Cancer AI’s next frontier: Turning better models into improved care

This article is sponsored by the SC26 Conference.

With the recent fanfare surrounding AI and the prevailing reliance on technology, it is no surprise that major initiatives are underway to develop AI models that support imaging, diagnostics, analysis, and treatment planning for all types of cancer. However, cancer AI models have accelerated beyond merely finding patterns or examining data. Today’s mission is to move cancer research closer to the people it is meant to serve: the patients. That said, a cancer AI model can perform well in research yet still be far from ready for implementation in patient care. 

A shift in career, a shift in data

Closing the gap between a promising AI result and a clinically tested, useful cancer AI model is a primary focus for Eric Stahlberg, executive administrative director of the Institute for Data Science in Oncology (IDSO) at The University of Texas MD Anderson Cancer Center. In December 2024, after years in national-scale cancer computing, Stahlberg moved inside a hospital for the first time when he joined UT MD Anderson Cancer Center in Houston. The move was part of his personal aim to “shorten and accelerate the path from important cancer research insights to actual patient impact.” 

Stahlberg has a unique perspective. Over the last three decades, he has been deeply engaged in data science and high performance computing (HPC), including directing cancer data science initiatives at the Frederick National Laboratory for Cancer Research, where he helped to establish and lead the collaboration between the National Cancer Institute (NCI) and the U.S. Department of Energy (DOE) in applying supercomputing to oncology at scale. With experience spanning cancer data science, HPC, AI in research, Stahlberg chose to spend the next chapter of his journey seeking solutions for patients. 

“We’ve done a great deal of development in AI applications in a research setting,” he said. “Now, the job is to figure out how we can leverage those capabilities to reach the patient.” 

Stahlberg’s move to IDSO also changed his approach to the work. In a research setting, the reward often is promising new data science approaches, AI tools and computing capabilities to enhance the next experiment. Yet in a hospital, new capabilities can lead to complicated questions. How do these tools improve the patient experience? Do they augment current operations or support clinical decision-making? Do the capabilities meet the considerable safety concerns for patient care? A higher standard emerges. 

Through the years, Stahlberg has observed that the greatest change is in how cancer research uses data, explaining that advanced scientific computing in cancer research had focused on reduction coupled with managing the growing abundance of data. To home in on an interpretable result, researchers had to filter, sift, and narrow the data, focusing down to only what mattered. Now, Stahlberg explained, two general classes of AI have added a different methodology to the mix: 1) predictive AI, in which models learn from data to provide answers, and 2) generative AI, which may also create or suggest new answers in response to inputs (ChatGPT being a popular example). 

“The higher volume and diversity of data that is now available adds both complexity but also significant opportunity to train, evaluate, and apply new AI models for cancer,” he said.   

For a cancer data science leader, that distinction is useful. Where the bottleneck once was primarily raw compute, data access, quality and its governance are now key hurdles. Quality, provenance, interoperability, and the discipline in managing data become paramount to maintaining trust. As AI becomes more capable, the bar is raised for disciplined data operations. 

Federal research provided important groundwork for this shift in cancer. For example, the CANcer Distributed Learning Environment (CANDLE), a software framework co-developed by national laboratories including DOE’s Argonne, Lawrence Livermore, Oak Ridge, and Los Alamos, Frederick National Laboratory for Cancer Research, and NCI piloted a new capability that helped expedite creating breakthrough machine learning models explicitly for cancer research. Efforts like this drove pilot projects at the cellular, molecular, and population levels and helped make large-scale computation usable for cancer questions. For Stahlberg, who spearheaded Frederick National Laboratory’s contributions to CANDLE, that groundwork changed what is now possible.  

“AI is providing an avenue to bridge the complexity of cancer and of cancer data, enabling so many new, creative use cases to help shorten the time from research insight to clinical impact,” he added. 

