Amy Abernethy, MD, defies the typical silos of health care. She has navigated the entire ecosystem of cancer, from the bedside at Duke University to the data-driven boardrooms at Verily and Flatiron and the regulatory halls of the U.S. Food and Drug Administration (FDA) as a Principal Deputy Commissioner. Recently, she moved on to launch an investment firm called Highlander Health. She is a visionary who doesn’t just theorize about the future of medicine and oncology but is building it.
Doug Flora interviewed Amy for “The State of AI in Precision Oncology,” the journal’s third annual virtual summit, broadcast in December 2025.
This interview was lightly edited for length and clarity by AI- and human hand.
Doug Flora: Amy, you are the co-founder of Highlander Health. When you stepped away from your last role, the world paused, wondering where you would land. We were excited by the announcement that you and Brad Hirsch were partnering again. Tell us about how your journey led you to this unique place in medicine.
Amy Abernethy: My entire career has been focused on determining how to get safe and effective treatments to the people who need them as quickly as possible. This focus guided my work from my melanoma clinic at Duke to Flatiron, the FDA, and Verily. I realized that making progress requires bringing all “swim lanes” together. My longtime partner, Brad Hirsch, and I stepped back to assess what was needed, concluding that we needed to apply capital and effort to solve this problem.
We started Highlander Health as an investment firm with two parts. First, the Highlander Health Institute makes non-profit investments, providing grants to health systems and others to test new ideas in clinical research. The second part is Highlander Health Partners, which makes commercial investments in companies contributing to the modern clinical research and personalized health care infrastructure of the future. As oncologists, Brad and I recognize the necessity of clinical research and working treatments to personalize cancer care, a concept we believe generalizes across all therapeutic areas.
Flora: How has your integrated optimism guided this? Most people view health care challenges as too hard, but your approach has always been the opposite.
Abernethy: The key is truly believing in the possible. It means seeing something and realizing it can be better for the patient in front of me and scaled up to benefit whole populations. This requires vision. Having worked across many different “swim lanes,” I know that solving big problems involves bringing diverse solutions to the table and finding the intersection among them. This involves valuing all methods of problem-solving—whether from the academic space, community health systems, tech companies, or formal academic research—and believing they all have value to contribute.
It also requires being able to cheerlead and build energy toward solving problems. Furthermore, to solve these issues, you must foster communication and build a common lingua franca. This positive energy comes from not only believing in the possible but also demonstrating to each actor that their tools and contributions, as well as those of others, are valuable and showing them how to collaborate.
The final element is developing mantras. During my time at the FDA managing COVID-19, I reminded myself and others that we can do hard things, and hard things are worth doing. Building simple sayings reminds us of the “why.”
Flora: One of your famous mantras is that trust is the glue that holds it all together. If data are the currency in AI, trust is the bank. How do we build or rebuild trust, both for hesitant providers and for patients, regarding our ability to move data and break down silos?
Abernethy: This is a big, but critical, hill to climb. There must be clarity for everyone regarding “what’s in it for me” and the “why.” A key challenge for care providers (physicians, nurse practitioners, etc.), patients, or citizens, is clearly understanding the benefits of leveraging their data and ensuring the data are not used contrary to their intent.
We must be transparent without using overly legalistic language or medical gobbledygook that hides the true intent and feels false. Another task is ensuring the outcome is tangible; if my data are leveraged, is there valuable, recurrent information provided back to me that helps my day? This recurrent currency for providers and patients is crucial for rebuilding trust. We must recognize that building trust is slow, but its erosion is fast. We need to ensure that as meaningful solutions emerge, we protect that nascent trust from rapid erosion.
Flora: You made leaps of trust yourself. In 2014, you left a prestigious academic post at Duke—where you had over 500 publications—for a tech startup. This must have felt like a gamble. What did you see that made you believe that real-world data at Flatiron could change oncology? That was prescient.
Abernethy: At the time, I think everyone was wondering what I was thinking! There were two key factors. First, serving on the board of directors of Athena Health showed me that cloud-based health care infrastructure enabled complex operations by bringing together structured, accessible, high-quality data. Second, while we were building data sets at Duke to streamline clinical trials and understand cancer care, we were constrained—not by grant funding, but by a lack of access to engineering and capital at scale. Seeing the potential of technical infrastructure, venture capital, and engineering talent (like at Alphabet Google), I realized Flatiron could be an accelerant to the work we were attempting at Duke. The move wasn’t about leaving my academic job but finding a necessary accelerant that was unavailable in the academic space.
Flora: Focusing on COVID-19, the pandemic offered instructive lessons, particularly when information was changing hourly. It forced the FDA and industry to abandon traditional timelines. You were integral to this with the Evidence Accelerator. What surprised you, and what did COVID-19 teach us about the speed we can achieve when necessary?
