Mapping the genealogy of medical device predicates in the United States

In the United States, medical devices are regulated and subject to review by the Food and Drug Administration (FDA) before they can be marketed. Low-to-medium risk novel medical devices can be reviewed under the De Novo umbrella before they can proceed to market, and this process can be fairly cumbersome, expensive, and time-consuming. An alternate faster and less-expensive pathway to going to market is the 510(k) pathway. In this approach, if the device can be shown to be substantially equivalent in safety and effectiveness to a pre-existing FDA-approved marketed device (or "predicates"), it can be cleared to market. Due to the possibility of daisy-chaining predicate devices, it can very quickly be difficult to unravel the logic and justification of how a particular medical device's equivalence was established. From patients' perspective, this minimizes transparency in the process. From a vendor perspective, it can be difficult to determine the right predicate that applies to their device. We map the connectivity of various predicates in the medical device field by applying text mining and natural language processing (NLP) techniques on data publicly made available by the FDA 78000 device summaries were scraped from the US FDA 510(k) database, and a total of 2,721 devices cleared by the 510(k) regulatory pathway in 2020 were used as a specific case study to map the genealogy of medical devices cleared by the FDA. Cosine similarity was used to gauge the degree of substantial equivalence between two medical devices by evaluating their device descriptions and indications for use. Recalls and complaints for predicate devices were extracted from the FDA's Total Product Life Cycle database using html scraping and web page optical character recognition to determine the similarity between class 1 recalled devices (the most severe form of device recall) and other substantially equivalent devices. A specific product code was used to illustrate the mapping of the genealogy from a De Novo device. The ancestral tree for the medical devices cleared in 2020 is vast and sparse, with a large number of devices having only 1-2 predicates. Evaluation of substantial equivalence data from 2003-2020 shows that the standard for substantial equivalence has not changed significantly. Studying the recalls and complaints, shows that the insulin infusion pump had the highest number of complaints, yet none of the recalled devices bore significant degree of text similarity to currently