Care-needs level prediction for elderly long-term care using insurance claims data

Owing to an aging population, the increase in the number of elderly people certified as requiring long-term care has become a critical social issue in Japan. This study aimed to construct a machine learning model predicting the maximum care-needs level required for long-term care within the next three years for persons aged over 75 years. The prediction model was constructed using features extracted from long-term care and healthcare insurance claims data. The study subjects were a total of 47,862 elderly individuals who had not received long-term care services in a large city in Japan. The prediction classes for outcome variable were categorized according to the criteria of the Japanese long-term care system: class 0 (no required), class 1 (support levels 1 and 2), class 2 (care levels 1 and 2), and class 3 (care levels 3–5). As explanatory variables, a total of 516 features were used, including age, sex, and 514 diseases classified under ICD-10. In this study, we focused on constructing a prediction model with the interpretability and adopted multinomial logistic regression (MLR) with L2 regularization as a machine learning algorithm. MLR allowed us to identify the characteristics influencing each prediction class of care-needs levels. In terms of overall predictive performance, MLR achieved weighted average precision, recall, F-value, and lift scores of 0.694, 0.505, 0.567, and 1.333, respectively. Compared to other machine learning algorithms, MLR demonstrated comparable performance to Support Vector Machine (SVM) and Random Forest (RF). From the factor analysis based on the magnitudes of coefficients of the MLR model, the top three features influencing each prediction class were as follows: Class1: female sex, hypertension, and gonarthrosis; Class 2: age, Alzheimer-type dementia, and neuromuscular dysfunction of the bladder; Class 3: age, Alzheimer-type dementia, and type 2 diabetes mellitus. In practical terms, the care-needs level prediction can be applied by local governments to identify high-risk areas by comprehensively and routinely predicting insured persons under public health insurance and long-term care insurance systems.