Long-Term Care Risk Prediction by Claims Data Analysis

In an aging society such as Japan and other OECD countries, the medical and non-medical costs of long-term care are increasing every year, and it is becoming more difficult for local governments to bear the costs. It is necessary to identify high-risk persons before they need long-term care. We built a prediction model for seven scenarios regarding the need for long-term care using medical claims, pharmacy claims, diagnosis procedure combination (DPC) payment system data, health examination results, and long-term care claims from a local government database. Based on the number of actual data, our proposed long-term care risk prediction is targeted at the elderly over 75 years old. Because there are many variables in the data, especially in medical claims, we used the heterogeneous mixture learning (HML) model because it can automatically optimize all combinations of explanatory variables. The explanatory variables we used are age, sex, 533 diagnosis codes as listed in ICD-10, 108 prescription drug groups under the therapeutic category of drugs in Japan, and 28 special health examination items. The results showed that the area under the curves (AUCs) for all scenarios was above 0.7. Since the recalls of HML were larger than that of other machine learning models, more high-risk persons can be identified by our model.