Are Nonimage Data Really Necessary for Disease Prediction With Graph Convolutional Networks?

Graph convolutional networks (GCNs) have been widely applied to automatic disease diagnosis based on neuroimaging data, and nonimage data are used to determine local connections in the GCN mode. However, previous studies reveal the GCN model may perform even worse than the linear model if nonimage data are inappropriate or unavailable. Considering that manually identifying disease-related nonimage data among numerous alternatives is time-consuming and nonimage data are usually not available in medical data sets for privacy reasons, this shortage limits the application of the GCN model. Besides, whether nonimage data are really necessary is also worth discussing, since much literature have revealed that neuroimaging data can well characterize nonimage data. To overcome the limitation, we apply a nonlocal operation (a special form of attention mechanism) to the GCN model (nonlocal GCN), which automatically determines local connection based on image data and no more relies on the nonimage data. The experiments on two public data sets show that the proposed model can perform better or achieve almost the same performance without using any nonimage data. The results on simulation data sets reveal that our model can perform better in feature-driven data sets if image data contain nonimage data, in which the nonimage data are not necessary.