Regina Barzilay, PhD
Regina Barzilay's diagnosis seemed to come out of nowhere. She had done all that was recommended by the medical establishment. Regular mammograms starting at age 40 never suggested anything suspicious. She had tested for any gene mutations associated with the disease and found none, and she knew of no family members who had been diagnosed with breast cancer.
So she was shocked when, in 2014, a routine mammogram revealed breast cancer. Barzilay, then 43, underwent two lumpectomies to remove the cancer, and had radiation therapy and chemotherapy. Throughout her treatment, Barzilay, a computer scientist, was struck by the lack of any useful data processing in breast cancer care, or health care in general, for that matter.
For example, there was no easy way to search hospital data to learn the cancer prognosis of women with characteristics similar to Barzilay's, she says. The information was there, but no one was using the full power of computers to look for patterns in the data.
"There was really an unbridgeable gap. Even in 2014, which was not the time of ChatGPT, there was so much you could do with data in all other areas of our life. But it was used so very little in health care," says Barzilay, PhD, a distinguished professor of AI and health at the Massachusetts Institute of Technology (MIT) and co-leader at the Jameel Clinic-MIT Initiative in Machine Learning and Health.
Barzilay thought there was an opportunity to do better.
Finding Another Way
When she finished treatment in December 2014, Barzilay set out to investigate how artificial intelligence (AI) might advance the identification of breast cancer. At the time, AI was hardly a household word, and it took some time to find an investigator willing to work on it with her, she says.
Finally, in 2016, Barzilay found a research partner in Adam Yala, PhD, now an assistant professor in computational precision health statistics and computer science at the University of California, Berkeley. Together, they studied how machine learning could help analyze images and make predictions about future outcomes.
Compared with a lot of health care data that's disorganized and "messy," Barzilay says imaging data is relatively clean, which makes it easier for AI to learn from.
In 2016, Barzilay and Yala crystalised the idea for MIRAI (which means "future" in Japanese), a deep-learning system to predict a patient's future cancer risk by analyzing their mammogram.
How? The AI system was trained on more than 200,000 mammography exams from more than 50,000 patients and learned to identify the subtlest of signs that a patient will develop breast cancer. MIRAI can spot a patient's future risk years before traditional methods can detect the disease.
As a human, we cannot really do this because our capacity to see things is limited. ... But a machine can capture much more subtle patterns.
-- Regina Barzilay, PhD
"As a human, we cannot really do this because our capacity to see things is limited," Barzilay says. "We need to have clear patterns which we can identify. But a machine can capture much more subtle patterns because it looks at all these pixels and has much better ability than our eyes. That's what enables it to capture these very subtle, complicated correlations, which can correspond to future cancer."
Getting Results
Starting in 2020, Barzilay and Yala began testing MIRAI's ability to accurately predict future cancer risk in patients across the world. They tested the accuracy of their model on test data from three hospitals in the U.S., Sweden, and Taiwan.
MIRAI was significantly better than current methods in predicting which people or groups of people were more likely to get breast cancer. At the U.S. hospital (Massachusetts General), MIRAI identified nearly twice the future cancer diagnoses, compared with the current clinical standard (the Tyrer-Cuzick model).
Would MIRAI have detected Barzilay's cancer risk earlier? Yes, she says. The AI system analyzed her own series of previous mammograms and detected a tiny spot in the images that grew over time. The algorithm would have flagged her as "high-risk" before she was diagnosed, she says.
Knowing breast cancer is on the horizon lets patients take earlier action and potentially lessen future risk — potentially saving lives, says Barzilay.
Armed with this knowledge, what are next steps?
"They need to have an MRI, or maybe instead of coming in every 2 years for screening, they need to come every year," Barzilay says.
And there are other possible steps. "Even now, you have preventative chemo drugs for breast cancer that you can give patients. Obviously, you don't want to give it to everybody, because they can have side effects, but maybe somebody who's really high-risk, they may consider taking it."
A Tool For All Races, Ages, and Ethnicities
Research shows Black women younger than 50 are twice as likely to die from breast cancer as White women. American Indian/Alaska Native women, meanwhile, are 17% less likely to be diagnosed with breast cancer as White women but are 4% more likely to die from the cancer. There's also some evidence that breast cancer screening methods and guidelines are based on biased data that contribute to the disparities.
Addressing these issues was a critical part of testing MIRAI's effectiveness, says Barzilay.
"It's really important that [AI] models are addressing the gaps that we already are suffering from in the current system," she says.
So Barzilay has been glad to see that, so far, MIRAI appears to perform similarly across a number of races, ethnicities, ages, and breast density groups.
Loading...
What's Next for MIRAI?
Barzilay and her team are working to make MIRAI part of standard care in hospitals and medical offices. MIRAI is already installed at Massachusetts General Hospital, where staff are working to integrate it into routine treatment. A large prospective trial of MIRAI is also underway at UMass Memorial Hospital in Worcester. Other trials are underway at more than 40 hospitals around the world, including health centers in the U.S., Canada, Mexico, Brazil, and India.
Barzilay, who has been cancer-free for 8 years, hopes MIRAI will become the norm for all women when they have breast cancer screenings.
"Not only for MIRAI, but for many AI tools, the question is, 'How do you bring them into the health care system in a way which changes patient outcomes?'" she says. "It's an extremely complex process of studying how people implement it, and what is the right way to make it a standard of care."
More in This Series
David Fajgenbaum, MD
After finding the cure for his own deadly rare disease in an existing drug, David Fajgenbaum, MD, MFA, devoted himself to using AI to do the same for others.
Suchi Saria, PhD
Sepsis kills -- as Suchi Saria knows from the tragic death of her own nephew. That's part of the driving force behind the AI tool she helped to develop to detect and stop sepsis in hospitals.