Improving adverse event detection related to biologic immunosuppressant use – a pilot study of the BERT deep learning model adapted to real-world clinical notes
Our goal is to deploy these models within EHR systems and support FDA’s ability to perform real-time surveillance of drug safety in the post-marketing setting.
CERSI Collaborator (University of California, San Francisco): Dana Ludwig, MD; Madhumita Sushil, PhD; James Buchanan, PharmD; Anna Silverman, MD; Balu Bhasuran, PhD; Atul Butte, MD, PhD; Vivek Rudrapatna, MD, PhD
FDA Collaborators: Jawahar Tiwari, PhD; Lauren Choi, PharmD; Nadia Habal, MD; Artur Belov, PhD; Rebecca Racz, PharmD; Samer El-Kamary, MD, MPH; Qi Liu, PhD; Ohenewaa Ahima, MD; Yan Li, Ph.D., B.Pharm
Project Start Date: February 23, 2021
Regulatory Science Challenge
The accurate detection of adverse events (AE) is critical to ensure that clinicians and patients can make well-informed treatment decisions that balance risks with benefits. This is true of immunosuppressants, a class of medications that are used for the treatment of autoimmune and immune-mediated diseases. These medications are commonly required long-term for the management of these conditions, many of which are otherwise incurable and lifelong. Many of these medications are biologics, newer therapies that are derived from living systems such as cultured cells and organisms. As a class, immunosuppressants are associated with many toxicities because they suppress the immune system. These can include infections, cancer, and other serious conditions. There is great interest in the study of these relatively new medications from a safety standpoint, in large part because there is less collective experience with newer drugs as to the consequences of their long-term use.
Clinical notes within electronic health record (EHR) systems are a rich source of adverse event (AE) data because the treating clinicians often document these events as well as the reasons for treatment discontinuation. However, these notes have been underutilized for surveillance due to their unstructured nature and methodological limitations to effectively examine them using natural language processing. Recently, there has been impressive advances in natural language understanding following the release of the context-aware deep learning model BERT (Bidirectional Encoder Representations from Transformers). However, its adaptation to clinical language has been limited in part due to the unavailability of platforms for processing these sensitive data.
Project Description & Goals
In this project, we are using over 100 million clinical notes at University of California, San Francisco to adapt BERT for clinical language and train it to perform adverse event detection. We will then compare this to existing methods for clinical text processing and structured data analysis regarding the task of accurate AE detection.
We will pilot this study in the context of patients with Inflammatory Bowel Disease who have received biologic immunosuppressive medications. If our model proves successful, we will extend it to other clinical domains and assess its performance on data from other health systems. Our goal is to deploy these models within EHR systems and support FDA’s ability to perform real-time surveillance of drug safety in the post-marketing setting.