Applying Natural Language Processing for Postmarket Surveillance: AI-driven Analysis of Adverse Events with the Woven EndoBridge (WEB) Device
Ibraheem Alkhawaldeh1, Hamza Alsalhi2, Yousef Hawas3, yasmeen Jamal Alabdallat2, Mostafa Hossam El Din Moawad4, Ibrahim Serag5, Mohamed Abouzid6
1Mutah university - jordan hospital, 2Faculty of Medicine, Hashemite University, Zarqa, Jordan, 3Faculty of Medicine, Tanta University, Gharbeya, Egypt, 46.Alexandria Main University Hospital, Alexandria, Egypt Faculty of Medicine, Suez Canal University, Ismailia, Egypt, 5Mansoura University, Mansoura, Egypt, 6Department of Physical Pharmacy and Pharmacokinetics, Faculty of Pharmacy, Poznan University of Medical Sciences, Rokietnicka 3 St., 60-806 Poznan, Poland
Objective:

This study evaluated real-world adverse events of Woven EndoBridge (WEB) devices using the FDA MAUDE database and demonstrated the utility of AI-based NLP for postmarket safety surveillance.

 

Background:

The Woven EndoBridge (WEB) device is widely used for intracranial bifurcation aneurysms, highlighting the need for real-world safety monitoring. The FDA’s MAUDE database provides valuable postmarket data, and AI-driven natural language processing (NLP) enables deeper analysis of its unstructured reports to uncover hidden safety patterns.

Design/Methods:

All adverse event (AE) reports involving WEB devices submitted to the FDA’s Manufacturer and User Facility Device Experience (MAUDE) database between March 2019 and September 2024 were systematically reviewed. Device model, event classification, anatomical location, and reporter type were extracted. AI-driven NLP methods –  including Latent Dirichlet Allocation (LDA), t-distributed Stochastic Neighbor Embedding (t-SNE), and term frequency–inverse document frequency (TF-IDF) – were applied to narrative fields to identify semantic themes. Heatmaps were generated to visualize AE patterns across device models.

Results:

474 reports met inclusion criteria. Median patient age was 62 years. Most reports (99.6%) originated from the manufacturer, with 50.6% classified as injury, 43.9% as malfunction, and 5.5% as death. Common aneurysm sites were anterior communicating (38%), middle cerebral (26.7%), and basilar arteries (12.9%). The WEB SL 17 model accounted for 54.7% of cases and most device-related issues, including difficult or delayed separation (34.6%) and separation failure (20.1%). Among 342 patient problems, thrombosis (15%), aneurysm remnant (12.6%), and hemorrhage (9.4%) predominated. NLP identified three themes: device deployment and detachment mechanics, clinical outcomes, and manufacturer investigations, linking mechanical failures with potential manufacturing factors.

Conclusions:

AI-powered NLP enhances postmarket surveillance by enabling scalable, nuanced safety signal detection, supporting regulatory monitoring and device improvement. Real-world WEB device adverse events are more diverse than those in trials, underscoring the need for improved imaging, operator training, and integration of real-world data to optimize safety.

10.1212/WNL.0000000000216345
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.