Project Lead: Aparna Ramanathan, UW Medicine

Data Science Lead: June Yang and Spencer Wood

Currently in the United States, postmarket surveillance of medical device safety is conducted via the Manufacturer and User Facility and Device Experience (MAUDE) database.  MAUDE is targeted at device manufacturers and does not engage patient or provider users of new medical devices; this lack of stakeholder engagement has led to the injury of hundreds of thousands of gynecology patients.  With the Essure® system of permanent birth control as an example, attention was not drawn to MAUDE database reporting pathways until after patients themselves gained traction by reporting complications through social media platforms. 

Using Essure® as a model, we aimed to acquire social media data related to patient experience with Essure®, build out a natural language processing pipeline to facilitate analysis of these data, and build and test models for identifying posts that express negative outcomes.  The ultimate goal is to prototype a model for detecting early warning signals for safety issues regarding Essure® and other implantable medical devices.  During the accelerator, we acquired Reddit achival data on subreddits most pertaining to Essure® and cleaned the data, filtering for posts related to Essure®.   We then processed the raw text and did some initial data analysis related to work freqencies and word embeddings.  We refined our research question and developed strategies for data labeling so that we could study trends in the data.  I learned how to use an LLM for data labeling.  We explored the ideas of sentiment analysis to try to understand how patient trust and perception changed in Essure® over time.  We then explored topic modeling strategies which we plan to implement in future work towards the question of identifying early warning signals.

Our preliminary data is exciting in that we found a substantial amount of conversation around Essure and trends in words related to known Essure ® complications that suggest that patients were identifying these issues prior to FDA awareness in 2015.  Furthermore, we observed text describing adverse experience with the device, mechanical/materials failure, and user error in the placement of the device – indicating that social media posts likely contain “early warning signals” of future medical device failure.  Importantly, we gained valuable time and experience working together as a multidisciplinary team on this question which has set us up nicely to pursue avenues for internal and external research funding to continue to explore this topic.  

The Github repository can be found here.