleep state classification method that operates on WFCI data alone is needed. A hybrid, two-step method is proposed. In the first step, spatial-temporal WFCI data is mapped to a multiplex visibility graph (MVG). Subsequently, a two-dimensional convolutional neural network (2D CNN) is employed on the MVGs to be classified as wake, NREM and REM.

System Requirements

• Linux
• MATLAB
• Miniconda >= 4.8.3
• Python 3.7.6.
• Tensorflow 2.2.0.
• NVIDIA driver >= 440.59, CUDA toolkit >= 10.0
• SciPy, NumPy, scikit-image, sklearn, matplotlib.
• Focal loss package.
• Fast natural visibility graph MATLAB package.

The conda environment including all necessaray packages can be created using file environment.yml:

conda env create --file environment.yml

Dataset

The WFCI data in this paper is available on PhysioNet:

Construction of multiplex visibility graph

The directory MVG contains the following sub-directories:

• atlas.mat: variables for defining Paxinos atlas
• define_rois.m: function used for defining which parcels will be used to construct MVG
• parcel2trace.m: function to compute the avarage time series for each parcel
• extrac_MVG.m: the main script to construct MVG

The construction of MVG is based on MATLAB, simply modify the data path highlighted in comment and run extract_MVG.m.

Classifying sleep on MVG with deep learning

The directory network training contains the top level scripts:

• dataloader_MVG.py: Script for loading the MVG representations from WFCI data epochs
• model_cnn2d.py: Script for the compact 2D CNN to classify sleep
• utils.py: script to compute evaluation metrics
• config.txt: example txt. file for defining list of subjects used in training and validation
• train.sh: Scripts for running the network training.
• checkpoints: a pretrained model in our paper is included.

To train the network, make sure to

• Modify the data path highlighted in comment in dataloader_MVG.py.
• Make you own config.txt listing the training/validation mouse name.
• Define parameters in bash script train.sh and run chmod u+x train.sh to make the script executable.
• Simply run ./train.sh to start your training

Citations

1. Landsness, E., & Zhang, X. (2021). Wide-field calcium imaging sleep state database (version 1.0.0). PhysioNet. https://doi.org/10.13026/jzqa-j937.
2. Zhang, X., Landsness, E. C., Chen, W., Miao, H., Tang, M., Brier, L. M., ... & Anastasio, M. A. (2021). Automated sleep state classification of wide-field calcium imaging data via multiplex visibility graphs and deep learning. Journal of Neuroscience Methods, 109421.

References

1. Iacobello, G., Ridolfi, L., & Scarsoglio, S. (2021). Large-to-small scale frequency modulation analysis in wall-bounded turbulence via visibility networks. Journal of Fluid Mechanics, 918.
Chicago	
2. Iacobello, G., Marro, M., Ridolfi, L., Salizzoni, P., & Scarsoglio, S. (2019). Experimental investigation of vertical turbulent transport of a passive scalar in a boundary layer: Statistics and visibility graph analysis. Physical Review Fluids, 4(10), 104501.
3. Paxinos, G., & Franklin, K. B. (2019). Paxinos and Franklin's the mouse brain in stereotaxic coordinates. Academic press.
Chicago