Not sure whether this is what you're looking for, but post it anyway!
For finding some hidden relationships between patients, symptoms and diseases, (like if they're from the same earning class, whether or not exercise, or smoke or ...), a study shows that the network features can be more representative than usual patient features. I think the advantage of graphs here is capturing the hidden structure or architecture within the data. Usually doctors go with age, gender and smoking status, but 5 more features added based on graph representation, and 2 of them are in top.
So in this case the hidden (latent) relationships using Graphs are extracting more info than without graph..
Here's the article's title:
A patient network-based machine learning model for disease prediction: The case of type 2 diabetes mellitus
The PDF and the feature importance diagram are attached.
