Flatiron Health Enables Patient-Level Data Sharing Across National Borders

By Deborah Borfitz 

March 18, 2025 | Patient-level real-world data (RWD) derived from harmonized electronic health records (EHRs) across four countries will help power multinational oncology research of the future, thanks to groundbreaking, collaborative efforts of Flatiron Health scientists working in the U.S., UK, Germany, and Japan. Although the research community was initially skeptical that this was even possible, investigators are now lining up to get into the secure, cloud-based environment where they can analyze the cancer-specific datasets from anywhere in the world, according to Blythe Adamson, Ph.D., the company’s international head of outcomes research and evidence generation who led the ambitious project.  

Building on the experience and learnings from the United States Flatiron Health Research Database, developed over the last 12 years, the team developed a common data model specific to each cancer type, she says. “We went cancer type by cancer type with oncologists from each of these countries... to [collaboratively] define what is clinically meaningful” on a global scale. 

“There is so much to learn from the differences between countries with different clinical guidelines and treatment pathways and many of the questions are unanswered about which treatments and pathways can give patients the best outcomes,” says Adamson. Another valuable opportunity with a multinational dataset is to provide drug regulators with a geographically diverse real-world control arm for assessing treatment effectiveness in lieu of a single more homogeneous control group. 

Creation of the datasets was a herculean effort, she points out, which is the main reason no other group has done this before. The regulatory and standardization barriers were “very high,” notably the EU’s General Data Protection Regulation (GDPR) and Japan’s Act on the Protection of Personal Information (APPI) that made cross-border patient-level data sharing a highly complex exercise. 

The value of integrating the EHR-derived datasets in support of multinational oncology studies was laid out in an article that was published recently in ESMO Real World Data and Digital Oncology (DOI: 10.1016/j.esmorw.2025.100113). In it, Flatiron scientists, oncologists, and engineers outline the methods ensuring the trustworthiness, reliability, and relevance of EHR-derived datasets—namely, alignment with the International Society for Pharmacoeconomics and Outcomes Research (ISPOR) EHR-derived data SUITABILITY checklist. 

“Our approach prioritized data cleanliness, security, and compliance, and this paper helps explain the systems of provenance and governance that allowed us to show full traceability of these data sources” and alignment with regulatory frameworks like GDPR and APPI, Adamson says. Rigorous techniques such as pseudonymization, obfuscation, redaction, and masking were used to protect data privacy while preserving the data’s usability. 

Flatiron Health spent years trying to figure out how to comply with all the local laws and regulatory environment in each country, which are “completely different” in the U.S., UK, Germany, and Japan, says Adamson. And this is to say nothing about the challenges related to the differing clinical depth of documentation, clinical visit cadence, and where and how data gets collected. 

Even within countries, data inconsistencies were more the rule than the exception. In Germany, for example, cancer care is delivered through a scattering of small clinics throughout the country and data is stored differently at every single one of them, Adamson says. 

Unique Resource 

Importantly, integration of the deidentified or anonymized patient-level datasets was enabled by the development of a secure, cloud-based “trusted research environment” with controlled access and standardized data harmonization, Adamson says. This provides a “secure room” for researchers to analyze the datasets from anywhere with an internet connection. Data sharing across national borders has traditionally been limited to aggregated data, making the idea of a multinational patient-level dataset unthinkable. 

It was decided many years ago that the UK, Germany, and Japan would be the geographic starting points, she adds. This was after carefully surveying many key countries based on their digital infrastructure, documentation patterns, privacy and regulatory environment, and healthcare provider landscape, as well as the interest of government healthcare technology assessment bodies in having local data and the demand from pharmaceutical companies for RWD to use in research. 

Creation of the oncology dataset was a first not only across countries but within them, notes Adamson. “Nothing exists like this from electronic health records that are curated longitudinally.” The information from all locations gets refreshed at least every 90 days, so patients’ experiences can be followed in real time. 

This includes access to all the rich, unstructured data in documents, including PDFs of 50-page genomic testing reports and short digital texts entered by physicians about patients’ visits. “It’s a dataset that’s not found in any other real-world data sources,” she says, a list that includes claims data from insurance records, information about clinical studies in the Protocol Registration and Results System hosted by the NIH National Library of Medicine, and data collected from one-off chart reviews to answer a single research question. 

Custom Integrations 

Among the many challenges in harmonizing the EHR-derived RWD datasets is that different countries use different biomarker tests to guide cancer treatment, Adamson offers as an example. Documentation practices, which are aligned with the unique reimbursement pathways in each country, also differ widely. 

