Predicting high cost patients with type 2 diabetes mellitus using hospital databases in a multi-ethnic Asian population

Abstract

Data analytics is playing an important role in health care because of the potential actionable insights that can be derived from individual-level medical records in the electronic health records (EHRs). This paper explores the utilization of EHR data for predictive analytics at an academic health system in Singapore to facilitate patient stratification for intensive case management among individuals with type 2 diabetes mellitus (T2DM). Though a multidisciplinary team approach, we developed a risk score for high health care utilizers with EHRs. A backward stepwise variable selection model building approach was performed to develop a risk score using the multiple logistic regression model and Akaike Information Criterion where the variables from 2010 was used to predict the top 10% health care spenders in 2011. The list of predictors in the risk score included sociodemographic, biochemistry, comorbidity and healthcare utilization variables. The Area under the Curve (AUC) of the risk score was 0.708, which was higher than having total cost in 2010 as the only predictor (AUC = 0.658). The absence of biochemistry measurements could either be a proxy of no regular follow-up for managing T2DM condition if their regular measurements were part of the clinical practice for T2DM, or be a proxy of a favorable perception of patient's medical condition otherwise. A close collaboration across multiple disciplines is important to ensure a holistic interpretation of a risk score.