How Federal Regulatory Strategies Are Spurring Innovation in the Use of AI in Cancer Care
Last August, the United States Food and Drug Administration (FDA) issued a guidance to provide recommendations for predetermined change control plans (PCCPs) tailored to AI-enabled devices. The new guidance offers guardrails for AI models while maintaining a reasonable assurance of their safety and effectiveness, and addresses modifications that would normally require a new marketing submission, allowing manufacturers to control “algorithm drift” on their devices within the predefined scope without the need for additional FDA reviews for each new iteration of the AI model.
The impact of this new guidance on oncology care can be far reaching, enabling faster access to more precise AI-enabled devices that can keep pace with the rapid acceleration of advances in cancer care.
“The issuance by the FDA of this PCCPs guidance is a big step forward in helping AI get into the oncology clinic,” said Travis J. Osterman, DO, MS, FAMIA, FASCO, Associate Vice President for Research Informatics at Vanderbilt University Medical Center and Director of Clinical Informatics at Vanderbilt-Ingram Cancer Centers.
In this interview with ASCO AI in Oncology, Dr. Osterman discussed how federal regulation is speeding AI use in the clinic, using augmented reality to assist in the treatment of cancer, and what’s on the horizon for the utilization of AI in cancer care.
Please talk about the ramifications of the recent FDA guidance on PCCPs on AI-enabled devices in oncology care, and the importance of maintaining quality care and patient safety in all regulatory decisions.
We have seen quite a bit of movement recently from the federal government across several agencies. The new guidance by the FDA on PCCPs is a welcome addition to federal regulation that we have been seeking for a long time, and fills a gap that oncology has had in the AI space.
The guidance outlines anticipated changes to a medical device that are expected to occur after the device receives FDA approval, and is especially relevant for software-based devices, in which updates and improvements can be used to add features. For example, if manufacturers are trying to add software to a device that tracks whether a patient’s tumor has grown, they had to wrestle with the lack of guidance on whether every update required going through another approval process to have the device renewed, which was a really big barrier to innovation.
This guidance gives vendors a greater sense that their devices will perform in the market a year later the same way they performed on the day they were approved. It also assures medical institutions that vendors are committed to making sure their AI models are updated over time, so it’s a win-win for manufacturers and for patients.
Your institution developed a head-mounted augmented reality system surgeons can use to guide tumor resection and communicate faster and more accurately with pathologists during surgery for patients with head and neck cancer. What have been the results so far; has the device resulted in fewer surgical margin resections to confirm clear margins at the edge of the removed tissue?
This is incredible work being done by Michael Topf, MD, MSCI, Associate Professor of Otolaryngology-Head and Neck Surgery at Vanderbilt University Medical Center, who has been a visionary in this field. The use of this technology is still being investigated in a clinical trial, so we don’t yet have published data on its performance, but safety and effectiveness are at the core of this research.
The idea stemmed from the lack of application for 3D scanning in oncologic surgery. Dr. Topf implemented a protocol to create 3D models of resected cancers for surgeons, pathologists, and oncologists. In this study, Dr. Topf is using augmented glasses that create a holographic overlay that highlights the precise region that needs further tissue removal in patients with head and neck cancer. In the head and neck region, every centimeter of tissue that is removed is very valuable. The goal is to remove all the tumor and make sure there are negative margins, and not extract more tissue than is necessary, and reduce the risk of cancer recurrence.
This technology is transferable to applications in other cancer types and will become more commonplace with the results from this research.
AI is being integrated into treatment decision-making by providing clinical decision support and into research. What FDA regulations are needed to utilize these types of AI software and devices in the clinic and research; and what guardrails should there be put in place to alert humans when AI isn’t equipped to make accurate treatment predictions, and to ensure that the model isn’t biased?
While it’s very important that our AI models are unbiased, we also need to ensure that we preserve access to all patients. Rural communities may benefit most from these models. Traveling to large cancer centers for treatment adds a huge burden on patients that AI could reduce.
— Travis J. Osterman, DO, MS, FAMIA, FASCO
These are all questions we are wrestling with today. I see the perspective of needing regulation on AI medical devices, and the PCCPs guidance gives us confidence that AI models will be updated and improved over time.
For some of these other applications—for example, in clinical decision support—there is push-and-pull on both sides of the regulation argument. In some instances, we need to get these devices into the clinic to understand their utilization and accuracy. And I worry about overburdening clinicians and researchers with too much FDA regulation, especially on the implementation side because there are so many patients with cancer that would benefit from this technology.
