Chronic disease prediction using administrative data and graph theory : the case of type 2 diabetes

Clinical diagnosis and regular monitoring of the population at risk of chronic diseases is clinically and financially resource-intensive. Mining administrative data could be an effective alternative way to identify this high-risk cohort. In this research, we apply data mining and network analysis technique on hospital admission and discharge data to understand the disease or comorbidity footprints of chronic patients. Based on this understanding we have developed a chronic disease risk prediction framework. The framework is then tested on Australian healthcare context to predict type 2 diabetes (T2D) risk. The dataset contained approximately 1.4 million admission records from 0.75 million patients. From this, we filtered and sampled the records of 2300 patients having comorbidities including T2D and another 2300 patients having comorbidities other than T2D. Along with demographic and behavioral risk factors for prediction, we propose several graph theory and social network-based measures which indicate the prevalence of comorbidities, transition patterns, and clustering membership. We use an exploratory approach to understand the relative impact of these risk factors and evaluate the prediction performance using three different predictive methods"”regression, parameter optimization, and tree classification. All three prediction methods gave the highest ranking to the graph theory-based 'comorbidity prevalence' and 'transition pattern match' scores showing the effectiveness of the proposed network theory-based measures. Overall, the prediction accuracy between 82% to 87% shows the potential of the framework utilizing administrative data. The proposed framework could be useful for governments and health insurers to identify high-risk chronic disease cohorts. Developing preventive strategies then, over a period of time, can reduce the burden of acute care hospitalization.