Nursing-Care Text Classification and Extraction of Important Terms for Classification Using Transformers

In the medical field, improving the quality of nursing-care by nurses is important for improving the quality of medical services. Therefore, a Web-based Nursing-care Quality Improvement System has been developed by the Japan Institute for Nursing Quality Improvement (JINQI) that evaluates the quality of nursing care in each hospital ward from an independent organization. Currently, we are developing a classification system for nursing-care texts using artificial intelligence technology to support nursing-care experts and enable more hospital wards to use this system. In our former studies, we investigated classification performance by UTH-BERT and visualized the important terms and phrases for classification using attention weights. In this paper, nursing-care texts are classified using models of JmedRoBERTa (a domain-specific model of RoBERTa), rinna-GPT2 (a general-purpose model of the GPT series), and ELECTRA japanese (a general-purpose model of ELECTRA), which are derivative models based on Transformers. We also visualize important terms using SHAP values. From our experimental results, JmedRoBERTa and ELECTRA showed better performance than previous studies. Terms extracted using SHAP values also showed bases close to the judgment bases by nursing-care experts.