AI analysis speeds FDA drug reviews
Connecting state and local government leaders
An artificial intelligence-based platform helping the Food and Drug Administration analyze adverse drug reports is expected to save reviewers about 7,500 hours per year.
The use of an artificial intelligence-based platform to analyze forms is expected to save Food and Drug Administration reviewers of adverse drug report about 7,500 hours per year.
ABBYY’s Digital Intelligence platform transforms as many as 120 fields in safety reports on potential adverse interactions during clinical drug trials into structured content that reviewers can check at a glance. Previously, reviewers had to manually study each of the 40,000 safety reports FDA receives each year before a drug goes on the market and review 2 million reports a year after it’s released, said Justin Scott, senior business informatics officer at FDA’s Office of Business Informatics.
To review the reports, analysts would open each file and spend about 10 minutes reading it. For about 10% of the forms, they would spend another 15 to 20 minutes writing a review about a likely relationship between the reaction and the drug, he said. That alone added up to about 1,000 hours, he added.
Now, the analysts get extracted and structured information so “they can basically go to a series of reports, take a glance at it and understand, ‘All right, this is how many people took this product and were hospitalized, this is how many of them were hospitalized and subsequently died,’” Scott said. Analysts can now quickly gather high-level information “instead of having to open dozens and dozens of documents and take notes,” he said. “Before, where they might get fatigued from reading all these PDFs and notes, now they have the data just presented right in front of them.”
The platform eliminates the need for a human reviewer to transcribe or enter the information manually, said Bruce Orcutt, ABBYY’s senior vice president of product marketing. “It delivers better accuracy and gives more information for analytics and other analysis that are required to either approve something or do risk profiles and other analysis,” Orcutt said.
FDA worked with ABBYY to train the platform on all the ways safety reports, each of which can consist of five to 100 pages, can be structured. They made sure it learned the exceptions and permutations that the forms might have and set up a “bucket” where anything that’s not recognized goes for additional analysis.
“We found that we can just drop 100,000 files in, have them process over the course of a week and almost all of them were recognized,” Scott said.
FDA uses the platform not only for current drug safety reports, but also to help process its 30-year archive on adverse events. The platform was 99.3% accurate out of the gate, Scott said.
One feature he pointed to is ABBYY’s FlexiLayout, “which enables a data capture application to locate data fields on the documents and extract information from these fields,” according to the company. That means if a safety report has a section that spills onto the next page, the platform recognizes that without Scott’s team having to build new templates for each one-off.
The platform has also contributed to overall efficiency, Scott said. “We have our reviewers actually working on the scientific analyses that they are hired and paid for as opposed to working on administrative tasks to gather the data,” he said. “If this person doesn’t have to spend 10 minutes reading a document, 20 minutes writing a review, but they can just take a look at it at a glance and understand what’s going on, that means that they can move on to the next object in the pipeline that much sooner.” And the faster the pipeline flows, the more treatments FDA can approve, he said.
The implementation of the platform is part of FDA’s larger Technology Modernization Action Plan, which came out at the end of 2019. The agency aims to address the silos that have grown around its component organizations based on their various responsibilities: oversight of human drug products, food, cosmetics, animal products and medical devices, for example.
“They all have over the years developed their own mom-and-pop IT departments, and they are very siloed,” Scott said, adding that the departments are often protective of their purpose-built IT, making it unscalable and unable to be shared.
FDA is in the process of moving to the cloud, but it must overcome some security and interoperability challenges first, he said.