A systematic review of networks for prognostic prediction of health outcomes and diagnostic prediction of health conditions within Electronic Health Records.

Background And Objective: Using graph theory, Electronic Health Records (EHRs) can be represented graphically to exploit the relational dependencies of the multiple information formats to improve Machine Learning (ML) prediction models. In this systematic qualitative review, we explore the question: How are graphs used on EHRs, to predict diagnosis and health outcomes?

Methodology: The search strategy identified studies that used patient-level graph representations of EHRs to utilise ML to predict health outcomes and diagnoses. We conducted our search on MEDLINE, Web of Science and Scopus.

Results: 832 studies were identified by the search strategy, of which 27 studies were selected for data extraction. Following data extraction, 18 studies used ML with patient-level graph-based representations of EHRs to predict health outcomes and diagnoses. Models ranged from traditional ML to neural network-based models. MIMIC-III was the most used dataset (n = 6, where n is the number of occurrences), followed by National Health Insurance Research Database (NHIRD) (n = 4) and eICU Collaborative Research Database (eICU) (n = 4). The most predicted health outcomes were mortality (n = 9; 21%), hospital readmission (n = 9; 21%), and treatment success (n = 4; 9%). Model performances ranged across outcomes, mortality prediction (Area Under the Receiver Operating Characteristic (AUROC): 72.1 - 91.6; Area Under Precision-Recall Curve (AUPRC): 34.8 - 81.3) and readmission prediction (AUROC: 63.7 - 85.8; AUPRC 39.86 - 84.7). Only one paper had a low Risk of Bias (RoB) that applied to our research question (4%).

Conclusion: Graph-based representations using EHRs, for individual health outcomes and diagnoses requires further research before we can see the results applied clinically. The use of graph representations appears to improve EHR representation and predictive performance compared to baseline ML methods in multiple fields of medicine.