Can a Machine Learning Model Help Predict Care Needs Among Older Adults?

The aging population in Japan has presented a significant societal challenge with an increasing number of older adults requiring long-term care. A recent study sought to develop a machine learning model to predict the maximum care needs level that individuals over the age of 75 would require within the next three years.

"Japan is the most aged country in the world, wherein 35.89 million people are aged 65 years and above," write study authors. "They accounted for 28.4% of the total Japanese population in 2019, and this percentage is expected to reach approximately 40% by 2065."

To construct this prediction model, the research team leveraged data from long-term care and health care insurance claims. The study included 47,862 older adults who had not yet received long-term care services in a major Japanese city. The outcome variable for prediction was categorized in alignment with the Japanese long-term care system: class 0 (no care required), class 1 (support levels 1 and 2), class 2 (care levels 1 and 2), and class 3 (care levels 3–5). A total of 516 features were used as explanatory variables, encompassing age, gender, and 514 different diseases classified under ICD-10. The study aimed to construct a prediction model that could provide insights into the factors influencing care-needs levels, thus emphasizing interpretability. For this purpose, the researchers adopted multinomial logistic regression (MLR) with L2 regularization as the machine learning algorithm.

The results of the study revealed that MLR exhibited a weighted average precision of 0.694, recall of 0.505, F-value of 0.567, and a lift score of 1.333. These metrics indicate that the model was reasonably effective in predicting care-needs levels for older adults. When compared to other machine learning algorithms, such as Support Vector Machine (SVM) and Random Forest (RF), MLR demonstrated comparable performance.

Furthermore, the research team conducted a factor analysis based on the magnitude of coefficients in the MLR model. This analysis unveiled the top three features influencing each prediction class. For class 1 (support levels 1 and 2), the key factors were female gender, hypertension, and gonarthrosis. For class 2 (care levels 1 and 2), age, Alzheimer-type dementia, and neuromuscular dysfunction of the bladder were the primary factors. Finally, for class 3 (care levels 3–5), age, Alzheimer-type dementia, and type 2 diabetes mellitus emerged as the most influential features.

In conclusion, this study offers a valuable tool for predicting care-needs levels in the older adult population, particularly those aged over 75, in Japan. The model constructed through MLR with L2 regularization demonstrated commendable predictive performance and interpretability. This model could be practically applied by local governments to identify high-risk areas and allocate resources more effectively. By routinely predicting the care needs of insured persons under public health insurance and long-term care insurance systems, policymakers and health care providers can better prepare for the growing demand for older adult care services in Japan.

"The predictive performances obtained in this study can indicate the limit of a prediction model using features based on claims data, such as disease diagnosis information," authors concluded.

"The care-needs level prediction may require further improvement in terms of predicting individuals, but in practical terms, it 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. Local governments in Japan implement community-based activities for long-term care prevention, including exercise, cognitive training, and nutritional guidance. Identifying high-risk areas would be helpful for local governments in prioritizing the implementation areas of these activities."

Reference: 

Fukunishi H, Kobayashi Y. Care-needs level prediction for elderly long-term care using insurance claims data [published online August 6, 2023]. Inform Med Unlocked. doi: https://doi.org/10.1016/j.imu.2023.101321