MedGCN: Medication recommendation and lab test imputation via graph convolutional networks

•We innovatively incorporate the complex associations between multiple medical entities into a graph called MedGraph, and develop a machine learning framework MedGCN to learn the representations for entities in MedGraph for medication recommendation and lab test imputation. The framework can also be generalized to other medical tasks.•MedGCN extends the general GCN model to heterogeneous graphs and missing feature values in medical settings.•We introduce cross regularization, an effective regularization strategy to reduce overfitting for the training of MedGCN.•MedGCN is an intrinsic inductive model that could use the learned model to efficiently generate representations for new data.