A Novel Disease Prediction Method Based on Inductive Representation Learning

Using rich imaging and non-imaging information to complete the disease prediction task, the model is required to simultaneously represent individual characteristics and data associations between subjects from a potentially large population. The graph provides a potential framework for such tasks. In this study, we extend GCN to inductive representation learning, combine imaging and non-imaging data for population brain image analysis, and propose an inductive representation learning method for disease prediction based on population graph. Specifically, the structure represents the population as a sparse graph, where the vertices are associated with image-based eigenvectors and the edges represent the population phenotypic information. Then the inductive representation learning model was trained on a partial labeled graph to infer the category of unlabeled nodes based on node characteristics and pairwise associations between subjects. The results demonstrate the improvement of our proposed approach on the ABIDE dataset and emphasize the importance of obtaining more implicit information from the diagram structure. This has an obvious impact on the performance of the prediction model, and makes the prediction accuracy of 77.2%, which is significantly better than the standard linear classifier which only considers a single feature and the previous graph representation algorithm.