There are likewise conflicting thresholds regarding how healthy a patient needs to be to receive treatment due to concerns about tolerability of the side effects, she continues. “In Japan that would be documented through handwritten notes, whereas in the UK it might be a structured field from a drop-down menu of numbers.” 

Local oncologists within each country helped Flatiron understand where and how the information was collected, so it could be harmonized in a way that provided content and concept validity of what the clinical variables meant, says Adamson. Differences in what doctors felt were most critical to document in support of their clinical decisions revealed itself through the missingness of different variables between countries, she points out. 

To remedy data inconsistencies between and within countries, software engineering teams worked on-site with hospitals and clinics to integrate their EHR data into the Flatiron system. “We had to build custom technical integrations to harmonize this data from diverse EHR systems,” Adamson explains. 

“It is a demonstration of the effort it takes to build this global health infrastructure,” she adds. Now that exacting work has been completed, oncology researchers everywhere might now reap benefit by accessing the resulting datasets. 

Team on the Ground

Three years ago, Flatiron Health expanded its footprint with the creation of subsidiaries in the UK (London), Germany (Berlin), and Japan (Tokyo) and Adamson says she visits them all regularly to work with teams the company built in those countries. These include medical oncologists, software engineers, data scientists, and data insights engineers. 

The software engineers are the ones who wrote the code to create the pipeline for the transfer of documents from multiple EHR systems into the Flatiron systems within each local country, she explains. Oncologists, being clinical data experts, helped define disease-specific data models and worked alongside cancer registrars who opened patient charts and pulled out the information. Data scientists were responsible for developing patient- and cohort-level quality assurance processes to ensure the data was accurate, reliable, and relevant for answering research questions. 

Once the datasets were built within each country and deidentified or anonymized, they were uploaded to the trusted research environment. A properly permissioned researcher anywhere in the world can virtually access the web-based platform to analyze the information in the R or Python programming language or interact in the dashboard to look at different subgroups of patients, says Adamson. “I think very soon you will start to see abstracts and manuscripts coming out that are answering important research questions with these datasets.” 

Flatiron’s multi-disciplinary team of experts worked together to answer every question on ISPOR’s SUITABILITY checklist about how the datasets were made and suited for research, she says, which was “harder than expected” to fill out. The framework was only formally published last summer, and Flatiron was the first user. 

Extraction and Transportability

Before expanding the project to any other nations, Flatiron Health is growing its site network within the UK, Germany, and Japan to ensure the datasets are capturing geographic variability within each of those countries, reports Adamson. The datasets are rapidly expanding to include about four times more patients powered by 10 or more new sites across all markets this year. 

Meanwhile, the company is expanding its use of artificial intelligence and machine learning to extract certain data elements from EHRs. In a paper Adamson and her colleagues recently published in AI in Precision Oncology (DOI: 10.1089/aipo.2024.004), large language models (LLMs) were used to extract PD-L1 biomarker data from EHRs, which has relevance across cancer types and plays a role in treatment decisions. This was done in the U.S., but it points to the possibilities within the multinational datasets, she notes. 

Researchers found that LLMs “fine-tuned” using high-quality labeled data consistently outperformed the zero-shot approach (no fine-tuning), achieving high accuracy across 10 cancer types and multiple biomarker details, including test and result dates, cell type, and percent staining. Particularly noteworthy was that the fine-tuned LLMs outperformed a deep-learning model trained on over 10,000 examples, despite using far fewer labeled examples (500–1,500). 

The reality is that “not every country will be able to have the data infrastructure to have high-quality real-world data to make decisions” and may therefore find it necessary to “borrow” information from another country, Adamson says in referencing Flatiron FORUM (Fostering Oncology RWE Uses and Methods). This is the company’s consortium for collaboration with sponsor companies to advance the impact and outcomes of real-world evidence, which launched back in 2023. 

As reported in the latest ESMO paper, Flatiron FORUM will be exploring the “transportability” of RWE across country borders—that is, when it is appropriate and suitable to take evidence created in one country and apply it to decision-making in another country. Members of the group will be doing a collection of validation studies specific to breast cancer, lung cancer, and multiple myeloma to determine the accuracy of predictions about drug effectiveness in the UK, Germany, France, and Austria compared to the truth of how patients fared. 

“If you can borrow information [from the U.S] and make adjustments to apply it to a different country’s patient characteristics, then it really opens up the opportunity to use synthesized data from all around the world to make decisions about drug effectiveness and stewardship of medicines in that country,” says Adamson. “We’re going through this matrix of different countries and different cancer types and these papers... will be referenced as supporting evidence for decision-making to justify whether or not it is appropriate to use information in that cancer type borrowing from a different country.”