# GraphSleepNet(2020)
>[Source](https://www.ijcai.org/proceedings/2020/0184.pdf)

- ### Abstract
- The model proposed GraphSleepNet learns the intrinsic connection between different EEG channels represented by an adjacency matric and passes it to Spatial Temporal Graph convolution network(ST-GCN).
- Each channel is corresponds to a node and the connection between the channels as edges.
- A limitation that other methods have is the input must be grid data and as brain regions are not in Euclidean space graph structures would be better.

- ### Model
- 
- 

- **Adaptive Sleep Graph Learning**
- 
- It dynamically learns the hraph structure instead of having a fixed graph.
- **Spatial-Temporal Graph COnvolution**
- Spatial-temporal graph convolution is a combination of spatial graph convolution and temporal standard convolution, which is used to extract both spatial and temporal features.
- Spacial convolution is based on spectral graph theory to extract spatial features, which uses Chebyshev extension to reduce complexity.
- Temporal concolutions capture sleep transition rules and uses CNNs.
- **Spatial-Temporal Attention**
- In the spatial dimension, different regions have different effects which dynamically changes during sleep.
- In the temporal dimension, there are correlations between different sleep stages.
- These have attention mechanisms applied to extract valuable features.

- ### Evaluation
- The MASS Dataset is used and is preprocessed with different bandpass ffilters for each channel.
- A 31 fold cross validation is used to evaluate the performance of GraphSleepNet.
- It has accuracy of 88.9 and N1 has F1 score of 60.3.