Abernethy: I learned so much at the FDA, which is a remarkable job serving the public at scale. Within the various government work streams, I focused on how to leverage existing health care data to improve understanding of the pandemic—who was getting sick, the severity, and resource needs like ventilators. We also aimed to use this data to identify emerging problems, such as COVID-related clotting challenges, plan clinical trials, and monitor the performance of medical interventions. We initiated the Evidence Accelerator to gather all stakeholders in the real-world data space—data holders, researchers, tech companies, and government groups—to determine what was possible.
What I learned set the stage for how the FDA views real-world data, informing guidance development for real-world data in clinical trials for 2025. A key realization was how difficult it is for the government to perceive possibilities outside its sphere—like what tech companies or academic researchers find intriguing, or what clinicians observe immediately. The accelerator created a crucible where rapid information exchange allowed regulators to quickly gain familiarity, shortening the cycle time and building trust through intersectionality.
The second lesson was that government regulations and guidance are not developed in a vacuum. Writing guidance requires concrete examples and proof points, which must be supplied by pharma sponsors, academics, and data companies. Finally, the Evidence Accelerator underscored the initial point about building a lingua franca. Different groups use different languages, and moving quickly requires a mechanism to foster shared understanding.
Flora: Moving into 2026, there is significant hype surrounding AI, which has become a turn-off for some. Yet, many of us believe the technology is actually underhyped in terms of the forthcoming changes. Considering the practicing oncologis, dealing with exciting changes like immunotherapies, what AI tools can they actually rely on today? What AI applications do you predict will be mainstream?
Abernethy: The most practical AI applications by the end of 2026 are those currently emerging but not yet fully integrated: AI in radiology, digital pathology, and multimodal biomarkers. These involve leveraging AI to read images, such as radiology scans, with unprecedented detail. AI makes sense for near-term scaling because the algorithms are largely developed and improving.
The challenges in 2025 are primarily rigid structural issues: Accessing radiology images, setting up algorithm flow in PAC systems, transitioning pathology workflows from microscope slides to digital images, addressing payment structures, and accessing and integrating the right multimodal biomarker data. I believe these infrastructure capabilities are gradually coming online. Since the algorithms and regulatory frameworks largely exist, “the kettle is ready to boil” in this space, and we should see major uptake by 2026…
While these tools improve the clinician’s life, they also enhance the patient’s experience. Notes become easier to read, and interpretation arrives faster, adding value for the patient. I believe 2026 will emphasize this benefit.
Flora: Data extraction is pivotal, converting unstructured data into structured clinical notes. This could be a turning point, allowing us to convert clinical notes into true research-grade data at scale. Imagine the potential if previous teams had access to such gigantic datasets for training.
Abernethy: Much of what we document exists as “digital paper,” primarily PDFs, or visual data like ECGs. The use of AI algorithms, including large or small language models, to transform the deep information within unstructured documents into analyzable structured data is rapidly changing and yielding incredible outcomes in the AI data curation space.
Clinicians understand the subtlety inherent in clinical notes. Therefore, algorithms must be trained alongside people who understand clinical nuances, yielding rich, nuanced clinical data. The irony is that AI algorithms require high-quality structured data to be built, and the AI algorithms curating unstructured notes are precisely providing that high-quality structured data, enabling the development of more and better-performing algorithms over time.
Flora: The promise of deep learning is exciting, especially because it is difficult to explain to an engineer why nuanced details—like a daughter’s wedding date or a patient’s three-and-a-half-hour travel time—influence chemotherapy selection. Deep learning is improving in understanding these nuances. Oncologists face a delicate balance: they demand full transparency (avoiding the “black box”) where an AI recommendation is supported by clear inputs, like a patient’s PD-L1 scores or a specific trial. However, they also want the AI to be invisible and seamlessly integrated to reduce click fatigue. Oncologists cannot adopt another tool unless it truly offloads existing burdens. Where do we find that necessary balance?
Abernethy: This focus on transparency versus invisibility is crucial. The reality is that we need both. There must be clarity regarding the algorithm’s purpose, performance, and inputs leading to a decision. Yet, clinicians do not have time for constant cross-checking. If AI is over-reading mammograms, you eventually want to trust the AI reading and allow the radiologists to move to the next task. The algorithms must be trusted “underneath the hood” so they can become invisible yet remain accountable.
The major concern is that algorithms can “go off the rails,” underperform, or hallucinate. To address this, we need performance monitoring for all active algorithms. This monitoring system—a “cockpit”—would ensure that clinicians don’t need to do extra work but are immediately alerted when an algorithm stops performing, allowing them to pause reliance on it until it is fixed. This monitoring is critical for enabling invisible operations over time.
