Cluster Analysis on Utilization Patterns of Patients with Chronic Diseases Based on Flattened Electronic Medical Records

Designing efficient strategies of reducing health care costs and optimizing facility utilization has long be a pursue of health systems, and stratifying patients into different expenditure/utilization patterns and further profiling these patient groups can provide useful clues for the strategy design. In this study, a data set including patients with chronic diseases (hypertension, diabetes, heart failure, chronic obstructive pulmonary disease (COPD) and/or stroke) was constructed, based on flattening massive electronic medical records and aggregating such records into the patient level. After outlier removal, the total cost and length-of-stay (LOS) of patients were used to cluster the patients into high-cost utilizers and low-cost ones, depending on k-medoids clustering with subsampling validation. To further profile or compare the two patient clusters, median-cost and median-LOS difference between the patients with a specific type or number of disease in the high-cost cluster and those in the low-cost one were investigated. Ultimately, COPD corresponds to the highest cost/LOS difference among the five diseases, implying that it is a potentially important disease that drives high costs/LOS. Moreover, patients with all five diseases correspond to the highest cost/LOS difference compared to those with one to four diseases. Overall, this study can usefully promote the development of healthcare-strategy design and implementation.