自然言語処理を用いた医療機器不具合報告からの回収推定

抄録

It has been reported that about 400 kinds of medical devices are recalled every year in Japan. Additionally, more than 10,000 kinds of defect reports are raised in each year, of which 100 to 200 kinds of devices result in recalls. In this study, we analyzed medical device defect reports in Ministry of Health, Labour and Welfare between 2008 and 2022 using text mining. We targeted total 4,278 case of generator of cardiac implantable electronic devices, of which 342 cases were recalled. We analyzed contents of problem status and health damage reports using text mining. Moreover, we attempted to estimate cases that result in recalls using Bidirectional Encoder Representations from Transformers (BERT). As a method, using tohoku-BERT that a pre-training model based on Japanese Wikipedia data, and JMedRoBERTa that a pre-training model based on medical research papers. We conducted a classifier with fine tuning on a dataset annotated by medical device defect reports. As a result, the recall rate was 0.918 and the F2-score was 0.659 using JMedRoBERTa with over sampling data.