During the National Comprehensive Cancer Network (NCCN) 2026 Annual Conference, Travis Osterman, DO, MS, FAMIA, FASCO, offered an overview of how AI can be utilized to improve oncology care and provided tips for how physicians should approach the adoption of AI tools in their own practice.

Dr. Osterman, who is Associate Professor, Biomedical Informatics; Associate Professor, Hematology and Oncology; Associate Vice President, Research Informatics, Vanderbilt University Medical Center; and Director, Cancer Clinical Informatics, Vanderbilt-Ingram Cancer Center, first described the overall field of AI and how it overlaps with the pre-existing data science and analytics field and natural language processing. He explained that it also encompasses areas of machine learning, deep learning, and the large language models or chatbots, like ChatGPT, that serve as a widely used example of AI.

He then explained that AI is already being used and explored for future uses in many applications across the entire care continuum from prevention, diagnosis, and treatment to survivorship and end-of-life care.

Common AI Implementations in Oncology Practice

Dr. Osterman gave examples of a number of AI uses that have already been implemented in many clinical practices, including at Vanderbilt. He also noted a number of potential places where technology could be used to improve patient care.

Ambient Scribes

Ambient scribes record the audio of a clinic visit, attribute each voice to the correct person, and format the recording as a usable note. These scribe services have been integrated into many health-care systems across the country already. However, he noted that oncology tends to lag behind other areas of medicine in adopting ambient scribes.

Chart Summary

Chart summary is another area where AI has been promised to simplify administrative burdens to simplify the reading and understanding of all the elements of a patient case. Several solutions to this issue have been proposed and have been integrated into some health-care systems already. However, Dr. Osterman noted, there are still challenges with AI chart summary, especially as physicians are still responsible for the result, even if AI makes errors or something critical is missed.

Radiology Critical Alerts

Many institutions have applied AI for initial reads of radiology scans to flag critical findings, which are then reviewed more quickly by radiologists.

“What I don’t want you to hear is ‘we’re using AI to replace radiologists’ jobs at Vanderbilt,’ because that’s not what’s happening. We’re using AI to help make sure that the radiologists read the critical studies faster. And by reading those studies faster it hopefully means that I get the notice more quickly and then I can notify my patient while they’re still in clinic,” Dr. Osterman explained.

Surgical Planning

AI and augmented reality are also being used to guide re-resection and simplify the relocation of positive margins during surgery. Surgical specimens are scanned ex vivo and the 3D file is uploaded to an augmented reality headset to show the surgeons where tumor still exists. A hologram of the 3D specimen is then aligned with the surgical defect to guide re-resection.

A feasibility study was performed in head and neck cancer showing that of 20 resections completed with augmented reality, the mean relocation error was 4 mm with a standard deviation of 3.9 mm. The average time of the overall protocol from 3D scanning to alignment of the hologram into the surgical bed was 25.3 minutes (± 8.9 minutes).

“I think this is a fantastic opportunity to leverage artificial intelligence and current technology to provide better care for all of our patients,” Dr. Osterman said.

Infusion Scheduling

Dr. Osterman explained that at Vanderbilt they have worked with outside vendors to build an infusion schedule that optimizes planning chemotherapy infusions for each patient in line with their chemotherapy cycles, the rate of each patient’s infusion, nursing staff availability, infusion chair availability, and more. He explained that this has resulted in narrowed estimated wait times for patients and improved efficiency of scheduling for both patients and nurses. Dr. Osterman noted that this has also improved the culture among nursing staff because shifts can be better planned in advance to accommodate busier times.

“This has given me a new lens when we’re implementing AI solutions to not just think about the technical wow, or not just the one infusion chair added, but to think about how this impacts our patients…and how this impacts our workflow,” Dr. Osterman said.

This has already been implemented in laboratory hematology, he noted. AI-based hematological analyzers automatically count cells and platelets, and Dr. Osterman said that many physicians take it for granted that systems in their practice can count cells for them.

Regulatory Updates

FDA

One area that has been a challenge for AI algorithms used as medical devices is the possibility of algorithmic drift. Dr. Osterman explained that there is concern over compliance risk as well as physician concern over the deterioration of the performance of the AI device or solution. However, commercially, vendors do not want to pay to re-approve each updated version of the model to account for and overcome algorithmic drift and physicians do not want to commit to a device or solution that is expected to get worse over time and is not updated and re-approved.

To manage this issue, the U.S. Food and Drug Administration (FDA) implemented the predetermined change control plan (PCCP) pathway in August 2025, whereby a company can submit a plan for how they will update their algorithm within a set of guardrails without needing additional FDA approval. The PCCP focuses on areas of modification description, modification protocol, and impact assessment.

Dr. Osterman recommended asking all vendors about their PCCP before implementing a new solution so that the change control plans can be implemented into the system early on and changed over time.

ONC

Vendors for electronic health record systems operate in accordance with the Office of the National Coordinator (ONC) for Health Information Technology, which has split areas of decision support intervention into two categories: evidence-based and predictive decision support interventions. The first category addresses rule- and guideline-based interventions, and the second category includes areas of machine learning, natural language processing, large language models, and more. “These are what we think about as more of the black boxes,” Dr. Osterman commented.

The ONC requires that predictive decision support interventions expose their source attributes in plain language at the point of use—either through direct display, hyperlink, drill-down, etc.—so that physicians can easily understand why the system is making a certain decision for a patient, for example. Dr. Osterman provided the example of a sepsis analysis tool to illustrate how the system needs to alert the clinician that the patients is at risk of sepsis and provide clear and easily accessible reasons for why it has come to that conclusion.

