Keywords: Graphs, GNNs, Multi-modal
Need: The complexity and depth of information captured in Electronic Health Records (EHR) requires a variety of methods to analyse. This project would look to explore graph representations as such a method to gain insight from EHR and other healthcare data sources, using graph databases and neural network approaches to better explore the data, and the interactions with standards such as Fast Healthcare Interoperability Resources (FHIR). Other modern approaches for encoding sequential information would be good to benchmark against e.g. entity embeddings, etc.
Current Knowledge/Examples & Possible Techniques/Approaches: Recent work using graph models to build simpler knowledge discovery systems e.g. to model electronic health records for improvements in diagnostic prediction, or to understand negative drug interactions - this is achieved by storing the information in a format that is closer to reality. Utilising Graph Data Structures, Embeddings, Graph Neural Networks as discussed in Scalable and accurate deep learning with electronic health records and Variationally Regularized Graph-based Representation Learning for Electronic Health Records, etc.
Related Previous Internship Projects: n/a as first year of the scheme.
Enables Future Work: These both constitute important steps towards better features for ML/AI solutions such as the recent interest in Graph Neural Networks. Supports work looking at interoperability.
Outcome/Learning Objectives: Demonstration of transforming EHR data to Graph based and the value this has for example pieces of analysis.
Datasets: Open healthcare datasets with relevant structures with a view to extend to other datasets if successful.
Desired skill set: When applying please highlight any experience around informatics, graph structures, deep learning, coding experience (including any coding in the open), and any other data science experience you feel relevant.
Return to list of all available projects.