Representation Learning for Medical Data

Karol Antczak Military University of Technology in WarsawInstitute of Computer and Information SystemsABSTRACT We propose a representation learning frameworkfor medical diagnosis domain. It is based on hetero - geneous network-based model of diagnostic data aswell as modified metapath2vec algorithm forlearning latent node representation. We comparethe proposed algorithm with other representationlearning methods in two practical case studies:symptom/disease classification and disease predic - tion. We observe a significant performance boost inthese task resulting from learning representationsof domain data in a form of heterogeneous network. INTRODUCTIONRepresentation learning is a group of machinelearning methods that aims to find useful represen - tations of the data. The "usefulness" is typicallyunderstood in terms of extraction of features thatare meaningful from the point of view of targetobjective.