From model to medicine

In a hospital, physicians see real patients, many with significant health issues and comorbidities. The test becomes whether the emerging AI and computational capabilities actually support better patient experiences and outcomes. Does data crunching and modeling improve clinical decision-making and care obligations? Often, the challenge is knowing where to start. 

“Given the number of potential applications, there are multiple potential avenues to move a given research capability into clinical or operational use,” Stahlberg said. “It is not a surprise that there is no one-size-fits-all path.” 

Stahlberg’s time at UT MD Anderson has sharpened his sense of what is keenly needed for effective translation, and some recognizable progressions have emerged from his experiences. 

“An important starting point is for the new capability to deliver recognized value, which is best achieved by working with clinicians to address a recognized challenge or opportunity for improvement,” he explained. “It is a team approach to cancer data science.” 

As one approach, the team works together to create and evaluate the AI model for its value to the process, confirm its soundness, and validate it against broader datasets to show whether it meets the rigorous conditions required for potential future clinical use. Assuming the first bar is met, the team introduces the AI capability for initial use in an observational clinical or educational context. As regulatory and operational requirements are met, the cancer AI model would then progress to prospective use in a clinical trial, and broad deployment would follow only thereafter. 

Stahlberg referenced something that has guided him for years: “If you can observe something, you can model it. It is then improving the model that becomes the focus.” The goal is to take the imperfect models, identify their value and their limits, and then work to improve them with new insights, information, and approaches, driving the field forward. 

Finding community to solve complex problems

According to the American Cancer Society, roughly two million people are diagnosed with cancer in the United States each year. Each diagnosis represents a unique individual with distinct needs, concerns, and potential treatments.  Due to such complexity, cancer was once believed to be too complicated to be organized for crunching on a supercomputer. 

“The common perception was that there were too many factors to realistically bring together into something that you could compute,” Stahlberg said. Instead, he saw an opportunity. Cancer is a problem that requires many fields to work together at once, making computation important beyond its raw horsepower. 

“Advanced scientific computing is a way to bring things together,” he said. “People can work in isolation on separate pieces, but the computation brings their work together in the same place, at the same time, to interact and create dynamic models that tackle complexity.”   

“HPC actually helps you unite people working across disciplines and domains in a network,” Stahlberg added. The shared ground of advanced computing gives clinicians, data scientists, computational scientists, and institutions a common framework for developing a shared language to work together and solve meaningful problems.  

“I have been fortunate to work with wonderful leaders in computing, science, and education across many organizations throughout my career,” Stahlberg commented. “IDSO provides the opportunity to continue this course with the important focus on improving outcomes for cancer patients.” 

For more than a decade, Stahlberg has fostered the growth of this cross-disciplinary computing community via the Computational Approaches for Cancer Workshop (CAFCW), held in conjunction with Supercomputing. Stahlberg co-founded CAFCW in 2015 with Patricia Kovatch, presently System Chief, AI Computing and Data at the Icahn School of Medicine at Mount Sinai in New York. Over the years, more than 1,000 people have participated in the workshop, building an HPC community devoted to accelerating cancer research to improve patient outcomes. The 12th workshop will be held on November 16th, as part of SC26 in Chicago. Its 2026 theme is “HPC Unites Patient, Clinic, and Research to Improve Cancer Care.” 

This workshop represents an opportunity for those at the forefront of cancer and AI to find common cause among like-minded colleagues working toward similar goals. The CAFCW workshop offers thought leaders in cancer data science and scientific computing opportunities to share innovative work and best practices in navigating AI adoption, data infrastructure, and the long road from model to clinical use. CAFCW helps to connect the advanced computing community working to push the next frontiers in cancer care. 

SC26 takes place from November 15 to 20, 2026, at McCormick Place in Chicago. For those working at this intersection, the CAFCW workshop is where the conversation continues. There is a student track as well as a professional track. The best student submission is competitively selected and will present at the workshop. To share your work at the CAFCW workshop, please submit an abstract for presentation or papers for the SC Proceedings at https://cafcw.org.   

SC26 Registration is now open.