IQVIA’s AI vision is to cut pharmacovigilance costs by 50% with superhuman accuracy

Pharmaceutical companies could slash their pharmacovigilance costs by half while significantly improving accuracy through generative AI, according to IQVIA’s Global Practice Lead of Pharmacovigilance Technologies, Uwe Trinks.

“The idea is: we can save 50% of the cost and get the quality to above 99%,” Trinks said. “That means when we have human verification, we only have to change data in one in 100 cases, which is unheard of because right now in QC, we probably have to change data in one in 10 cases. So we humans, in a really good environment, get to about 90% quality.”

What traditionally takes a week could be completed in a single day. “This is a vision we’re working toward,” Trinks added.

Making sense of a mountain of data

Despite advances in techniques such as natural language processing, the pharma sectors continues to face significant challenges in adverse event case processing, with current manual workflows taking up to a week per case. “A large company could have between 3,500 and 4,500 different reporting rules to 800 or so partners worldwide.”

Complicating matters further are disparate inputs. There are informal adverse events reported on social media and discussion groups. Large medical information centers where customers contact call centers and ask questions about pharmaceuticals and sometimes report adverse events in the process, which then must be extracted from voice files.

“Mining those, going through those, and asking, was there any adverse event (AE)? Because the people in this global medical information are not really trained to analyze and get AEs,” Trinks explained. “They have procedures that say, if the patient says anything like that, take the data down and then forward it to drug safety, but they can miss a lot of things.”

To address this challenge, IQVIA developed an AI-powered detection system. “One of the modules of our software, IQVIA Vigilance Platform, is called ‘Detect,’ and that is a Gen AI-based tool that actually reads through billions of data points and identifies those AEs,” Trinks said.

The system also processes structured data through established channels. “The second way we get data is through structured formats from either license partners that we have, from regulators, from other companies, from pharmacies, that use a standard XML file called E2B,” he noted. “This is a standard file from the International Conference for Harmonization of Technical Requirements for Pharmaceuticals for Human Use (ICH), which has been around since 1995, and we use that for a lot of data exchange.”

A third input stream comes through direct collection methods. “Then from standardized collection forms that go out to countries, to other sources that fill them out and send them in, to physicians that fill them in, or to all kinds of electronic portals, either web-based or cell phone-based, that are built to collect adverse events,” Trinks explained. For these structured inputs, traditional coding remains most efficient. “Those forms and portals, because they are highly standardized data collection forms, use standardized code, and there’s no good reason for AI to be involved there.”

Modularizing pharmacovigilance

Just as adverse event data comes through distinct channels requiring different handling approaches, IQVIA’s implementation strategy reflects this need for specialized processing. The vision is segmented into multiple connected elements that Trinks refers to as “bubbles”—discrete AI-powered and traditional manual code processes working in concert. In the diagram below, the light blue bubbles represent where generative AI can be most effective, while dark blue indicates where traditional coding continues to have the upper hand.

“We have these distinct bubbles that do distinct things in these databases and actually put data in, and then the next bubble goes in and does that,” Trinks explained. This modular approach ensures each component is highly specialized and efficient.

The system operates within a carefully controlled environment. “We have our own closed system with our own IQVIA IP that is totally secured from public access,” he said, emphasizing the importance of data security and privacy compliance.

“Privacy of the information is very, very important,” Trinks said. “We have regions like Europe with GDPR… if you violate the privacy of patient data, and you do it maliciously, you can be sued for up to €20 million or 4% of annual sales, whichever is higher.”

Building on its closed ecosystem approach, IQVIA has established strategic partnerships to enhance its AI capabilities. “We have a partnership with Microsoft and OpenAI. What we’ll do is we take the native Gen AI, which is very, very good, and we add to the native AI our internal data,” Trinks explained.

This integration leverages IQVIA’s sizable data resources. “We have billions and billions of data on pharmaceutical products… We have a huge amount of data for mappings,” he noted. The system incorporates multiple standardized databases including “MedDRA… SNOMED from the National Library of Medicine… FDA’s Drug Database.”

