Graph-Based Modeling of Diabetic Patient Data for Readmission Risk and Care Pattern Analysis

Abstract

Electronic health records (EHRs) contain rich relational data such as individual patient data, encounters, diagnoses, medications, etc. Healthcare systems often store EHR data in tabular form. However, traditional flat representations (“bag of features”) can lose critical context. For example, treating a patient encounter as an unordered set of codes obscures the fact that a specific combination of drugs might have caused an adverse outcome. Knowledge graphs offer a robust alternative by organizing medical data into interconnected entities and relationships, capturing complex associations (e.g. between symptoms, treatments, diagnoses) for a more holistic understanding of patient history. In this work, we transform the Diabetes 130-US Hospitals dataset (a collection of ~100,000 inpatient encounters from 130 hospitals over 10 years) into a labeled property graph (LPG), and demonstrate the advantages both conceptual and quantitative of graph-based analysis, in a medical informatics context. Each encounter in this dataset includes patient demographics, diagnoses (ICD-9 codes), lab results (e.g. HbA1c), and 24 diabetes-related medications with change indicators (“up”, “down”, “steady” or “no change”) among other features. Notably, the original study focused on 30-day readmissions, highlighting that poor glycemic control and suboptimal inpatient diabetes management lead to higher readmission rates and complications. Our graph model makes these clinical relationships explicit, enabling multi-hop reasoning (e.g. linking a patient’s lab result to medication changes and subsequent readmission outcome) that is cumbersome with relational tables. We show that converting such EHR data into a graph can improve predictive modeling of readmissions and uncover insightful patterns of comorbidities and care processes that would be difficult to extract using SQL alone, aligning with recent trends in biomedical informatics to leverage networks for clinical data analysis.