“If you’re getting Black Box Warnings, you should go back to your IT teams and ask them whether or not this is on their road map to implement. Chances are, it probably is,” Dr. Osterman said.

CMS

The Centers for Medicare & Medicaid Services (CMS) has issued the CMS Interoperability and Prior Authorization Final Rule to make healthcare systems and payers provide application programming interfaces to help expedite prior authorizations. Dr. Osterman expects that it will take at least another year for the rule to be fully implemented, but expects that it will improve areas that delay prior authorizations such as when a patient changes insurance companies.

“This really opens the door for patient-facing applications where a patient can authorize their health system to give them the information and tell them where they are in the prior authorization workflow,” Dr. Osterman commented. “We may be years away from seeing the applications actually come to fruition, but the first step is forcing the vendors to make [these] data available to the patients. So I think this is a fantastic place that we are laying some new ground.”

Challenges With AI Adoption

Although AI has and can provide many helpful solutions and improvements in oncology, there are still ongoing areas of concern and challenge for the implementation of AI in health care.

Reliance on AI

Research has increasingly demonstrated over the past few months that a reliance on AI can lead to “deskilling” or the reduction of specialized expertise needed for a job or function due to the use of technology. Outside of medicine, researchers showed that when students used a large language model—ChatGPT in this case—to help them write an essay or complete a writing task, many of those students could not accurately retain the written information compared with students who did not use a large language model. “The results are not shocking,” Dr. Osterman quipped.

In medicine, on the other hand, a study showed that endoscopists who routinely used AI support in their analysis of colonoscopies had a reduced adenoma detection rate when the AI assistance was removed. The detection rate decreased from 28.4% before AI exposure to 22.4% after (P = .0089).

“This is interesting because this isn't what we've seen [in areas where we’ve used] other kinds of clinical decision support in health records. And I think this is interesting for us to think about as we consider how we implement AI tools,” Dr. Osterman said. He explained that in these situations, the AI system is making decisions before the physician has had a chance to address it without assistance, and thus may not learn from the interaction.

“Does this mean that I think we should not use AI systems? Certainly not,” he said. “But I do think this is something we’ll have to consider.”

Medical Student Paradox

Dr. Osterman explained that he has spoken with many AI companies that have pitched a decision support intervention in the form of an excellent fourth year medical student. However, he believes that this is not the best end-goal for an AI system as it puts greater burden on the physician to review all of the AI’s recommendations and decisions as the risk of malpractice is still on the physician and his own license.

“AI systems do best at the middle of the bell curve,” Dr. Osterman explained, meaning that they perform best in areas that are routinely seen in practice because there are more examples for the system to learn from. However, he explained, these are not the areas most in need of assistance. Instead, he noted, physicians need more help with the more unknown areas, where there are fewer examples to give the system.

Future AI Applications in Oncology

Dr. Osterman gave a look into the future for where AI could be heading in the oncology space. AI licensing, or allowing AI to practice medicine at any level, has been typically discussed as one possibility for the future of AI in health care in terms of medication refills, viral workups, and administrative tasks such as filling out back to school/work letters. Instead, Dr. Osterman proposed something more provocative: that in the future, physicians will supervise AI agents that are capable of semi-independent medical practice.

Already, several large language models have demonstrated the ability to pass the U.S. Medical Licensing Examination and specialty board certifications. Additionally, there is an expected shortage of physicians in the United States that cannot keep up with the growing demand of a growing population. The Association of American Medical Colleges has projected the shortage to reach up to 86,000 physicians by 2036.

Over the past several decades, physicians have evolved from solo practitioners, supported only by nurses, to a leader of a multidisciplinary team of nurses, clinical pharmacists, nurse practitioners/physician assistants, etc. Dr. Osterman expects that in the future, that list will extend to include AI agents.

He explained that it will be one of many choices going forward to determine how much responsibility physicians are willing to give AI. “I think one of the choices we have as thought leaders in our field is how do we want to think about adopting that or not, and how do we want to help our regulators think about what that has to look like for the cancer practice. Because I would argue again that [oncology practices are] a great practice model to think deeply about this because of how interdisciplinary our practices are,” Dr. Osterman commented.

Practical Tips for Leveraging AI

Dr. Osterman concluded with a few practical tips for how to leverage AI in oncology practices. He noted first and foremost that not every vendor product is needed. Also, many AI products or solutions will introduce many new questions, which would then only add to the cognitive burden of physicians.

He noted that learning how to use AI is not dependent upon a younger, more technologically inclined generation. Instead, the skills of AI—including prompting strategies, and how to identify when an output is wrong even though it is confidently declaring its decision—need to be learned at all ages and levels.

AI agents can be used to help in many areas of oncology, but Dr. Osterman stressed that physicians can use and work with agents—they will not replace physicians. “We’re already facing the physician shortage. None of this makes me think that any of us—radiologist, pathologist, infectious disease, oncologist—are less likely to have a job going forward. But I do think that our job will continue to change, just like it’s changed over the last several decades,” he commented, concluding that he expects these changes to occur more slowly than many believe.

DISCLOSURES: Dr. Osterman has disclosed receiving grant/research support from GE HealthCare and Microsoft; receiving consulting fees from AstraZeneca Pharmaceuticals LP, Biodesix, COTA Healthcare, eHealth Technologies, Flagship Biosciences, Inc., GenomOncology LLC, MDoutlook, LLC, and Outcomes Insights, Inc.; and serving as a scientific advisor for AstraZeneca Pharmaceuticals LP, Biodesix, COTA Healthcare, eHealth Technologies, Flagship Biosciences, Inc., GenomOncology LLC, MDoutlook, LLC, Outcomes Insights, Inc., and Tempus AI, Inc.

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.