Experts predict rising use, acceptance of AI in cancer treatment

SAN FRANCISCO — Leading oncologists said this week that artificial intelligence will one day be as integrated into cancer care as it is in smartphones and self-driving cars — and that this is a change we should welcome.

Their comments, made at STAT’s Breakthrough Summit West on Wednesday, reflected an optimistic view for how the health care system can use AI across nearly all aspects of cancer care, from matching patients with clinical trials to predicting how they might fare on a given treatment. Some of this work is already happening. The panelists noted that AI has the potential to offer deep expertise across a growing number of precisely defined cancer indications, and that the technology can generate insights research focused on individual hypotheses might miss.

Whether these tools will be widely accepted by patients and health care providers, however, is a different question. The panelists stressed that physicians need better ways to understand the strength of AI-powered findings, adding that these tools can aid human judgement but will never replace it.

Despite those caveats, Karen Knudsen, CEO of the Parker Institute for Cancer Immunotherapy, stressed that AI’s widespread integration into clinical practice is both a net positive and unavoidable.

“I think we have massive opportunity in cancer care to get patients to the right care, the most advanced care, earlier by taking those workforce shortages and using AI to get to solutions,” she said. “I’m completely bullish on AI. It’s inevitable.”

In some ways, the future is already here. Clinicians increasingly use AI-powered notetakers to free them to engage with patients during appointments, said Knudsen. And Danielle Bitterman, clinical lead for data science at Mass General Brigham, noted that she used AI to help her treat a patient just two days before the panel discussion. Bitterman is a radiation oncologist, and clinicians in her field use AI to develop treatment plans that optimize efficacy and safety.

But the panelists framed these advancements as small splashes compared to the tidal wave of adoption on the horizon. Clifford Hudis, CEO of the American Society of Clinical Oncology, pointed to the iPhone as a useful analogy, adding that most new smartphones have AI prompts when you open your email or search the web — regardless of whether you use or ignore those tools.

“I think that’s the ubiquitous way in which this is going to seep its way into every minute of the clinician’s day,” he said. “It’s going to make our lives better and our patients’ health better.”

As an example, Hudis said that AI could help “nudge” busy clinicians with patient-specific reminders and notes during appointments: “That’s a use case that we should be begging for.” The ASCO head, who for decades treated breast cancer patients, noted that researchers can now define cancer types far more precisely than was once possible due to advances in genomics. That comes with the ability to better target treatments to patients, but it’s also becoming harder for clinicians to have expertise across a growing number of narrowly defined indications. That’s another area where he believes AI can help.

Knudsen argued that, while many people are worried AI will replace jobs, these tools have the potential to fill critical workforce gaps. In her previous role as head of the American Cancer Society, the organization invested in Paradigm Health, a firm that uses AI-powered tools to help clinicians plan clinical trials. Knudsen, who was also previously executive vice president of oncology at Jefferson Health, added that AI can also help better match patients to the right studies.

“I would have killed as the head of Jefferson to have those workforce gaps filled,” she said.

Knudsen also drew on her previous experience at the nonprofit health system to argue that AI has the potential to scour large datasets in an unbiased way and spot signals that experts might miss. While Knudsen was at Jefferson, the health system ran a study to see what factors increased a patient’s risk of returning to the hospital within seven days of a major cardiac event. Researchers expected that BMI, blood pressure, or family history might prove to be the main risk factor. The actual answer: credit score. And while researchers didn’t fully understand why that is — for instance, whether a lower score was linked with living in high-stress environments — they’re now able to screen patients discharged from the hospital and offer additional support based on the finding.

While physicians will need some baseline level of familiarity with AI, experts said, they won’t need extensive training in the technology to use it any more than you need to understand alternating current to use an electrical outlet. Bitterman pointed out that while radiation oncologists are trained in the physics behind radioactive isotopes and linear accelerators used to deliver therapy, much of that training isn’t essential to delivering day-to-day care.  

“Does everyone need to become an expert in artificial intelligence? No, and that’s not realistic, I don’t think, to expect of oncologists,” she said.

But Bitterman, whose lab has spent years researching the potential usefulness of large language models in health care, added that one barrier to AI’s adoption is that it’s still hard for patients and providers to understand how these tools generate outputs — and how much they can trust them.

“It’s just a lot to ask a clinician to say, ‘This person has X risk of survival, incorporate that into your treatment recommendation,’ without a lot more backing behind it,” she said.

Bitterman added that there haven’t been enough studies of patients’ views on using AI in health care, and that most existing research focuses on physicians’ attitudes. She said that, for now, many of the tools in development haven’t shown rock-solid evidence of improving patient outcomes, reducing health care costs, or reducing doctor burnout compared to current practices. 

Bitterman said Open AI’s HealthBench, a benchmark recently unveiled to evaluate health care AI, has found that top models have scored around 60% when their responses are graded against a set of physician-written criteria.

Still, panelists said, at some point, these tools may advance to the point that it’s unethical not to use them.

“It’s going to be both visible and invisible, and our community is going to get very quickly over the hump,” Hudis said of AI’s adoption. “I’m not scared; I’m really excited.”