Navigating the Intersection of AI, Clinical Pathways, and Value-Based Cancer Care

The quest to deliver optimal cancer care to the right patient at the right time is a complex, ever-evolving challenge, one that demands constant innovation. Today, that innovation is being driven by advanced therapeutic science, value-based care, and the potential of new technologies.

A panel at the 2025 Patient-Centered Oncology Care meeting brought together leading experts to dissect this evolution, exploring how to operationalize pathways in the busy clinic, manage financial toxicity, integrate vast quantities of precision medicine data, and cautiously yet optimistically embrace artificial intelligence (AI). Clinical pathways, through the use of technology, are evolving from static documents to a tool that delivers the right information to the provider at the right time, explained panel moderator Edward “Ted” Arrowsmith, MD, MPH, medical oncologist and medical director for pathways at OneOncology.

“I would say that the intellectual work of producing a PDF that lists treatments for stage and molecular features is just the beginning of the work that we try to do to get the right care for the right patient at the right time in the clinic every day,” he said.

The New Frontier of Pathways: Integrating Technology and Precision

AI has become so pervasive that Arrowsmith quipped it feels like there is a law that “all panels have to start that way.” In oncology, the large amount of data available from genomics, imaging, and remote monitoring is a natural fit for AI’s analytical capabilities.

David Jackman, MD, medical oncologist and medical director for pathways and strategic alliances at Dana-Farber Cancer Institute in Boston, Massachusetts, highlighted that AI is not a future concept but a present reality in oncology operations.

“We have AI listening to our visits, summarizing notes. We have AI getting discrete data elements out of EMRs [electronic medical records], or ingesting notes from outside and creating summaries,” Jackman said. “Already, we’re using AI to figure out what the clinical setting is. We’re using it.”

The critical challenge, he said, lies in the core decision-making and deciding how much to allow AI to do while still keeping humans in the loop.

The scale of AI implementation at major cancer centers provides a window into its transformative potential and the necessity for structured governance. Fernanda Polubriaginof, MD, PhD, associate chief health informatics officer at Memorial Sloan Kettering Cancer Center (MSK) in New York, New York, detailed a centralized governance structure for AI at her institution, which includes an AI registry with 83 different models.

MSK takes a risk stratification approach when rolling out AI models, where high-risk tools are monitored closely and often to ensure safety as new technologies are deployed. Low-risk tools are still monitored but less frequently.

There are some use cases that are being explored but aren’t ready just yet. “Some of these are just turned on behind the scenes so that we can collect data and really evaluate if this is a use case that we can then later expand,” Polubriaginof explained.

Operationalizing Value and Minimizing Financial Toxicity

A core driver for the evolution of pathways has been the spiraling cost of cancer treatment and the mandate for value-based care. When The US Oncology Network’s pathways programs began 20 years ago, it had a straightforward goal of using generic drugs to save money, explained Aimee Ginsburg Chesnick, PharmD, BCPS, director of clinical content strategy at McKesson Specialty Health, which supports network practices. This strategy worked for years, but it had to change as therapies began to get more expensive.

“We had to standardize the approach,” Chesnick explained. This standardization involved evaluating treatments for equivalent efficacy and toxicity and then examining cost, but first they had to grapple with the cost for each stakeholder group: patient, payer, provider, and institution.

The inability of the US to draw a line in the sand on cost-effectiveness meant the Network had to take a more nuanced approach that integrated incremental cost-effectiveness, total cost of care, and advanced payment models into clinical guidance.

This focus on financial reality is moving into the realm of supportive care, where pharmacists are pioneering real-time decision support, said Candy Peskey, PharmD, BCPS, BCOP, assistant professor of pharmacy at Mayo Clinic in Rochester, Minnesota. Her organization integrated a decision support tool into the EMR that addresses financial toxicity at the moment of prescribing.

For instance, a medical oncologist may have a patient arrive in the clinic with a rash after receiving immunotherapy. If the oncologist attempts to order their preferred topical steroid cream, the EMR will incorporate payer data and inform the oncologist whether the patient’s insurance covers the cream and what the co-pay will be. The oncologist will be given some alternatives that may be more cost-effective for the patient.

At some point, Peskey said, the goal of this tool will be to “incorporate [these] payer data into oral oncolytics that maybe need a prior authorization, but maybe we can shave a few days off by routing it to the correct specialty pharmacy right out of the gate, rather than delaying therapy by sending it to one pharmacy and then forwarding it on to another pharmacy.”

Overcoming Data Fragmentation and Decision-Making Boundaries

In oncology, there are vast amounts of data, but the use of those data is still very simplistic, Jackman said. He envisions a time in the near future in which AI can analyze genomics and consider appropriate therapeutics in a more nuanced way, but this will provide challenges in planning pathways.

Polubriaginof likened the shift to a greater use of technology, such as wearables, to the transition from mostly inpatient to more outpatient treatments. MSK has built up programs to administer drugs and send patients home with devices to monitor them 24/7 and catch it early if they start to develop something like cytokine release syndrome and need to go back to the hospital.

Having data from a variety of sources still presents a challenge, Chesnick explained. A patient might receive a diagnosis at one institution, then a second opinion at another, but they are the only one with all the information gathered. To get to a day when AI can help make treatment decisions, all of those structured and unstructured data, plus the historical data on the patient, need to be collected in one place.

“I think too often, what we have is a scan that somebody has…interpreted inappropriately, and then we’re trying to make decisions based on inaccurate data,” she said. “And that’s the part I think that needs to be [standardized and stabilized] before we move forward with AI.”

A prime example of using AI to conquer data chaos is in clinical trial matching, which Chesnick called a “no-brainer” because it involves searching for specific, quantifiable eligibility criteria. Peskey noted that Mayo Clinic is beta testing the use of AI for certain tumor groups to search through patients’ charts and pair them with studies where they meet eligibility criteria.

However, there remains a philosophical boundary between augmenting a process and making a definitive clinical recommendation. There is a difference, Jackman said, between using AI for trial identification or for suggestions and using it for specific directives. He argued that a more comfortable next step is offering a model where AI presents data on how similar patients did with different treatment options, allowing the physician to make the final decision. Then, if a pathway has 3 options, the AI can provide information about how patients with a similar genomic subset of a similar age and race have done on these treatment options. The key, though, is having the tool be transparent about where it got the data from, he said.

“Maybe it is informative, maybe it’s not, but I think that’s an easier next step vs the black box of ‘We think you should do this, and we don’t have a transparent way of telling you why,’” Jackman said.

Regulatory Hurdles and the Challenge of Trust

The rapid adoption of AI introduces significant barriers concerning liability, regulation, and trust. Polubriaginof emphasized the necessity of providing a safe environment for clinicians to explore AI. MSK has created a Health Insurance Portability and Accountability Act–compliant AI portal so clinicians can have a place to explore new ideas instead of just going “to their ChatGPT account.” She drew a historical parallel, noting that during the transition from paper records to electronic health records, the development of clinical decision support took decades. However, the adoption of ChatGPT took only months.

Looking Ahead: Augmentation, Not Replacement

The panelists agreed that while the field is evolving at an unprecedented pace, the immediate future of AI in oncology will be one of augmentation rather than replacement.

“[AI] will never replace the doctors, in my opinion, but it will augment and help support with some of these decisions, especially when you have…such a complex patient with data from everywhere,” Polubriaginof said. “Maybe it can help us make sure that we’re not missing some of those test results that are in some PDF scanned somewhere in the chart.”