Abstract
Background: This research aims to develop an AI-powered care management support system that predicts declines in activities of daily living (ADL) for older people requiring long-term care in Japan, addressing the current lack of information sharing among multiple home care service providers and potentially improving care coordination and patient outcomes.
Approach: Our research team is developing a website that collectively manages care information of older people requiring long-term care, predicts declines in ADL based on artificial intelligence (AI) analysis, and automatically sends reports to service providers and care managers when necessary. To verify the feasibility of this care management support system, our research team collected information on 50 older people requiring care over six months and made predictions using AI-based machine learning. The research team collaborated with a care management specialist and a website development expert while actively involving elderly individuals and their families in the design process.
Ten experienced home care managers each randomly selected five users and entered data about those users once a month for six consecutive months, ensuring real-world applicability. A Google Form was distributed for data entry, allowing the input of 12 monitoring items and the Barthel Index score. A machine learning environment was built in Python to predict Barthel Index scores from the collected data.
For the learning process, 85% of the collected data was used for training, while 15% was reserved for validation and evaluation. Since the 12 learning factors are categorical variables, one-hot encoding was applied. A support vector machine (SVM) served as the learning model, with grid search employed to optimize hyperparameters. Additionally, recursive feature elimination (RFE) extracted 10 key factors for prediction, weighting these factors in the learning process to identify essential items for predicting the Barthel Index score.
Results: By searching for hyperparameters and extracting factors using RFE, it was possible to predict the Barthel Index score with a learning accuracy of 41%. From the 10 factors extracted by RFE, it was shown that matters such as worsening of dementia symptoms, activity level, bowel movements, falls, personal complaints, fluid intake, and food intake are closely related to predicting the Barthel Index score.
Implications: Developing a website that uses AI to predict declines in ADL and sends reports to care providers addresses the lack of information sharing among multiple service providers in Japan's long-term care system. This solution could enhance the coordination and delivery of care and be adapted to other countries facing similar challenges. As improving prediction accuracy is related to the amount of training data, future studies will require an increase in the number of subjects and continued data collection.