2022 Volume 42 Issue 2 Pages 47-59
In the field of medical information, the amount of text information generated is increasing day by day, and the realization of automatic processing by computers is an urgent issue. However, medical texts, such as medical records generated in hospitals and medical diaries posted by patients on social networking services, are difficult to handle through natural language processing because of the use of special styles and expressions.
The Kanagawa Cancer Center keeps text data records of the cancer telephone counseling service that was previously provided. However, these text data are unsuitable for natural language processing because they are described using special expressions, like other medical texts. Therefore, in this study, we first conducted a qualitative analysis, and then used text mining to extract and visualize the main complaint of the consultant.
First, based on the results of the qualitative analysis, the chief complaints “worry,” “anxiety,” “request,” “distrust and mistrust,” and “dissatisfaction” that appeared cross-sectionally were selected as the targets of text mining. Next, the object keywords of these chief complaints were extracted using extended Backus-Naur form and visualized using a graph line drawing tool.
The extraction results were subjected to a performance evaluation using the F-measure. As a result, the F-measure for “worry,” “request,” “distrust and mistrust,” and “dissatisfaction” all exceeded 0.7, with the F-measure of “request” reaching approximately 0.8. In addition, this method was also confirmed to improve performance in comparative experiments with general text mining methods. On the contrary, the F-measure of “anxiety” was approximately 0.62 because the object keywords were frequently ambiguous. Dealing with ambiguity is a work for the future.
These results demonstrate the effectiveness of our method, and the findings of this study may be useful for processing other medical texts.
