A Deep Learning Approach for Predicting FDA’s 510(k) Medical Device Recalls Using Device Citation Relationships

More than 90% of medical devices in the United States are approved through the Food and Drug Administration’s 510(k) pathway, primarily based on demonstrating the equivalence of new devices (known as applicant devices) to previously cleared devices (known as predicate devices). However, safety concerns are raised as applicant devices cleared this way may be more prone to recalls that relate to substantial patient harm and financial strain on the healthcare system. In response, this work introduces a data-driven information technology approach to predict medical device recalls, aiming to alleviate these safety concerns by augmenting human decision making. The approach primarily uses the characteristics of the network formed by predicate device citation relationships (predicate network). It uses deep learning to tackle three design challenges: learning the predicate network structure, capturing the temporal patterns of predicate network characteristics, and accounting for dependencies across the predicate citation history. Based on 45,398 medical devices cleared between 2003 and 2020, the approach substantially improves recall prediction accuracy and timeliness compared with existing state-of-the-art approaches. The improved recall-prediction performance and insights into performance variations across device categories provide opportunities to preemptively react to potential recalls and improve the safety of devices cleared through the 510(k) pathway.