Marzyeh Ghassemi was fascinated by video games and puzzles as a baby, but from a young age she was also serious about health. Luckily, she found a method to mix the 2 interests.
“Although I had considered a profession in healthcare, the pull of computer science and engineering was stronger,” says Ghassemi, associate professor in MIT’s Department of Electrical Engineering and Computer Science and the Institute for Medical Engineering and Science (IMES) and provost Researcher on the Laboratory for Information and Decision Systems (LIDS). “When I noticed that computer science basically and AI/ML particularly could possibly be applied to healthcare, it was a convergence of interests.”
Today, Ghassemi and her Healthy ML research group at LIDS are working to comprehensively explore how machine learning (ML) might be made more robust after which used to enhance safety and equity in healthcare.
Growing up in Texas and New Mexico in an engineering-oriented Iranian-American family, Ghassemi had role models to follow in his STEM profession. Although she loved puzzle-based video games—“solving puzzles to unlock other levels or advance was a really attractive challenge”—her mother also exposed her to more advanced mathematics at an early age, which led her to view mathematics as greater than just arithmetic .
“Addition or multiplication are basic skills which can be emphasized for good reason, but the main target can obscure the concept that higher-level math and science are largely about logic and puzzles,” says Ghassemi. “Thanks to my mom’s encouragement, I knew I used to be in for something fun.”
Ghassemi says that along with her mother, many others have supported her mental development. When she earned her bachelor's degree at New Mexico State University, the director of the Honors College and former Marshall Scholar — Jason Ackelson, now a senior adviser to the U.S. Department of Homeland Security — helped her apply for a Marshall Scholarship, which she pursued She went to Oxford University, where she earned a master's degree in 2011 and first became serious about the brand new and rapidly developing field of machine learning. During her doctoral work at MIT, Ghassemi said she received support “from professors and colleagues alike,” adding, “It is that this environment of openness and acceptance that I try to copy for my students.”
During her doctoral research, Ghassemi also got here across the primary indication that machine learning models can hide biases in health data.
She had trained models to predict outcomes using health data, “and the considering on the time was to make use of all the information available. We had seen that neural networks for images learn the best features to perform well, eliminating the necessity to manually develop specific features.”
During a gathering with Leo Celi, senior research scientist on the MIT Laboratory for Computational Physiology and IMES and a member of Ghassemi's dissertation committee, Celi asked whether Ghassemi had checked how well the models worked on patients of various genders, different insurance types, and independent patients. reported races.
Ghassemi checked and there have been gaps. “We now have almost a decade of labor to point out that these modeling gaps are difficult to shut – they’re resulting from existing biases in health data and standard engineering practices. Unless you concentrate on it rigorously, models will naively reproduce and extend biases,” she says.
Since then, Ghassemi has been coping with such topics.
Her biggest breakthrough in her work got here in several parts. First, she and her research group showed that learning models can detect a patient's race from medical images corresponding to chest X-rays, something radiologists cannot do. The group then found that models optimized for “average” performance didn’t perform as well amongst women and minorities. Last summer, her group combined these results to point out that the more a model learned to predict a patient's race or gender from a medical image, the greater the difference in performance could be for subgroups in these populations. Ghassemi and her team found that the issue could possibly be mitigated if a model were trained to account for demographic differences somewhat than specializing in overall average performance – but this process have to be carried out at every site where a model is deployed becomes.
“We emphasize that models trained to optimize performance (balancing overall performance with the smallest equity gap) in a hospital setting aren’t optimal in other settings. “This has a vital impact on how models for human use are developed,” says Ghassemi. “A hospital can have the resources to coach a model after which have the opportunity to exhibit that it performs well, even perhaps with specific fairness constraints in mind. However, our research shows that these performance guarantees don’t hold in recent environments. A model that’s well balanced in a single location may not work effectively in one other environment. This impacts the usefulness of models in practice and it is crucial that we work to handle this issue for those developing and deploying models.”
Ghassemi's work is formed by her identity.
“I’m a visibly Muslim woman and mother – each of which have helped shape the best way I see the world, which influences my research interests,” she says. “I work on the robustness of machine learning models and the way lack of robustness might be combined with existing biases. This interest is not any coincidence.”
As for her thought process, Ghassemi says inspiration often comes from being outside—biking in New Mexico as a student, rowing in Oxford, running as a graduate student at MIT, and lately, walking along the Cambridge Esplanade. She also says that when solving a sophisticated problem, she has found it helpful to think concerning the parts of the larger problem and understand how her assumptions about each part could possibly be improper.
“In my experience, the limiting factor for brand spanking new solutions is what you think that you already know,” she says. “Sometimes it’s difficult to get past your partial knowledge of something until you actually dig deep right into a model, system, etc. and realize that you could have not fully or fully understood a sub-area.”
As passionate as Ghassemi is about her work, she consciously keeps an summary of the larger picture of life.
“If you’re keen on your research, it could possibly be difficult to forestall that from becoming your identity – that’s something I believe many academics need to pay attention to,” she says. “I attempt to be certain that I actually have interests (and knowledge) that transcend my very own technical expertise.
“One of one of the best ways to prioritize balance is to work with good people. If you could have family, friends, or colleagues who encourage you to be a full human being, hold on to them!”
Ghassemi has won many awards and recognition for his work, which spans two early passions – computer science and health – and professes his belief in seeing life as a journey.
“There is a quote from the Persian poet Rumi that translates to ‘You are what you seek,’” she says. “At every stage of your life, you will need to reinvest find yourself and aligning that with who you should be.”