2020 Volume 40 Issue 2 Pages 73-82
[Background] Pain management and the sharing of information on pain between medical staff and patients, are significant elements in the medical domain. However, information on patients’ pain is generally included as free text in clinical notes, making it difficult to extract the required information from clinical notes. [Purpose] We aimed to extract pain information from clinical notes and to conduct a factuality analysis. [Method] Clinical notes, generated in July 2016, were obtained from the University of Tokyo Hospital. Sentences were denoted with spaces, indents and commas in the notes ; 20,000 sentences were randomly selected for analysis, and medical experts annotated pain-related information in the data. [Result] We obtained an average F1 of score of 56.4 with BERT, which was pre-trained using clinical notes. [Discussion] The experimental result indicates that BERT could improve accuracy in extracting pain information from clinical notes.