I had the opportunity to attend and present at the European Society for Medical Oncology’s (ESMO) first congress dedicated to artificial intelligence (AI) and digital oncology, ESMO AI, in Berlin this year and am really excited to share some of my perspectives. This event aligned perfectly with two areas we knew would be essential for Flatiron’s clients — and for our own growth — in 2025 and beyond: using the latest advances in AI to uncover novel patterns and extract insights from the data, and predict future trends through advanced modeling, such as digital twins; and expanding our evidence solutions across Europe so we could better understand local data challenges, conduct transportability studies, and explore early opportunities for cross-border data harmonization.
More than 1,000 attendees from over 60 countries gathered for this inaugural event and I left with a deep sense of confidence: first, that our AI research at Flatiron is firmly aligned with where the field is heading, and second, that we are contributing meaningfully to a broader, industry-wide movement to transform cancer research and redefine patient care.
Here are three key takeaways from the conference that I believe will shape the field in 2026:
Takeaway #1: Expect AI innovation across the oncology landscape
I was struck by the breadth of AI applications emerging across the oncology ecosystem. We’re all focused on relatively narrow applications — in my case, it’s testing and validating novel uses of multimodal real-world data and AI, starting with predictive modeling and digital twins, to strengthen the ability to extract insights from EHR-based real-world data — and it’s really inspiring to see what everyone else is doing with AI.
In the study that I presented in Berlin, models were trained using real-world data from nearly 10,000 patient records in the U.S. to predict overall survival for patients with advanced NSCLC receiving standard of care treatments, unlocking the potential for digital twins use cases. These models incorporated custom LLM to extracted variables (like autoimmune comorbidity and other reasons for immunotherapy omission) as features into the digital twin model to reduce confounding by indication and predict overall survival more accurately. While I did see other researchers present work on similar topics including AI-driven EHR curation, modeling disease progression and predicting treatment response, there were also completely non-overlapping use cases such as predicting biomarker mutation status from digital pathology, reducing inter-observer variability in RECIST measurements, and streamlining patient-trial matching, to name just a few. There was also breadth in the AI approaches being used to tackle these use cases, from task specific models to foundation models to multi-agent systems.
Takeaway #2: Wider AI adoption won’t happen without trust
Despite all the excitement around AI innovation, a clear theme ran through nearly every stage presentation and most conversations on the exhibit floor: trust in AI remains unfinished business. Clinicians and researchers are, by nature, appropriately skeptical, and they won’t adopt AI tools if they doubt the underlying data, can’t follow the model’s reasoning, worry about hallucinations, or see performance that falls short of established standards of care.
ESMO is paving the way with ELCAP and EBAI, its recently published guidances on the use of LLMs and AI biomarkers in clinical practice, and it’s in that same spirit that we at Flatiron developed the VALID framework, a comprehensive methodology for assessing the quality of LLM-extracted real-world data. It’s a transparent, multi-pronged framework that allows us to benchmark LLM performance against expert human abstraction at the variable level, detect internal inconsistencies or implausible values, and test the fitness of LLM-extracted data through rigorous replication studies. It’s the first framework of its kind, and we believe it will play an important role in building confidence in AI-extracted data and raising the profile of RWE in oncology.
Takeaway #3: From siloed studies to global insights
My third and final takeaway stems from seeing so much compelling research emerging from so many countries, each shaped by its own patient demographics, regulatory frameworks, reimbursement models, treatment options, and clinical patterns. It raises an age-old question: can insights generated in the US truly inform case studies in the UK, Germany, or Japan? Or are we destined to stay siloed, replicating the same studies over and over — with all the associated costs, delays, and implications for patient access?
Historically, study transportability has been a formidable challenge, but AI is beginning to change that calculus. At Flatiron, AI is helping us expand the scale and granularity of our RWD assets around the world and harmonize much of that data across diverse health systems — and that progress is opening the door to a new class of research questions. We showcased a dozen such multinational studies at ISPOR Europe 2025 a few weeks ago, and in Berlin we took another step forward by testing our VALID framework on pseudonymized, LLM-abstracted German and British patient data using a GDPR-compliant private data network. You can expect to see far more of this kind of cross-border, collaborative work in 2026 — not just from us, but across the research community.
Would you like more information about how Flatiron is applying AI methodologies to generate global real-world evidence? Please reach out to the Flatiron team on our website.