Whether those systems need more or less regulation still needs to be worked out. But I want to be very careful about stifling innovation because the very patients we are trying to protect, especially those living in rural areas, where we need to ensure the delivery of high-quality oncologic care, are going to pay the price.
While it’s very important that our AI models are unbiased, we also need to ensure that we preserve access to all patients. Rural communities may benefit most from these models. Traveling to large cancer centers for treatment adds a huge burden on patients that AI could reduce.
You were the keynote speaker at the National Comprehensive Cancer Network (NCCN) 2025 Oncology Policy Summit, which explored the evolving landscape of AI and its role in cancer care. During your presentation, you acknowledged a widening gap between the use of AI in National Cancer Institute–designated cancer centers compared with community cancer centers. How can this gap between the two types of cancer centers be closed?
We have already started to close the gap in some areas. For example, with the use of ambient scribe technology, which is now experiencing rapid, widespread adoption across both large comprehensive cancer centers and smaller community oncology centers and practices. Oncologists feel that the technology saves them time and alleviates high documentation burdens, a major contributor to clinician burnout. We can argue about whether the technology actually does save time, early studies have shown mixed results.
My hunch is the greater value is in having more face-to-face time with patients, which is what we all want. We came to the field of oncology to spend more time with patients, not to spend time typing notes. So, this technology has been a great advancement in patient care.
Another key area of innovation with AI models that can be used in both large and small cancer centers is in optimizing AI-powered infusion scheduling. Optimizing infusion room scheduling is one of those economy-of-scale advantages that might benefit very large 100-chair infusion centers more than a smaller four- or five-chair infusion center, because the technology can manage highly complex workflows, optimizing hundreds of daily appointments, increasing chair utilization by 10% to 20%, and reducing patient wait times. However, small- and medium-size centers that operate with tighter margins and less staff may also benefit from the technology because they are able to increase patient volume, improve throughput, and enhance staff satisfaction without having to invest in physical expansion.
The use of AI-powered radiology critical alerts is another innovation that benefits both large medical cancer centers and smaller community centers and plays a key role in compensating for the high demand of radiologists, and is widely adopted across the field.
As more technology companies standardize these AI-powered tools and reduce the implementation burden on clinicians, we’ll see more of them make their way into community practices.
What would you like to see AI accomplish in cancer care over the next 3 to 5 years?
A decrease in disease burden for both patients and clinicians. For example, AI-powered patient-facing portals that allow patients to be matched to clinical trials, and AI-powered systems that enable the same level of care at both large cancer centers and smaller community centers, so patients don’t have to travel long distances to receive their care.
On the clinician side, these tools can help standardize care and assist physicians in being aligned with NCCN and ASCO guidelines to support best practices. They will also help monitor patients for undue toxicity risks that might be different from those seen in patient populations in clinical trials. These are real cognitive burdens busy clinicians wrestle with. We all welcome AI to help reduce clinical burden, especially the administrative burden of caring for patients.
AI-powered tools that can assist in treatment decision-making is a longer time frame away. I understand why we want to focus on this aspect of patient care, but it’s the hardest to accomplish for a couple of reasons: One, is these models require training on vast, diverse, and high-quality data sets to provide precise, personalized recommendations, and must be validated in prospective, randomized clinical trials, which take a long time to complete. Second, the area of care where we need the most help is with edge cases—those patients with rare anatomical variations, atypical tumor behaviors, or unusual comorbidities that challenge standard treatment protocols. By definition, those are the patients we have the least amount of data on to train AI models.
Structurally, using AI-powered tools to help in treatment decision-making is a great goal, and we are making progress in that direction, but I don’t expect that these tools will be fully tested and implemented over the next 3 to 5 years.
Identifying patients for clinical trial matching and identifying what trials a large center should open to meet the needs of its patients are the fertile areas of implementation of the next 2 to 3 years. Looking out a little further to the 3-to-5-year horizon, we’ll see AI-powered diagnostic tools transition from experimental research to operational, clinical-grade systems, with significant clinical adoption. A little further out from there, we’ll get to the adoption of treatment decision-making support systems.
DISCLOSURES: Dr. Osterman receives research funding and has partnerships with GE Healthcare; Microsoft; Melinda Stead Fund; IBM; Conquer Cancer, the ASCO Foundation; the National Cancer Institute. He is an advisor to AstraZeneca, eHealth Technologies, MD Outlook, Biodesix, Medscape, Dedham Group, MJH Life Sciences, Tempus AI, COTA Healthcare, Flagship Biosciences, Inc., GenomOncology LLC, and Outcomes Insights, Inc. Dr. Osterman also has ownership in FacultyCoaching.com
ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.