Constructing a Classification Model for Elderly Care Records Using Natural Language Processing

Abstract

In long-term care facilities for the elderly, understanding physical anomalies or daily living conditions is important to improve the quality of care. Internet of Things (IoT) sensors are used to monitor the conditions of users, but it remains difficult to understand the full aspects of users' daily lives. Facilities often maintain care records as text data. Therefore, it is difficult to use sensor data systematically to understand physical anomalies or living conditions. In this study, we constructed a system that uses natural language processing artificial intelligence to analyze nursing care records and automatically determine whether an abnormality exists and the type of abnormality. This study aimed to understand the condition of the users using both care records and IoT sensors. We also studied a system for creating summaries of care records based on a condition judgment system. We collected care records from one nursing home between January 2021 and November 2022; built a natural language processing model that classifies the presence or absence of abnormalities and the types of abnormalities into multiple classes using XLNet; and evaluated the classification performance. The results of normal cases achieved recall of 0.70 and precision of 0.92. In cases of dementia-related symptoms (disquiet), the recall and precision were 0.80 and 0.39, respectively. There were many cases in which records from another data class contaminated the disquiet group. We plan to develop methods to improve judgment performance. In addition, we compared the generated summaries with the actual notes. In some cases, the system was able to generate almost the same contents as the summary records prepared by humans.