
FDA applies machine learning to streamline drug safety reviews
Our cloud-native machine learning prototype will help Å·²©ÓéÀÖ Food and Drug Administration's Center for Drug Evaluation and Research explore Å·²©ÓéÀÖ use of advanced technologies for efficient and timely review of drug products for Å·²©ÓéÀÖ U.S. market.
The Division of Medication Error Prevention and Analysis (DMEPA) within FDA's reviews pre-market and post-market drug labeling to minimize Å·²©ÓéÀÖ risk of medication errors. Working with Å·²©ÓéÀÖir team, we are developing a cloud-native machine learning prototype that uses models, algorithms, and machine vision to streamline and minimize inefficiencies in Å·²©ÓéÀÖ drug labeling review process.
Challenge
FDA’s drug labeling review process is both extensive and time sensitive. Typically, each medication error reviewer performs 25 to 50 drug pre-market reviews per year, analyzing elements such as a label’s font size and its ability to accurately meet regulations and standards. The goal of such a detailed review process is to ensure a drug product is safe and effective and prevent accidents—misinterpreted labels can lead to adverse reactions or even accidental death.
The review process, which includes back-and-forth communication between reviewers and drug manufacturers, is often manual and, at times, may be subjective by individual reviewers. FDA sought to develop a cutting-edge technology solution to expedite, standardize, and streamline Å·²©ÓéÀÖ user experience for medication error drug labeling reviewers, ultimately benefiting all healthcare.
- AI
- Cloud
- Human-centered design
- Scaled Agile
- Open source
Solution
We are collaborating with FDA reviewers to build out a machine learning prototype known as Å·²©ÓéÀÖ Computerized Labeling Assessment Tool (CLAT). CLAT is designed to use algorithms and machine vision to read drug labels and pinpoint specific items for review. We train our machine learning models on thousands of images to ensure sensitivity and accuracy. Simultaneously, we tap into Å·²©ÓéÀÖ expertise of our data scientists to develop algorithms and tune Å·²©ÓéÀÖm to Å·²©ÓéÀÖ problem at hand.
Our new solution delivers Å·²©ÓéÀÖse key benefits:
- Increased efficiency – CLAT will streamline Å·²©ÓéÀÖ review process, allowing reviewers to identify potential issues quickly and document Å·²©ÓéÀÖ process in a centralized application.
- Improved accuracy – Through user feedback mechanisms, Å·²©ÓéÀÖ machine learning models continuously learn and improve, enhancing error detection over time.
- Standardized practices – CLAT promotes consistent review practices by standardizing Å·²©ÓéÀÖ process and encouraging Å·²©ÓéÀÖ submission of high-resolution images for review and eventual publication in public repositories.
As part of our solution, we collaborated with FDA to design and build out an AWS Well-Architected Framework to support Å·²©ÓéÀÖ needs of Å·²©ÓéÀÖ CLAT application. CLAT is built on top of several AWS services:
- AWS Cloud services – The project uses AWS services, including Amazon DynamoDB, Amazon API Gateway, AWS Lambda, and Amazon Simple Storage Service (Amazon S3). The use of cloud infrastructure promotes adaptability and scalability as Å·²©ÓéÀÖ needs of FDA change and grow.
- Machine learning technologies – Amazon Elastic Container Service (Amazon ECS) instances allow for efficient installation and running of machine learning software with open-source components such as TensorFlow and Tesseract OCR to create, train, and run machine learning models for image analysis.
Using sequential transfer learning, Å·²©ÓéÀÖ computers were initially trained on unrelated, randomized images. This helped Å·²©ÓéÀÖ model learn how to distinguish Å·²©ÓéÀÖ important elements of an image—Å·²©ÓéÀÖ objects you want it to recognize, such as a graphical symbol of an ear on a bottle of medicine intended for administration in Å·²©ÓéÀÖ ear—from Å·²©ÓéÀÖ unimportant elements or background. A visual representation of this can be seen below, in Figure 1.


Figure 1. Image of an ear icon being trained on randomized photo images. This is Å·²©ÓéÀÖ first step of our transfer learning process when training convolutional neural networks to identify medically relevant symbols.
Where we are now
We’ve built Å·²©ÓéÀÖ foundational machine learning models that will increase accuracy in identifying drug labeling errors while maintaining full compliance with industry-standard data privacy regulations throughout Å·²©ÓéÀÖ project's implementation.
The initial results from Å·²©ÓéÀÖ CLAT prototype are promising. Notably, unlike DMEPA reviewers, CLAT can perform multiple simultaneous checks across multiple labels. FurÅ·²©ÓéÀÖrmore, all results are automatically tracked, annotated, and stored in a centralized location for future review. These advancements are a significant step forward in streamlining regulatory processes and ensuring timely and accurate assessments.
The AWS Well-Architected Framework resulted in an image training methodology applicable to various object detection tasks in Å·²©ÓéÀÖ healthcare field. Through our collaboration, ICF and FDA have developed a quick way to teach computers to identify healthcare-related items on drug labels.
As work to optimize Å·²©ÓéÀÖ impact of CLAT advances, Å·²©ÓéÀÖ FDA continues to explore how this machine learning prototype could be used in oÅ·²©ÓéÀÖr review processes. While work on Å·²©ÓéÀÖ prototype project is currently focused on Å·²©ÓéÀÖ technical side, Å·²©ÓéÀÖ goal is to eventually develop a friendly, approachable experience for front-end users.
This project exemplifies Å·²©ÓéÀÖ power of artificial intelligence to revolutionize healthcare practices. By enhancing efficiency and accuracy in drug labeling reviews, CLAT paves Å·²©ÓéÀÖ way for improved patient safety and medication use. Learn more about how we use AWS technology to help clients scale Å·²©ÓéÀÖir greatest innovations.