Skip to main content

Autocoding adverse events to MedDRA – time to throw out the manual

MedDRA autocoding with natural language processing

In an ideal world, adverse event reports would fit neatly into MedDRA, the standardized medical terminology for all regulatory submissions, so they could be easily reported and assessed for patterns to inform safety decisions. To someone outside of the pharma industry, this probably seems achievable. But as anyone working in drug safety today will tell you, capturing a holistic view of a product’s safety profile has never been that straightforward, and it is only getting more challenging. The reason is simple: adverse events are reported in natural language, and the reporters aren’t MedDRA coders. They are nurses, physicians and patients, each with a unique way of expressing themselves. To add to the complexity, these reporters have more reporting routes available than ever before, creating a deluge of natural language safety events that must be fully captured and understood.   

One adverse event, many expressions

To illustrate the problem, consider how a patient might describe their experience after taking a drug. Perhaps they would tell their physician, “I had a horrible headache and couldn’t sleep for two days.” The word “headache” is a one-to-one match with MedDRA, so it will auto code quite nicely. But “couldn’t sleep” would not be understood. It needs to be coded as “sleeplessness” in MedDRA. Finding that correct code requires a manual database search that takes up valuable time – it may take a minute or less but can be up to 30 minutes for a single event. Now consider that in some pharma companies, about 70 percent of their adverse events require manual coding, and the flood of new reports never stops. 

There’s no negotiating on patient safety, and yet drug companies can’t sustain this level of manual reporting. Whether you're working in clinical safety or post-market safety and surveillance, you need help to build a systematic, comprehensive view of the relevant data. You need Natural Language Processing, or NLP.

Harness your internal and external safety data

Today’s NLP technology can free you from manual coding by effectively “reading” adverse event reports in their natural language and standardizing them to MedDRA. That means that even if a single medical concept is expressed in multiple different ways across reports, NLP can, as a human would, understand the nuance in language and properly code the event.

Here in the NLP universe, we have built solutions that can understand what we call morphological variants, or predictable changes a word undergoes as a result of syntax (think of words like patient/patients/patients’/patient’s or takes/took/is taking). Using NLP, we can extract the appropriate context of each use, as well as matching across conjunctive words (like capturing “liver and kidney toxicity” as two distinct events despite their conjunctive link). Today’s NLP can capture common spelling mistakes and code them appropriately using the surrounding context without manual intervention.

There are some powerful broader implications once a pharma company embraces this technology that go beyond manual effort reduction. At every stage, critical data are being both generated and sought from unstructured text – from internal safety reports, scientific literature, individual case safety reports, clinical investigator brochures, patient forum, social media, conference abstracts. Only about 20 percent of that data is structured and easily analyzed. Intelligent NLP search across these hundreds of thousands of pages is the only feasible way today to capture the other 80 percent. By unlocking the full spectrum of information about how patients experience products, NLP delivers an essential tool for key decision support across pharma.

Use case: A top 50 pharma company

As I’ve mentioned, the uses of NLP across the safety product lifecycle are broad, but I’d like to focus on how one of our clients has used it to solve a common challenge today in pharma: coding adverse event verbatims into MedDRA. This client, a top 50 pharma company, was only able to autocode about 30 percent of their verbatims when they came to us, meaning 70 percent had to be manually coded. This client works in the rare disease space, so about 90 percent of the verbatims they encountered were unique, or something the manual coders had not come across before and wouldn’t be repeated in the future.

Prior to advances in NLP, what most companies would do in this situation is create a synonym list to try to capture words that are different but have the same meaning. But for this client, with 90 percent of verbatims never recurring, creating such a table was a tall order. It also needed to be maintained every half year for MedDRA reversion. The client needed another solution.

The IQVIA NLP team (Linguamatics) worked with this client to deploy our award-winning NLP in a way that improved their auto-coding capabilities. We collaborated with their medical team to create MedDRA mappings and compare those to manual coding for both the adverse event and the indication. In the end, we were able to double the level of auto-coding, moving from 30 to 60 percent with a very low level of mismatch. As a result, we not only improved the coding consistency, but also reduced risk for case processing in medical evaluation.

Today, this client is already thinking about the next version of the tool and how they can partner with us to reap even more benefit from NLP. To that end, we are working with them to create a MedDRA mapping NLP tool that can be more deeply integrated into their internal safety systems and workflows. In the beta program, the system will learn with each intake, making the algorithms better every day. This is something we see time and again with our NLP clients – once we show them what our NLP can do, they are eager to invest more deeply to harness their data in a truly transformational way.

What are the ‘hidden meanings’ in your safety data?

Marti Hearst, a professor at the University of California Berkeley, once explained text mining as the discovery by computer of new, previously unknown information by automatically extracting and relating information to reveal otherwise hidden meanings. I like this explanation because it gets straight to the heart of the value of NLP: making connections and revealing truths that may have otherwise gone unnoticed. Patient safety is paramount at all stages in the life history of a drug, and with today’s technology, gathering, analyzing and transforming data into actionable information has never been easier.

Contact us today to learn how you can use natural language processing to synthesize information from many sources, provide evidence for business decisions, and identify those novel connections that drive better value from your safety processing.

Ready to get started?

Request a Demo

Questions? Ask our experts