Captivated by video games and puzzles since she was a child, Marzyeh Ghassemi was also fascinated by health from an early age. Luckily, he found a path in which he could combine both interests.
“Although I had considered a career in healthcare, the pull of computer science and engineering was stronger,” says Ghassemi, an associate professor in MIT's Department of Electrical Engineering and Computer Science and the Institute of Medical Engineering and Sciences ( IMES) and Research Director of the Information and Decision Systems Laboratory (LIDS). “When I discovered that computing in general, and ai/ML specifically, could be applied to healthcare, there was a convergence of interests.”
Today, Ghassemi and his Healthy ML research group at LIDS are working on an in-depth study of how machine learning (ML) can be made more robust and subsequently applied to improve health safety and equity.
Growing up in Texas and New Mexico in an engineering-oriented Iranian-American family, Ghassemi had role models in a STEM career. While she loved puzzle-based video games (“Solving puzzles to unlock other levels or progress further was a very engaging challenge”), her mother also involved her in more advanced mathematics early on, encouraging her to see mathematics as something more than arithmetic.
“Addition or multiplication are basic skills that are emphasized for good reason, but the focus can obscure the idea that much of higher-level math and science is more about logic and puzzles,” says Ghassemi. “Thanks to my mother's support, I knew fun things were ahead of me.”
Ghassemi says that besides his mother, many other people supported his intellectual development. While earning her undergraduate degree at New Mexico State University, Honors College director and former Marshall Scholar Jason Ackelson, now a senior advisor to the U.S. Department of Homeland Security, helped her apply for a Marshall Scholarship that allowed her to the University of Oxford, where he earned a master's degree in 2011 and first became interested in the new and rapidly evolving field of machine learning. During his doctoral work at MIT, Ghassemi says he received support “from both professors and peers,” adding, “That environment of openness and acceptance is something I try to replicate for my students.”
While working on his PhD, Ghassemi also found the first clue that biases in health data can be hidden in machine learning models.
He had trained models to predict outcomes using health data, “and the mindset at the time was to use all the data available. “In image neural networks, we had seen that the right features would be learned to achieve good performance, eliminating the need to manually design specific features.”
During a meeting with Leo Celi, a senior research scientist at MIT's Computational Physiology Laboratory and IMES and a member of Ghassemi's thesis committee, Celi asked whether Ghassemi had checked how well the models worked in patients of different genders, insurance types and self-assessments. reported races.
Ghassemi checked it and there were gaps. “We now have almost a decade of work showing that these gaps in models are difficult to address: they arise from existing biases in health data and from default technical practices. Unless carefully thought through, models will naively reproduce and amplify biases,” he says.
Ghassemi has been exploring these questions ever since.
Your favorite breakthrough in the work you've done came in several parts. First, she and her research group showed that learning models could recognize a patient's race from medical images such as chest x-rays, something radiologists cannot do. The group then found that models optimized to work well “on average” didn't work as well for women and minorities. Last summer, his group combined these findings to show that the more a model learned to predict a patient's race or gender from a medical image, the worse its performance gap would be for subgroups of those demographics. Ghassemi and his team found that the problem could be mitigated if a model was trained to take demographic differences into account, rather than focusing on overall average performance, but this process must be done at each site where a model is deployed.
“We are emphasizing that models trained to optimize performance (balance overall performance with the smallest equity gap) in a hospital setting are not optimal in other settings. This has a major impact on how models are developed for human use,” says Ghassemi. “A hospital might have the resources to train a model and then be able to demonstrate that it works well, possibly even with specific equity constraints. However, our research shows that these performance guarantees do not hold in new environments. A model that is well balanced at one site may not work effectively in a different environment. “This affects the usefulness of models in practice and it is essential that we work to address this issue for those developing and implementing models.”
Ghassemi's work is based on his identity.
“I am a visibly Muslim woman and a mother, both of which have helped shape the way I see the world, which informs my research interests,” she says. “I work on the robustness of machine learning models and how lack of robustness can combine with existing biases. That interest is not a coincidence.”
As for his thought process, Ghassemi says inspiration often strikes him when he's outdoors: biking in New Mexico as a student, rowing at Oxford, running as a PhD student at MIT, and, these days, walk along Cambridge Esplanade. He also says that when tackling a complicated problem, he has found it helpful to think about the parts of the larger problem and try to understand how his assumptions about each part might be wrong.
“In my experience, the most limiting factor to new solutions is what you think you know,” he says. “Sometimes it's hard to overcome your own (partial) knowledge about something until you dig really deep into a model, system, etc., and realize you didn't understand a subpart correctly or completely.”
As passionate as Ghassemi is about her work, she intentionally keeps track of the bigger picture of life.
“When you love your research, it can be difficult to prevent that from becoming your identity; it's something I think a lot of academics need to be aware of,” he says. “I try to make sure I have interests (and knowledge) beyond my own technical experience.
“One of the best ways to help prioritize balance is with good people. If you have family, friends or colleagues who encourage you to be a fulfilled person, hold on to them!
Having won many awards and much recognition for work spanning two early passions (computer science and health), Ghassemi professes a belief in seeing life as a journey.
“There is a quote from the Persian poet Rumi that translates as 'You are what you are looking for,'” he says. “At every stage of your life, you have to reinvest in finding who you are and pushing it toward who you want to be.”
