MIT researcher held up as model of how algorithms can benefit humanity
The study found that a version that incorporated both the mammography data and traditional risk factors for breast cancer — such as age and family history of cancer — was equally accurate for white and African American women.
The workings of the algorithm, which predicted that her risk was low, were familiar: Barzilay helped build that very model, after being spurred by her 2014 cancer diagnosis to pivot her research to health care.
Barzilay’s work in AI, which ranges from tools for early cancer detection to platforms to identify new antibiotics, is increasingly garnering recognition: On Wednesday, the Association for the Advancement of Artificial Intelligence named Barzilay as the inaugural recipient of a new annual award honoring an individual developing or promoting AI for the good of society. The award comes with a $1 million prize sponsored by the Chinese education technology company Squirrel AI Learning.
While there are already prizes in the AI field, notably the Turing Award for computer scientists, those existing awards are typically “more focused on scientific, technical contributions and ideas,” said Yolanda Gil, a past president of AAAI and an AI researcher at the University of Southern California. “We didn’t have any that recognized the positive impact that AI is having in our lives.”
With the new award, AAAI aims to counterbalance the widespread messages of concern circulated in the news media and by other commentators about the potential negative impacts of AI. “What we wanted to do with the award is to put out to the public that if we treat AI with fear, then we may not pursue the benefits that AI is having for people,” Gil said.
With the selection of Barzilay, AAAI’s award committee is honoring work in health care — widely seen as one of the most promising fields in which AI is being applied, but also a realm in which plenty can go wrong.
Barzilay has done pioneering work in developing methods for processing language data, including deciphering dead languages, that earned her a “genius grant” from the MacArthur Foundation in 2017. But it was after 2014, the year she was diagnosed with breast cancer, that Barzilay began to focus her attention on health and medicine.
Barzilay’s treatment was successful, and she believes her clinical team at MGH did the best they could in providing her with standard care. At the same time, she said, “it was extremely not satisfying to see how the simplest things that the technology can address were not addressed” — including a delayed diagnosis, an inability to collect data, and statistical flaws in studies used to make treatment decisions.
“Going through it and seeing how much one can do really opened my eyes — I have to contribute,” Barzilay said.
Barzilay said she thinks it’s incumbent not just on the AI community, but also people outside of it, to turn the abundance of research on AI in health care into tools that can improve care.
“We have a humongous body of work in AI in health, and very little of it is actually translated into clinics and benefits patients,” she said.
To try to change that, Barzilay has delved into drug development, building a machine learning platform that was used to identify a novel antibiotic that effectively treated a gastrointestinal bug in mice in a study published earlier this year in the journal Cell.
She’s also co-leading the team developing the AI model for assessing breast cancer risk that was used on her own mammography data in June. In a study published last year in the journal Radiology, Barzilay and her team trained, validated, and tested their model on historical data from about 40,000 women who were screened for breast cancer. They found that their model could discriminate risk better than an older, widely used risk evaluation tool, known as the Tyrer-Cuzick model, that relies on breast density to assess risk.
Barzilay’s model appears to have another important advantage: The study found that a version that incorporated both the mammography data and traditional risk factors for breast cancer — such as age and family history of cancer — was equally accurate for white and African American women. By contrast, the Tyrer-Cuzick tool, developed and validated largely on data from white women, is less effective in African American women.
An updated version of the model described in the Radiology paper is now being implemented in the clinic at MGH. People who come in for a routine screening mammogram automatically have their data analyzed by the model, which spits out four data points: the woman’s breast density and her risk of developing breast cancer in the next one, two, or five years.
For now, with the model still in development, only a subset of MGH radiologists are actually viewing those predictions. They typically do not share the results with individuals who are screened. (There are occasional exceptions, such as in cases in which a woman has read about the model in the news media and makes a specific request to her radiologist to pull up those numbers.)
Another exception was Barzilay, because of her scientific interest and her role in the model’s development. When she went in to MGH for her most recent mammogram, she and Constance Lehman — the hospital’s director of breast imaging who is co-leading the project with Barzilay to develop the model — pulled up the model’s output and discussed the predictions.
Barzilay doesn’t remember the precise probabilities that the model spit out about her personal risk of recurrence, but the overall forecast was reassuring: “There was nothing remarkable there,” she recalled.