Anomaly Detection from Diabetes Similarity Graphs using Community Detection and Bayesian Techniques

Abstract

Diabetes mellitus is a multifactorial chronic disease with many possible contributing factors. Performing anomaly detection on datasets collected from large epidemiological diabetes studies has the potential to unearth previously unknown factors contributing to the pathogenesis of diabetes mellitus. This paper proposes a novel method for detecting anomalous blood glucose trajectories of individuals in a longitudinal diabetes dataset. We formulate the anomaly detection problem as the problem of discovering contextually homogeneous communities in diabetes similarity graphs, from which individuals exhibiting unexpected progression of blood glucose could then be identified. Specifically, we employ community detection and Bayesian techniques to identify communities with the highest degree of anomaly. Our results successfully pointed to individuals with contrasting blood glucose trajectories, even though they demonstrated highly similar social demographics and lifestyle characteristics. Domain expert evaluation supports the efficacy of our proposed method.