Online Program Home
  My Program

All Times EDT

Abstract Details

Activity Number: 75 - Contributed Poster Presentations: Biometrics Section
Type: Contributed
Date/Time: Monday, August 3, 2020 : 10:00 AM to 2:00 PM
Sponsor: Biometrics Section
Abstract #312356
Title: Classification of Adverse Events from the Apple Heart Study Using Natural Language Processing
Author(s): Rebecca Gardner* and Santosh Gummidipundi and Mai Nguyen and Sushmitha Tallapalli and Vidhya Balasubramanian and Justin Lee and Ariadna Garcia and Haley Hedlin and Christoph Olivier and Justin Parizo and Kenneth Mahaffey and Marco Perez and Mintu Turakhia and Manisha Desai
Companies: Stanford University, Quantitative Sciences Unit and Stanford University, Quantitative Sciences Unit and Stanford Center for Clinical Research and Stanford Center for Clinical Research and Stanford University, Quantitative Sciences Unit and Stanford University, Quantitative Sciences Unit and Stanford University, Quantitative Sciences Unit and Stanford University, Quantitative Sciences Unit and Stanford Center for Clinical Research and Stanford Center for Clinical Research and Stanford Center for Clinical Research and Stanford Division of Cardiovascular Medicine and Stanford Division of Cardiovascular Medicine and Stanford University
Keywords: Natural language processing; Clinical trials; Adverse events; Deep neural networks; BERT; NLP
Abstract:

The Apple Heart Study was a prospective, pragmatic, virtual study of the Apple Watch’s irregular pulse notification application. 419,297 participants enrolled in the study and data was collected from November 2017 to February 2019. Participants reported a total of 2248 potential adverse events, of which 1038 (46%) were classified as qualifying adverse events. The decision to classify a reported event as qualifying was determined by the study’s safety desk. With a large pragmatic trial, this process can be burdensome. Using Natural Language Processing (NLP) with the pre-trained model BERT (Bidirectional Encoder Representations from Transformers), we can mitigate this burden and leverage patterns in the adverse event text in order to accurately classify reported events. We will discuss the properties of our algorithm in terms of sensitivity, recall, precision, F1 score, and accuracy. Further, we will compare its performance to other contemporary models, including Clinical BERT, distilBERT, RoBERTa, and XLNET. NLP can be useful for future low-risk, pragmatic clinical trials to aid in identification of adverse events.


Authors who are presenting talks have a * after their name.

Back to the full JSM 2020 program