A phased approach to implementation

While IQVIA’s vision is comprehensive, the implementation is still in its early stages. The company has completed specific pilots, including one for literature cases, with encouraging results. Some components are already operational, such as the ‘Detect’ module for automated identification of adverse events from unstructured data sources.

However, more complex elements—including case assessment, validity checking, and causality determination—remain in development. “We have some test pilots running, but this is not something that you will see in the next year,” Trinks said. “This will take a lot more refinement.”

The company is taking a measured approach to rolling out different components, with some features expected in the next software release while others are part of a longer-term vision extending to the coming years.

Aligning human and machine talents

The key in IQVIA’s strategy lies not in wholesale automation but in a hybrid approach that strategically deploys AI for tasks like data extraction while maintaining human oversight where it matters most.

Not every process is suited for AI automation. While genAI excels at tasks like data extraction and translation, traditional coded solutions remain superior for highly structured processes like standardized reporting and product coding. “This is done with traditional code. This works extremely well already,” Trinks noted regarding the reporting processes. The key is knowing where each approach provides the most value.

“There is a huge difference if you use GenAI to extract data into data systems or GenAI to actually generate data,” Trinks said. “You can do really bad prompt engineering that creates hallucinations, and you can do really good prompt engineering that reduces some hallucinations.”

Trinks illustrated the point with a comical example: “There’s this famous thing from this image recognition software that was really, really good until they showed it a cow on the beach, and the software repeatedly said, ‘This is a camel.’ When they did the training, the sand overrode the features of a cow.”

Because AI systems can sometimes misinterpret patterns, governance approaches are in place to identify and mitigate bias. “That is what we’re currently having lots of discussions about—how we do that in our governance, what kind of analysis tools can be put on the system to identify bias,” he noted.

Human oversight remains crucial. “Because we have human review, and even if the human review is not perfect, we will hopefully catch all of these [errors],” Trinks emphasized.

To streamline the review process, the system employs a sophisticated confidence scoring mechanism. “The case processor sees immediately, yeah, this is a green one; we have a hard match… Oh no, this is an orange one; we think this is very soft,” Trinks explained. This color-coding system helps prioritize human review efforts, ensuring attention is focused where it’s most needed.

A learning curve

The interplay between humans and generative AI creates a virtuous cycle of continuous improvement. Human verification of AI outputs generates a massive dataset for refining the system.

“If we run this system for a year, we have three-quarters of a million cases going through that system, so that’s a pretty good verification,” Trinks observed. “We have this feedback against the real world, and we feed that back into control tables, and we feed that back into our prompt engineering.”

This ongoing refinement enhances the AI’s performance over time, leading to greater accuracy and reliability.

Keeping genAI accurate and fair

While hallucinations are a common concern with generative AI, IQVIA sees systematic bias as the bigger threat. “When GenAI hallucinates, it usually hallucinates badly,” Trinks explained. Such obvious errors are easily caught by human reviewers.

The greater challenge lies in subtle biases that might consistently steer the system in the wrong direction. “That is what we’re currently having lots of discussions about—how we do that in our governance, what kind of analysis tools can be put on the system to identify bias,” he noted.

The system includes built-in controls to maintain quality when new versions of the AI models are released, using test datasets to “see if anything materially has changed.” Additionally, the software is designed to assess its own confidence levels. “We intend to build the prompt engineering in a way that the software tells you how confident it is that what it’s doing is actually right,” Trinks said.

Human oversight remains crucial, especially in the initial years. “Because we have human review, and even if the human review is not perfect, we will hopefully catch all of these [errors],” he emphasized.

---

Filed Under: machine learning and AI, Pharmacovigilance, Uncategorized
Tagged With: adverse event processing, generative AI pharma, IQVIA drug safety, Microsoft OpenAI pharma, pharmacovigilance automation
 

---