Can Staff Distinguish Falls: Experimental Hypothesis Verification Using Japanese Incident Reports and Natural Language Processing

Falls are generally classified into two groups in clinical settings in Japan: falls from the same level and falls from one level to another. We verified whether clinical staff could distinguish between these two types of falls by comparing 3,078 free-text incident reports about falls using a natural language processing technique and a machine learning technique. Common terms were used in reports for both types of falls, but the similarity score between the two types of reports was low, and the performance of identification based on the classification model constructed by support vector machine and deep learning was low. Although it is possible that adjustment of hyper parameters during construction of the classification model was required, we believe that clinical staff cannot distinguish between the two types of falls and do not record the distinction in incident reports.