Building and Evaluating Prediction Models for Geriatric Syndromes Using Health Insurance Claims

Abstract

In today's Japan, the number of elderly people is increasing, and accordingly, geriatric syndromes are becoming a social problem. The main cause of geriatric syndromes is aging, but they may also be caused by concomitant use of drugs. Thus, identifying the cause of geriatric syndromes is not easy in many cases. In this study, we build machine learning models to predict the onset of geriatric syndromes using the records of health insurance claims (HICs). Specifically, we first extract a couple of datasets in different forms from the HIC records. Then, we build machine learning models using these datasets and evaluate their prediction accuracy. In this process, we adopt XGBoost as a model class and use Optuna for hyperparameter tuning. Finally, we use SHAP to interpret the predictions of the models and verify their validity from a medical viewpoint. The experimental results show that prediction accuracy can be improved when the training dataset include both the number of drugs known to cause geriatric syndromes and the first three digits of the therapeutic category code of drugs. Moreover, from the outputs of SHAP, we identified several therapeutic categories of the drugs that are likely to cause geriatric syndromes.