Patient outcome prediction using knowledge graph representation learning

Abstract

The project focuses on using knowledge graphs in a healthcare setting, classifying patient re-admissions. Knowledge graphs are a type of heterogeneous network consisting of entities and relations. Knowledge graph embedding method aims to generate lower-dimensional latent vector representation of these entities and relations while preserving their relational properties. The data consists of patient admission details along with their underlying diagnoses, prescriptions consumed and procedures performed. To exploit the true nature of knowledge graphs, more information to the patient graph is added by combining various biomedical databases to obtain a richer set of relationships. Cleaning patient records and converting the information in more standardized form, as well as gathering information and create a knowledge graph structure in the form of triplets are conducted. The generation of latent vector representations of the entities and relations are done with various embedding methods, where the final phase is to classify patient re-admissions. The methods investigated achieves to represent entities and relations in latent vector form when evaluating the embeddings based on the proposed loss functions. However, the embeddings generated doesn’t supply enough information that can accurately predict the patient readmission status in an extended down-stream fashion. The potential problems could be either of not enough features that explains the variability, not enough rich information regarding the different data sources used, or the effect of class imbalance. A stratified test subset was created from the same excerpt of training data to quantify the results.