[Adobe Stock]
That drive for swifter, more comprehensive insights has also shaped Flatiron’s next chapter. In May 2023, the company expanded from its strong foundation in real-world data curation and expertise in oncology-specific data to offer an integrated approach to real-world evidence, providing end-to-end support for oncology research—spanning everything from early-phase development through commercialization. And just last week, Flatiron Health announced a strategic partnership with Exact Sciences to accelerate clinical evidence generation for molecular residual disease (MRD) testing.
Chipping away at uncertainty
Blythe Adamson
Uncertainty is one of the biggest hurdles in drug development, which can contribute to billion-dollar write-offs, stall promising therapies, and delay patient access to life-changing treatments. “Real-world data,” Adamson said, “is our most powerful weapon against that uncertainty.” She explained: “Whether it’s sharpening our focus on the right targets, quantifying benefits for diverse populations, or rigorously assessing safety signals, RWD helps us make data-driven choices.” In practice, that can streamline the journey from lab bench to bedside, accelerating the delivery of promising therapies. By incorporating evidence from real-world settings, stakeholders can better decide which treatments to advance. And Flatiron Health is strategically positioned to harness its database of more than 3.5 million de-identified patient records—helping to generate real-world evidence that supports faster decision-making.
Developing new medicines is “a huge risk,” financially and in terms of time, and “there’s a lot of uncertainties that go along the way,” said Adamson, who we included in the article “By 2025, clinical research and patient care converge: Data integration, validation, and evolving markets.” To state the obvious, drug failures can occur at any stage and become more costly as they approach potential commercialization. “For me, it’s incredibly rewarding to identify the right data, the right methods, that can reduce that uncertainty at each phase.”
Unlocking trapped data
While real-world data can reduce uncertainty, it often comes in unwieldy forms. “When I joined Flatiron seven years ago,” Adamson recalls, “I saw how much clinically meaningful information was trapped in unstructured documents—physician notes, faxes, and long PDFs of genomic results—none of which epidemiologists were really trained to handle.”
Two AI technologies now helping Flatiron make sense of this data deluge are traditional natural language processing and generative AI. Combined, they extract critical insights from millions of patient records, including those for rare cancers. The goal is to liberate the untapped potential in unstructured EHR data and turn it into regulatory-grade evidence that guides treatment pathways and accelerates real-world research.
“I think a lot of groups still use both classic NLP and large language models. There’s a workflow where NLP identifies the documents of interest, then we send only those relevant documents into the LLM,” Adamson explains. This hybrid method excels at locating data points—such as biomarker percentages—buried in hundreds of pages of clinical records. “These tasks can be incredibly time-consuming for a human abstractor—sometimes 30 to 45 minutes just to locate the right data. With LLMs, once you’ve narrowed it down, the model can interpret tables, figure out negative versus positive controls, and extract precise information much faster.”
Validating AI with “golden data”
According to Adamson, Flatiron’s decade of “golden data” painstakingly curated by human chart reviewers is critical for verifying AI outputs. “We can compare how accurately the model extracts information to what a trained abstractor would find,” she says, “and that puts us in a unique position to validate the AI for regulatory or clinical decision-making.”
Still, Adamson underscores that human expertise is indispensable. “Our AI is only as good as the data we train it on,” she explains. “Ultimately, AI accelerates the process, but human judgment—resolving ambiguities, spotting patterns, and applying clinical reasoning—remains the foundation of reliable evidence generation.”
Multinational data poses standardization questions
With Flatiron’s expansion into the UK, Germany, and Japan, Adamson has gained a global perspective on evidence generation. “Each country has different standards of care, treatment guidelines, and data privacy requirements,” Adamson said. “While we want a common data model for multinational studies, we also have to respect local healthcare contexts and partner with the people who know them best.”
Adamson underscored that guidelines, privacy laws, and attitudes toward data sharing vary widely from country to country. Adapting AI and data models to these distinct realities, she explains, is crucial to producing meaningful, globally relevant evidence.
Yet as AI-driven platforms like Flatiron Health expand globally, they confront a paradox: the very complexity of healthcare systems that demands innovation also resists universal approaches. “With every new country, we’re reminded: There’s no one-size-fits-all approach,” Adamson said. While AI offers tools to streamline this complexity, its success hinges on respecting local nuance. “You can’t just copy-paste your data model from the U.S. into Japan or Germany,” she said. “Success means partnering with local clinicians who know the system inside and out.” In fact, a new publication from Adamson and colleagues provides the first-ever characterization of Flatiron’s EHR-derived real-world data in the UK, Germany, and Japan—showing how local regulatory requirements, data governance, and anonymization processes can help support effective multinational oncology research.
Filed Under: machine learning and AI, Oncology
