April 2022 to March 2023

　The preparation of care plans is one of the most burdensome tasks in care management, and it has been pointed out that these plans can vary depending on the care manager preparing them. As a result, there are high expectations that the use of AI can help make improvements in this area. In Phase 1 (FY2017 to FY2019) of the project, we studied ways to analyze data in the form of care plan tables (two) written in text in the development of AI that supports the creation of care plans. In the current research under Phase 2 (FY2020 to FY2023), we reinforced approaches favoring both quantity and quality of data by collaborating with in-home nursing care support providers and NEC, which possesses cutting-edge technologies in white-box AI. We summarized the results of our empirical research.

　We created a multilinear model for visualizing the thought flow of care managers for three diseases; namely, cardiac disease, cerebrovascular disease, and femoral neck fracture. We then trained AI by setting the common items obtained from this model as rules. This enabled classification into new groups based on criteria similar to judgments made by highly specialized care managers. Also, reflecting the deterioration of the mind and body caused by aging as “adjustment variables” made it possible to more accurately reflect the changes in the condition of users in the AI analysis. We were therefore able to achieve a significant transition from a disease-based AI model in Phase 1 to a general-purpose AI model based on eight categorized clusters in Phase 2.

　This fiscal year, particularly focusing on Cluster No. 4 out of the eight clusters, we built a prototype system for a care planning support AI that can be tested by care managers. The system enabled displaying "AI-recommended care" and "warnings of omissions" using appropriate care management methods as options, depending on the user's assessment input information. Although covering only a limited number of subjects, we also verified the prototype system through experiments with care managers. We were able to obtain positive responses indicating the usefulness of the suggestions generated by the system.

　Also, we created a leaflet summarizing our findings to date to widely disseminate the results of our research.