Electronic health records (EHR) are a connected data structure that can be modelled as a graphical structure. Research has shown that using graphical EHR is superior on predictive tasks than simply assuming no data connectivity. However, EHR data doesn’t always contain structural information making it difficult to actually create graphical EHR. The authors propose the Graph Convolutional Transformer (GCT), a novel approach to jointly learn the hidden structure while performing various prediction tasks when the structure information is unavailable. The proposed model consistently outperformed previous approaches empirically, on both synthetic data and publicly available EHR data, for various prediction tasks such as graph reconstruction and readmission prediction, indicating that it can serve as an effective general-purpose representation learning algorithm for EHR data.

• the importance of structure in eHR data
• self attention via transformers
• the use of KL divergence as a regularizer
• training issues with regards to sparsity

• Structure in EHR data is important can be modelled as a graph data structure for any downstream tasks
• Structural information is not always provided but can be learned jointly with the prediction tasks
• Transformers can be used to learn the adjacency matrix which are interpreted as attention coefficients
• Instead of starting from a dense adjacency matrix, domain knowledge can be used to incorporate priors into the adjacency matrix
• To ensure stability in the learning coefficients, the KL divergence can be used to prevent the attention coefficients from jumping sporatically inbetween GCT layers