In an electronic medical record, nursing records describe not only the patient’s condition but also a series of nursing practice processes and are considered to reveal the thoughts and actions of nurses. Nursing records are also an important source of information for other healthcare workers as one of the patient’s information;therefore, writing nursing records accurately and in high quality is extremely important, and several hospitals perform regular audit of nursing records. Most nursing records are often written in free text;however, since the announcement of BERT (Bidirectional Encoder Representations from Transformers) in 2018, the performance of natural language processing tasks has improved. In this study, we attempted an automatic audit of nursing records using the BERT model. As a method, using tohoku-BERT that a pre-training model based on Japanese Wikipedia data, and UTH-BERT that a pre-training model based on clinical text, we constructed a classifier with fine tuning on a dataset annotated by an audit nurse to extract self-removal of tubes in nursing records. Concretely, the construction of the classifiers was trained with several changes in the rate of under-sampling. As a result, a high recall rate was shown in all cases when the ratio of positive cases:negative cases was 1:1, 1:5, 1:10, 1:20, 1:50, and 1:100. In particular, when using UTH-BERT with a ratio of 1:100, the accuracy was 0.995, the precision was 0.572, the recall rate was 0.952, and the F1-value was 0.713. Therefore, in the auditing work to extract descriptions of self-removal of tubes in nursing records, the classifier we have constructed may greatly reduce the actual audit target, although there is a small risk of oversight.
View full abstract