- Common risk factors for breast cancer include family history and breast density. But oncologists say these don’t paint the complete picture of a patient’s risk.
- In a new study, AI was more accurate than a standard risk model for predicting breast cancer.
- Experts say better understanding a patient’s risk factor is critical to bettering outcomes.
Breast cancer is the
Arasu wanted to change that and give patients a clearer picture of their risk.
“Traditional risk factors — which we’ve known about for decades — include a woman’s age, family history, prior benign biopsies, estrogen exposure, and breast density,” says Arasu. “Identifying new risk factors would help us identify women who could benefit from more cancer screening with the goal of decreasing advanced breast cancer diagnoses and breast cancer deaths.”
But how?
AI, the same technology which most recently has generated headlines for ChatGPT, might be a critical aid for predicting a person’s breast cancer risk, according to a new study led by Arasu and published Tuesday in Radiology,a journal of the Radiological Society of North America (RSNA).
The study includes thousands of mammograms and indicated that AI could outperform one of the standard clinical risk models currently used to predict a person’s five-year risk of developing breast cancer, known as the Breast Cancer Surveillance Consortium.
“This suggests that AI used alone or combined with current risk prediction models provides a new avenue for future risk prediction,” Arasu says.
Breast cancer specialists not involved in the study hailed the research as promising for healthcare providers and their patients.
“AI holds promise in aiding radiologists detect subtle breast cancer, as well as potentially flag patients who may be at increased risk of breast cancer within the next decade,” says Liva Andrejeva-Wright, MD, a Yale Medicine breast imager (radiologist) and associate professor at Yale School of Medicine.
The study also presents a new use case for AI.
“It’s a new way to look at artificial intelligence,” says Nina Stuzin Vincoff, MD, the chief of breast imaging at Northwell Health in New York. ”We always thought of it as a way to make findings. Now, this study is not about finding cancer there now. It’s about finding out who is at higher risk of developing cancer in the future. It’s a really interesting and important way for artificial intelligence to play a role.”
Arasu explains that the study was retrospective, which means it looked back at what had already occurred.
Arasu and his team started by identifying more than 324,000 women who had a mammogram at Kaiser Permanente Northern California in 2016 and didn’t have a sign of breast cancer.
The team narrowed the participant pool to a random subgroup of 13,628 to analyze.
“We then looked to see which women developed breast cancer between 2016 and 2021,” Arasu explains. “We found there had been 4,584 women with a breast cancer diagnosis. We compared these women to a subgroup that included 13,435 of the 324,000 women who did not develop breast cancer.”
Researchers followed every participant through 2021.
“We evaluated five artificial intelligence algorithms and generated a score for these women’s negative mammograms from 2016,” Arasu says. “These scores are intended for breast cancer detection, but we now evaluated if these same scores could predict future cancer risk out to five years.”
“We also used the Breast Cancer Surveillance Consortium BCSC clinical risk model to assess their breast cancer risk based on their traditional risk factors from 2016,” Arasu added.
The Breast Cancer Surveillance Consortium (BCSC) is a commonly-used model to predict breast cancer risk. It uses self-reported information from the patient and other factors, such as age, family history of breast cancer, birth history, and breast density, and calculates a risk score.
One critical gap?
“There are a lot of factors that factor into whether you are at an increased risk for developing cancer, and someone may not know them,” Vincoff says.
For example, a person may not know their full family history of breast cancer if they were adopted or are estranged from a parent.
Could AI help change that? That’s what Arasu assessed next.
We looked to see whether AI or the BCSC had done a better job at predicting which women would have a breast cancer diagnosis,” Arasu says.
It did.
“The study demonstrates that AI risk assessment models may enhance the identification of average-risk patients who are more likely to develop breast cancer within a five-year time interval,” says Andrejeva-Wright. “In addition, the study suggests that the application of BCSC risk assessment models in combination with AI risk assessment models may lead to enhanced identification of possible patient cohorts within the average risk population who may benefit from enhanced screening.”
As promising as the results of the study are, Arasu says there’s more he’d like to know, evaluate, and improve.
“Further research is needed to see if we can make the algorithms even more accurate,” Arasu says. “We will also need to identify the appropriate way to use this information in clinical practice.”
One radiologist agrees that the findings are exciting but says questions about whether they can translate into doctor’s offices still remain.
“What has not been proven is whether these AI applications can be fully and effectively integrated into mainstream women’s health care,” Richard Reitherman, MD, Ph.D., board-certified radiologist and medical director of breast imaging at MemorialCareBreast Center at Orange Coast Medical Center in Fountain Valley, Calif. “This publication is based on what is called a retrospective analysis of past cases but requires validation in appropriate prospective clinical trials.”
Vincoff doesn’t know precisely if or when patients can expect to see this tool used as part of mammograms. But she says that the fact that researchers didn’t exactly reinvent the wheel of predicting cancer risk holds promise for faster implementation, should the time come.
“It doesn’t require any additional tests,” Vincoff says. “It uses mammography in a whole new way to predict risk. What’s amazing about that is we already have mammograms. You’re adding artificial intelligence to them and getting new information.”
But the add-on factor in predicting, rather than detecting, a cancer that has already developed, is critical.
“The interesting message of this article is that AI might be used to go beyond assisting the radiologist in interpretation to identify mammographic features that are not yet cancer — and therefore cannot be currently diagnosed – but may develop into cancer in the next five years,” says Reitherman.
Better understanding a patient’s risk factor is critical to bettering outcomes.
“The earlier breast cancer is detected, the chances of cure are greater, and the treatments are less onerous and costly,” says Reitherman.
Vincoff also finds this aspect exciting and says it could reduce the need for more intensive procedures, such as mastectomies, in more patients.
But under the current model, Vincoff says patients are being given less customized care.
“We treat everyone like they are average,” Vincoff says. “This study suggests a way we can personalize women’s screenings so it’s not a one-size-fits-all for screening.”
More broadly, Vincoff says AI, though perhaps controversial in other fields like writing, could have life-saving impacts on the future of medicine and breast cancer risk assessment, detection, and care.
“This [study] treats women as the individuals they are,” Vincoff says. “That’s where we want to be in medicine in general, where everyone is getting the care and screening tests that are appropriate to them and their own personal needs.”