Free-text MAUDE narratives provide a source-robust representation layer for biomaterial-device surveillance

Implantable biomaterial devices require effective post-market surveillance because clinically important failure patterns often emerge only after widespread use. However, surveillance workflows often rely on structured coded summaries that compress heterogeneous adverse-event narratives into coarse categories. This study compares coded and free-text narrative representations across 1,500 FDA MAUDE reports from three biomaterial device classes (coronary stents, bone cement, and surgical mesh) to test whether narratives preserve a more source-robust surveillance representation. Under manufacturer-held-out evaluation, narrative TF-IDF features outperformed structured code-only features (macro F1 0.925 versus 0.827), while delexicalized narratives retained strong grouped performance after masking device-class, manufacturer, brand, and legal-template tokens (F1 0.897). Narrative topics resolved reported events into procedural, anatomical, host-response, and reporting-context patterns, and an interpretable classifier recovered code-derived complication phenotypes from narrative text alone (mean F1 0.902, AUC 0.967). These findings support free-text adverse-event narratives as a complementary representation layer for post-market device surveillance, while remaining bounded by passive adverse-event reporting limitations and requiring validation across additional years, device classes, and independently adjudicated outcomes. Author SummaryWhen an implanted medical device fails inside a patient, the event is reported to the FDAs MAUDE database. Each report includes both a standardized code and a written narrative describing what happened. We asked whether these two representations carry the same information. Using 1,500 reports covering coronary stents, bone cement, and surgical mesh, we found that coded fields lose much of the clinical detail present in narratives. Importantly, narrative-based classifiers remained accurate even when tested on reports from manufacturers not seen during training, while code-based classifiers dropped substantially. This matters because real-world surveillance must generalize across different reporting sources. We also found that narrative text can recover clinically meaningful complication patterns that are defined by codes, and that most reports never name the specific biomaterial involved. These findings suggest that narrative text deserves a more central role in post-market device monitoring, complementing the coded fields that current surveillance pipelines rely on.