Automating Safety Surveillance for Software-Based Medical Devices: Insights from FDA MAUDE Data

Digital health interventions (DHIs) use digital technologies to deliver and support healthcare. When registered as medical devices incidents associated with them have to be reported. This study examines errors and incidents involving software-based medical devices as recorded in the U.S. Food and Drug Administration's MAUDE database. Using natural language processing methods, specifically BERTopic clustering and large language model-based classification, we analysed narrative reports to identify event types, contributing factors and temporal patterns. BERTopic effectively revealed thematic trends and topic evolution over time; however, many clusters reflected linguistic conventions in regulatory documentation rather than issues specific to the software. Themes that emerged related to cybersecurity, patient-device communication, and pandemic monitoring suggest that post-market surveillance is expanding to encompass broader human-device and public health contexts. While MAUDE narratives offer valuable insights, their limited contextual detail restricts root-cause analysis. Enhanced reporting structures are needed to support the safer design and oversight of DHIs.