Using NLP to extract predicate history from medical device approvals

The FDA approves new moderate-risk medical devices through the Pre-Market Notification (510(k)) process based on their similarity to previously cleared devices known as “predicates”. It is unknown how the features of predicates are associated with the safety of new devices. To address this issue, we employ Natural Language Processing (NLP) techniques to extract the complete list of predicates for each new device from their 510(k) documents and create a predicate database, based on which we assess the association between features of predicates and the likelihood of new devices' recalls. The results help answer questions such as whether new devices with longer predicate history chain are more likely to be recalled and whether new devices with more predicates are more likely to be recalled. Our proposed data-driven approach for analyzing the role of predicates in the 510(k) process helps researchers explore how the process promotes the development of safe medical devices.