Flora: How do you address the skeptical oncologist who fears AI will only create more work, and how do we demonstrate the “power of the possible”?
Abernethy: It is difficult. One method is encouraging them to test AI in other aspects of their lives—like using it to quickly cross-check a diagnosis or treatment plan for a personal issue. This self-testing helps build familiarity, similar to the goal of the Evidence Accelerator.
Another aspect is not demanding that everyone leverage AI universally. We must identify the specific suite of accessible tools and let individuals determine how to integrate them into their workflow. I believe that even skeptical oncologists will eventually find tools, like ambient documentation, digital pathology tools, or efficient treatment choice algorithms, that improve daily efficiency. When that happens, it will become second nature, like using a calculator instead of an abacus.
Flora: Physicians are discovering ambient note-taking enhances their quality of life, making them more curious and willing to investigate. Let’s look ahead to 2030. What will a medical oncologist’s day look like in 2030? What will they be doing with AI, and what will they do with the time AI frees up?
Abernethy: By 2030, the daily workload of an oncologist—including ambient listening tools and back-office processes like billing and payer reconciliation—should be much more streamlined. A major category of change is the reliance on algorithms and tools that integrate diverse, complex data types to facilitate increasingly complex decisions. These tools will combine traditional scans and pathology images with multi-omics and spatial biology data.
Furthermore, 2030 should bring the first waves of output from AI-based drug discovery. We will see new targets in the clinic with corresponding drugs, and more precise study designs will accelerate their arrival. For patients, care will be optimized daily and made more fluid through the integration of chatbots and real-time data. This integration will feed back to the clinic, not by creating noise, but by providing a signal when attention is needed.
Regarding the extra time AI frees up: I hope every oncologist achieves more rest and resilience. This means a greater focus on the oncologist as a “whole person.” It also provides the necessary mental space to grapple with complex elements of oncology, such as spatial biology, when AI handles many other tasks.
Flora: I hope that by 2030, we have more robust clinical decision support. Patient-facing tools like ePROs and navigation apps need to be more fully integrated. Your point about changes in clinical trials is important. Trials currently take 6–7 years, but we have been “drowning in data and probably lacking in wisdom.” How will clinical trials speed up in this new era?
Abernethy: A key focus at Highlander Health is accelerating clinical research timelines by 50% or more. This acceleration relies partly on AI capabilities: AI for data curation, reduced documentation requirements, better study designs, and improved patient finding and recruitment.
By 2030, we will design clinical trials in new ways. Longitudinal data, possibly super-powered registries, will serve as the foundation upon which clinical trials operate. The trial activities will sit on top of the normal data collection activities in oncology. I believe we will shift toward the concept of nested clinical trials, already emerging in Scandinavia, which will speed up our processes. We will also adopt more modern study designs, like seamless clinical trials that flow naturally from Phase 1 through Phase 3 and into postmarket surveillance. Finally, I foresee 24-h amendments for protocols, ensuring everything moves swiftly.
Highlander Health Partners’ first acquisition was Target RWE, a longitudinal data company focused on liver disease, gastroenterology, and dermatology. We acquired it because we believe longitudinal data sets are the necessary foundation for the new clinical trial research infrastructure. Target RWE was recently rebranded to Pedestal Health to better reflect a shared conviction that the evidence system powering today’s drug development was not built for what modern science demands.
Flora: What specific accomplishment in 5 years would define success for Highlander Health?
Abernethy: By 2030, we aim to showcase hybrid clinical trials. In this model, concurrently flowing data can be analyzed to determine the next study, allowing direct transition into nested clinical trials within the longitudinal data infrastructure. We will measure the study time and demonstrate how innovations have significantly shortened clinical research timelines. Furthermore, we will have invested in the flow-through mechanisms, ensuring that learning from clinical trials automatically updates clinical decision support, guidance committees, guidelines, and patient-facing applications.
Flora: What if it all goes right? How do you think this story ends 5–10 years from now, providing reassurance that we are responsibly introducing these technologies?
Abernethy: If it all goes right, oncology is going to showcase the way. Oncologists will demonstrate how to focus on biology, discover new targets, and develop treatments even for the most difficult diseases. We will learn how to use those treatments in the clinic to personalize care for patients. This will be coupled with making daily life easier for oncologists and cancer patients. Most importantly, by 2035, the advances made in oncology will translate to other fields, such as neurodegenerative and rare diseases. We will apply the lessons learned treating glioblastoma to conditions like Alzheimer’s disease. That is what the story will look like in the next 10 years.
